| Beide Seiten der vorigen RevisionVorhergehende ÜberarbeitungNächste Überarbeitung | Vorhergehende Überarbeitung |
| p:ki:fische_ki [2025/04/13 17:01] – [4. Fazit] Ralf Kretzschmar | p:ki:fische_ki [2025/09/09 15:11] (aktuell) – [3. Limitationen] Ralf Kretzschmar |
|---|
| |
| ====== 🐟 Was können wir von den Fischen über KI im Allgemeinen lernen? ====== | ====== 🐟 Was können wir von den Fischen über KI im Allgemeinen lernen? ====== |
| [{{ :p:pasted:chatbot-4071274_640.jpg?320|KI-Chatbot(([[https://pixabay.com/de/illustrations/chatbot-bot-assistent-unterst%C3%BCtzung-4071274/|Picture by Mohamed_hassan]] on Pixabay, [[https://pixabay.com/de/service/license-summary/|Pixabay Licence]])) }}] | <figure right>{{:p:pasted:chatbot-4071274_640.jpg?320}}<caption>KI-Chatbot(([[https://pixabay.com/de/illustrations/chatbot-bot-assistent-unterst%C3%BCtzung-4071274/|Picture by Mohamed_hassan]] on Pixabay, [[https://pixabay.com/de/service/license-summary/|Pixabay Licence]])) </caption></figure> |
| |
| Nachdem du 👩🦰 Sigrún erfolgreich helfen konntest, erweiterst du deinen Horizont, indem du überlegst, was du aus der Fischklassifikation für künstliche Intelligenz im Allgemeinen lernen kannst. Dafür vergleichst du als Zwischenschritt das neuronale Netz für die Fischklassifikation (Fisch-NN) mit aktuellen KI-Chatbots. | Nachdem du 👩🦰 Sigrún erfolgreich helfen konntest, erweiterst du deinen Horizont, indem du überlegst, was du aus der Fischklassifikation für künstliche Intelligenz im Allgemeinen lernen kannst. Dafür vergleichst du als Zwischenschritt das neuronale Netz für die Fischklassifikation (Fisch-NN) mit aktuellen KI-Chatbots. |
| |
| \\ | \\ |
| ~~INTOC~~ | |
| |
| | ~~INTOC~~ |
| \\ | \\ |
| ===== - Funktionsweise ===== | ===== - Funktionsweise ===== |
| |
| In den nachfolgenden zwei Unterkapiteln werden beide Modi beschrieben. | In den nachfolgenden zwei Unterkapiteln werden beide Modi beschrieben. |
| | |
| | \\ |
| ==== - Anwenden ==== | ==== - Anwenden ==== |
| [{{ :p:pasted:fischanwenden.png?313|Anwendung Fisch-NN | <figure right>{{:p:pasted:fischanwenden.png?313}}<caption>Anwendung Fisch-NN |
| ((Eigene Darstellung [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]], der Fisch ist übernommen: Hering | ((Eigene Darstellung [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]], der Fisch ist übernommen: Hering |
| [Citron, Public Domain, [[https://commons.wikimedia.org/wiki/File:Clupea_harengus1.jpg|Clupea harengus]] by [[https://commons.wikimedia.org/wiki/User:Citron|Citron]] on wikimedia))}}] | [Citron, Public Domain, [[https://commons.wikimedia.org/wiki/File:Clupea_harengus1.jpg|Clupea harengus]] by [[https://commons.wikimedia.org/wiki/User:Citron|Citron]] on wikimedia))</caption></figure> |
| |
| ; 🐟 Fisch-NN\\ \\ | ; 🐟 Fisch-NN\\ \\ |
| == ✍ Auftrag == | == ✍ Auftrag == |
| - Ersetze im Text die ''🤖'' in den ''%%[🤖]%%'' durch folgende Wörter, welche du natürlich grammatikalisch an den Text anpasst:\\ ''Ausgabe'', ''Eingabe'', ''Gewicht'', ''Training Set''. \\ ⚠️ Jedes Wort kann hier nur einmal verwendet werden, die ''['' '']'' sollten stehen gelassen werden, dann ist auch nach dem Ausfüllen klar, wo die Lücke war. | - Ersetze im Text die ''🤖'' in den ''%%[🤖]%%'' durch folgende Wörter, welche du natürlich grammatikalisch an den Text anpasst:\\ ''Ausgabe'', ''Eingabe'', ''Gewicht'', ''Training Set''. \\ ⚠️ Jedes Wort kann hier nur einmal verwendet werden, die ''['' '']'' sollten stehen gelassen werden, dann ist auch nach dem Ausfüllen klar, wo die Lücke war. |
| {{gem/match?0=N4IgLgpgHmIFwgCIEMDOACA6gS1aiAduhNgRBmQK4BOA9gcgDaHoByEYAXuermL0XykWAbUB8G4DRdgLroA5uQAOJAMYALCNTAA6dAEkCqpmEbY16AGZpU-ACbZyPHHhY206cdPSdKqZAFs-QgAadBstVR07CAArQmx5IhRIAhC-bFtCFh4wamRSe010AHcaGwgdHGobUIg-dBy8ggLk9Co6BmYidi5iMCK3EgJ5eQAjYPRUWnQAHRANUjJNWYnTVVC3NvomCDYOTlCaMyEydABxCCLVyHRKIYhzABPVanXclOKIRjUd7DqAFVy+SG6AAyhxiAQwIZGMlRJIZGUXhpobCiABRUiyZBjdD+dBPAgmNTWFj9agsW7VS47KJeSjyVQAN+o8Thnl+6ABjUxoPByB8vnUBB0KFylHM6FUpDk2BGtg2EBoW2Yux6HH6GH5qBCrjw6D8jNkRNU1nxqkJqzpoXs+CIs3mBEWyWWAEEfFixkQ+FoQEEQOZGMhZKh4CBsL6QOT5FAFKGtAAqY6EBPnS4Jpy2hNc0gJ7y+ALJ+OEKFGTLC+MYobIXP0iBMlkJBNu1BYhOi-nmBMgAC+QA#ab9098101ec11101}} | {{gem/match?0=N4IgLgpgHmIFwgCIEMDOACA6gS1aiAduhNgRBmQK4BOA9gcgDaHoByEYAXuermL0XykWAbUB8G4DRdgLroA5uQAOJAMYALCNTAA6dAEkCqpmEbY16AGZpU-ACbZyPHHhY206cdPSdKqZAFs-QgAadBstVR07CAArQmx5IhRIAhC-bFtCFh4wamRSe010AHcaGwgdHGobUIg-dBy8ggLIdCo6BmYidi5iMCK3EgJ5eQAjYPRUWnQFK2aQ1FICEexGGxMzeVQFOjVx1wwAMVpGBLKdFAoIGnomCDYOTmJqVQATxn5Bu48ZSgJqgDcNLJmKZVLZ7Ph0Gl+EIyNR0ABRaiLDS-WToZBEJjMRGkWTIMZEbB1AAquXyQ3QAGUOOdMcgzEUXqp3sQMKQJp94UiUdQ0aF7LEmgl0GTGnjUGNcAp7DjkD4QkUIIxdjVuXiCXd-K0XmprCwANa0AIsX7VDS8mGg0LIXIpYr2AV3bzyVQAN+RCRCxNF5KalJpHwIYEM7xYAEEfPjCRifL51AQdAAhNAFQjVeXmJ34Ig8ghwtEscy0ah+SgmDS2NxtG447qPWFfSRSHSsZBgAA-1HWqnQjBelHM4Ic8IACnRuHhWqD+EVHaoCugxqg1MiIDGV6pkWA+iRIQT5vQytQsYxmBhzKQyvxC+ZmfDgSpVNx0vEJmB29oQEEQOZGMhZFQeAQGwb8QGoCB5CgBRgK0AAqRtEzggBxCA53gyw8G0OCxVIeCFnzZZVh7eCEQ1DCSzLCtNHgyNUHxeDx1oSdUHgkAAF8gA#800f4ea5b8ce874d}} |
| </WRAP> | </WRAP> |
| | \\ |
| ==== - Trainineren ==== | ==== - Trainineren ==== |
| [{{ :p:pasted:fischtrainineren.png?290|Training Fisch-NN | <figure right>{{ :p:pasted:fischtrainineren.png?290}}<caption>Training Fisch-NN |
| ((Eigene Darstellung [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]], die Fische sind übernommen: Hering | ((Eigene Darstellung [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]], die Fische sind übernommen: Hering |
| [Citron, Public Domain, [[https://commons.wikimedia.org/wiki/File:Clupea_harengus1.jpg|Clupea harengus]] by [[https://commons.wikimedia.org/wiki/User:Citron|Citron]] on wikimedia; Lodde [Fb78, Public Domain][[https://commons.wikimedia.org/wiki/File:Mallotus_villosus.gif|Mallotus villosus]] by | [Citron, Public Domain, [[https://commons.wikimedia.org/wiki/File:Clupea_harengus1.jpg|Clupea harengus]] by [[https://commons.wikimedia.org/wiki/User:Citron|Citron]] on wikimedia; Lodde [Fb78, Public Domain][[https://commons.wikimedia.org/wiki/File:Mallotus_villosus.gif|Mallotus villosus]] by |
| [[https://commons.wikimedia.org/wiki/User:Fb78|Fb78]] on wikimedia.))}}] | [[https://commons.wikimedia.org/wiki/User:Fb78|Fb78]] on wikimedia.))</caption></figure> |
| |
| ; 🐟 Fisch-NN\\ \\ | ; 🐟 Fisch-NN\\ \\ |
| == ✍ Auftrag == | == ✍ Auftrag == |
| - Ersetze im Text die ''🤖'' in den ''%%[🤖]%%'' durch folgende Wörter, welche du natürlich grammatikalisch an den Text anpasst:\\ ''Ausgabe'', ''Gewicht'', ''Training Set''. \\ ⚠️ Jedes Wort kann mehrfach vorkommen, die ''['' '']'' sollten stehen gelassen werden, dann ist auch nach dem Ausfüllen klar, wo die Lücke war. | - Ersetze im Text die ''🤖'' in den ''%%[🤖]%%'' durch folgende Wörter, welche du natürlich grammatikalisch an den Text anpasst:\\ ''Ausgabe'', ''Gewicht'', ''Training Set''. \\ ⚠️ Jedes Wort kann mehrfach vorkommen, die ''['' '']'' sollten stehen gelassen werden, dann ist auch nach dem Ausfüllen klar, wo die Lücke war. |
| {{gem/match?0=N4IgLgpgHmIFwgOoBOAWAnCA7AJgAhwgGc8AVdAQwEssaBzEgdwnUKwKojwG1A+DcDRdgLp4INYniwQArugD2WCgBtseAHIQwAL3EAjCkU6ZceClIBmBFQBEKkdlQC2eAG4tmuDSr5CTWOhAAHfSIwADo8GxJJGXklLnUtPGV0LDACCh1RS3snb2FsMFQlOxUALSkiCgcHbCK-PE1GKiIAY1QVNjwAUXoMlSljTv9GAB+sVtQ7HgFhIlk8Oik0pRIHADe6RSo28JAAGhAzRQoGeBAqfZBMfygAs9CAKnJqLEeAcQgmnYebO0fnmiPZKpf6UQEPcqVaqPACCFToFEe60221QIAAvkA#0810b65aa898237f}} | {{gem/match?0=N4IgLgpgHmIFwgOoBOAWAnCA7AJgAhwgGc8AVdAQwEssaBzEgdwnUKwKojyoFs8sIAV3QB7LBQA22PADkIYAF55sYVJMgD2AbUB8G4DRdgLp4K7HJzwARChu58Abi2a5503YeNY6EAA4UiRMAA6Sz9+IVFxKVl5JSl0LDACCgAjCCoCaV48NyMVNQkNaQAtQSIKHh5sNU88BUYqIgBjVGk2PABRehTpQVwM9i9GAB+sJtQbHLwiETw6QUTJEh4AN7oJKmbgkAAaEAAzCQoGeBAqHZBMLyhvE8CAKjz1bGx7gHEIes27h3QnQg17uRqFh7iYvL5-EE7kCaPcSmUKlUwfcAIKlOgUe4rNYbVAgAC+QA#48f6cd78fe8678a9}} |
| </WRAP> | </WRAP> |
| |
| |
| 💡 Das Konstruieren, Trainieren und Anwenden einer KI ist in der Praxis mit zahlreichen Herausforderungen verbunden. Diese werden in den folgenden vier Unterkapiteln genauer vorgestellt. | 💡 Das Konstruieren, Trainieren und Anwenden einer KI ist in der Praxis mit zahlreichen Herausforderungen verbunden. Diese werden in den folgenden vier Unterkapiteln genauer vorgestellt. |
| | |
| | \\ |
| ==== - Datensatz zusammenstellen ==== | ==== - Datensatz zusammenstellen ==== |
| [{{ :p:pasted:fischdatenset.png?266|Fisch Datensatz | <figure right>{{ :p:pasted:fischdatenset.png?266}}<caption>Fisch Datensatz |
| ((Eigene Darstellung [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]], die Fische sind übernommen: Hering | ((Eigene Darstellung [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]], die Fische sind übernommen: Hering |
| [Citron, Public Domain, [[https://commons.wikimedia.org/wiki/File:Clupea_harengus1.jpg|Clupea harengus]] by [[https://commons.wikimedia.org/wiki/User:Citron|Citron]] on wikimedia; Lodde [Fb78, Public Domain][[https://commons.wikimedia.org/wiki/File:Mallotus_villosus.gif|Mallotus villosus]] by | [Citron, Public Domain, [[https://commons.wikimedia.org/wiki/File:Clupea_harengus1.jpg|Clupea harengus]] by [[https://commons.wikimedia.org/wiki/User:Citron|Citron]] on wikimedia; Lodde [Fb78, Public Domain][[https://commons.wikimedia.org/wiki/File:Mallotus_villosus.gif|Mallotus villosus]] by |
| [[https://commons.wikimedia.org/wiki/User:Fb78|Fb78]] on wikimedia.))}}] | [[https://commons.wikimedia.org/wiki/User:Fb78|Fb78]] on wikimedia.))</caption></figure> |
| |
| ; 🐟 Fisch-NN\\ \\ | ; 🐟 Fisch-NN\\ \\ |
| |
| ; 💬 KI-Chatbot\\ \\ | ; 💬 KI-Chatbot\\ \\ |
| : Für das Training eines KI-Chatbots, der freundlich und qualitativ hochwertig antworten soll, werden freundliche und qualitativ hochwertige Texte benötigt. Wird ein KI-Chatbot mehrheitlich mit fremdenfeindlichen Texten trainiert, so wird er auf die meisten Fragen mit fremdenfeindlichen Aussagen antworten.\\ \\ | : Für das Training eines KI-Chatbots, der freundlich und qualitativ hochwertig antworten soll, werden freundliche und qualitativ hochwertige Texte benötigt. Wird ein KI-Chatbot mehrheitlich mit fremdenfeindlichen Texten trainiert, so wird er auf die meisten Fragen mit fremdenfeindlichen Aussagen antworten. KI-Chatbots werden mehrheitlich mit Texten aus dem Internet trainiert. Diese sind häufig bezüglich Meinungen und Ansichten nicht neutral. Als Folge davon kann ein KI-Chatbot ebenfalls nicht als neutral bezeichnet werden.\\ \\ |
| : Für das Pre-Training des dafür genutzen, riesigen neuronalen Netzes wird eine Unmenge von Texten benötigt. Werden zu wenige Texte verwendet tendiert ein so grosses neuronales Netz dazu, die Trainingsdaten auswendig zu lernen (das als Overfitting bezeichnet). Die benötigte Textmenge ist zu gross, um diese von Hand zusammenstellen oder aussortieren zu können. Daher kommen auch viele qualitativ schlechte Texte beim Pre-Training zum Einsatz. Es ist unklar, ob alle geeigneten, von Menschen verfassten digital verfügbaren Texte genügen, um in Zukunft weitere noch umfangreichere KI-Chatbots zu trainieren. Hinzu kommt, dass Menschen mittlerweile auch von KI-Chatbots verfasste Texte im Internet veröffentlichen, welche somit voraussichtlich auch für das Training zukünftiger KI-Chatbots berücksichtigt werden. Es ist ebenso unklar, inwieweit sich die Qualität der KI-Chatbots verringern wird, wenn für das Training zusätzlich eine grössere Menge KI-generierte Texte zum Einsatz kommt.\\ \\ | : Für das Pre-Training des dafür genutzen, riesigen neuronalen Netzes wird eine Unmenge von Texten benötigt. Werden zu wenige Texte verwendet tendiert ein so grosses neuronales Netz dazu, die Trainingsdaten auswendig zu lernen (das als Overfitting bezeichnet). Die benötigte Textmenge ist zu gross, um diese von Hand zusammenstellen oder aussortieren zu können. Daher kommen auch viele qualitativ schlechte Texte beim Pre-Training zum Einsatz. Es ist unklar, ob alle geeigneten, von Menschen verfassten digital verfügbaren Texte genügen, um in Zukunft weitere noch umfangreichere KI-Chatbots zu trainieren. Hinzu kommt, dass Menschen mittlerweile auch von KI-Chatbots verfasste Texte im Internet veröffentlichen, welche somit voraussichtlich auch für das Training zukünftiger KI-Chatbots berücksichtigt werden. Es ist ebenso unklar, inwieweit sich die Qualität der KI-Chatbots verringern wird, wenn für das Training zusätzlich eine grössere Menge KI-generierte Texte zum Einsatz kommt.\\ \\ |
| : Im Fine-Tuning Prozess wird versucht, den KI-Chatbots mit relativ wenigen, qualitativ hochwertigen Texten "nachträglich" ein gewünschtes Verhalten einzuimpfen. Leider ist davon auszugehen, dass die Menschen, welche diese Daten zusammenstellen, häufig schlecht bezahlt (z.B. wenige Franken Stundenlohn) und leistungstechnisch unter Druck gesetzt werden. Hinzu kommt, dass sie insbesondere auch unerwünschte, zum Teil sehr belastende Inhalte sichten und kennzeichnen müssen. Seit der Veröffentlichung leistungsfähiger KI-Chatbots ist davon auszugehen, dass KI-Chatbots von den Betroffenen genutzt werden, um die Arbeit schneller und erträglicher erledigen zu können. Das würde bedeuten, dass auch im Fine-Tuning die KI manchmal von einer KI trainiert wird. | : Im Fine-Tuning Prozess wird versucht, den KI-Chatbots mit relativ wenigen, qualitativ hochwertigen Texten "nachträglich" ein gewünschtes Verhalten einzuimpfen. Leider ist davon auszugehen, dass die Menschen, welche diese Daten zusammenstellen, häufig schlecht bezahlt (z.B. wenige Franken Stundenlohn) und leistungstechnisch unter Druck gesetzt werden. Hinzu kommt, dass sie insbesondere auch unerwünschte, zum Teil sehr belastende Inhalte sichten und kennzeichnen müssen. Seit der Veröffentlichung leistungsfähiger KI-Chatbots ist davon auszugehen, dass KI-Chatbots von den Betroffenen genutzt werden, um die Arbeit schneller und erträglicher erledigen zu können. Das würde bedeuten, dass auch im Fine-Tuning die KI manchmal von einer KI trainiert wird. |
| {{exorciser/jspg?javascript=%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%2F%2F%20Diese%20Trainings-%20und%20Updateparameter%20d%C3%BCrfen%20ge%C3%A4ndert%20werden.%0Aconst%20LERNRATE%20%3D%201%3B%20%2F%2F%20Dezimalzahl%20%3E%200%0Aconst%20ANZAHL_HIDDEN_NEURONS%20%3D%203%3B%20%2F%2F%20ganzzahliger%20Wert%20%3E%200%0Aconst%20ANZAHL_EPOCHEN%20%3D%201000%3B%20%2F%2F%20ganzzahliger%20Wert%20%3E%200%0Aconst%20EPOCHEN_FUER_GRAFIK_UPDATE%20%3D%20100%3B%20%2F%2F%20ganzzahlig%20Wert%20%3E%200%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20CONSTANTS%20AND%20GLOBAL%20VARIABLES%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%20constants%0Aconst%20DATA_RANGE%20%3D%2010%3B%0Aconst%20DATA_WINDOW_SIZE%20%3D%20200%3B%0Aconst%20DATA_SCALE%20%3D%20Math.round%28DATA_WINDOW_SIZE%20%2F%20DATA_RANGE%29%3B%0Aconst%20NUM_DATA_PIXEL%20%3D%2050%3B%0Aconst%20PIXEL_SIZE%20%3D%20Math.round%28DATA_WINDOW_SIZE%20%2F%20NUM_DATA_PIXEL%29%3B%0A%0A%2F%2F%20training%20data%0Aconst%20trainInputArray%20%3D%20%5B%0A%09%5B2.5%2C%206.0%5D%2C%0A%09%5B3.5%2C%201.5%5D%2C%0A%09%5B3.5%2C%204.0%5D%2C%0A%09%5B4.0%2C%206.0%5D%2C%0A%09%5B4.0%2C%207.0%5D%2C%0A%09%5B1.5%2C%203.0%5D%2C%0A%09%5B1.5%2C%204.5%5D%2C%0A%09%5B1.5%2C%207.0%5D%2C%0A%09%5B2.0%2C%201.5%5D%2C%0A%09%5B3.0%2C%203.0%5D%2C%0A%09%5B3.0%2C%205.5%5D%2C%0A%09%5B3.0%2C%206.5%5D%2C%0A%09%5B4.0%2C%208.5%5D%2C%0A%5D%3B%0A%0Aconst%20trainOutputArray%20%3D%20%5B0%2C%200%2C%200%2C%200%2C%200%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%5D%3B%0Aconst%20trainInputMax%20%3D%20%5B4.0%2C%208.5%5D%3B%0Aconst%20trainInputMin%20%3D%20%5B1.5%2C%201.5%5D%3B%0A%0A%2F%2F%20validation%20data%0Aconst%20validInputArray%20%3D%20%5B%0A%09%5B2.5%2C%207.0%5D%2C%0A%09%5B3.5%2C%201.0%5D%2C%0A%09%5B3.5%2C%207.0%5D%2C%0A%09%5B4.0%2C%203.5%5D%2C%0A%09%5B4.5%2C%206.5%5D%2C%0A%09%5B4.5%2C%202.5%5D%2C%0A%09%5B4.5%2C%205.0%5D%2C%0A%09%5B6.0%2C%202.5%5D%2C%0A%09%5B6.0%2C%204.5%5D%2C%0A%09%5B6.5%2C%205.5%5D%2C%0A%09%5B7.5%2C%204.5%5D%2C%0A%09%5B8.5%2C%205.0%5D%2C%0A%09%5B1.5%2C%205.0%5D%2C%0A%09%5B1.5%2C%206.5%5D%2C%0A%09%5B2.0%2C%201.5%5D%2C%0A%09%5B2.5%2C%203.5%5D%2C%0A%09%5B2.5%2C%206.0%5D%2C%0A%09%5B3.5%2C%206.0%5D%2C%0A%09%5B4.0%2C%208.0%5D%2C%0A%09%5B4.0%2C%207.0%5D%2C%0A%09%5B5.5%2C%208.5%5D%2C%0A%09%5B6.5%2C%207.0%5D%2C%0A%09%5B7.5%2C%208.5%5D%2C%0A%09%5B8.5%2C%207.0%5D%0A%5D%3B%0Aconst%20validOutputArray%20%3D%20%5B0%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%5D%3B%0A%0A%2F%2F%20global%20variables%0Alet%20model%2C%20trainInput%2C%20trainOutput%2C%20validInput%2C%20validOutput%2C%20testInput%2C%20actualEpoch%2C%20actualLogs%2C%20trainBuffer%2C%20validBuffer%2C%20decisionBuffer%3B%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20NEURAL%20NETWORK%20FUNCTIONS%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20generate%20data%20tensors%0A%2F%2F%0Afunction%20generateTensors%28%29%20%7B%0A%09let%20normalizedTrainInputArray%20%3D%20%5B%5D%3B%0A%09let%20normalizedValidInputArray%20%3D%20%5B%5D%3B%0A%09let%20normalizedTestInputArray%20%3D%20%5B%5D%3B%0A%0A%09%2F%2F%20normalize%20training%20and%20validation%20inputs%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20trainInputArray.length%3B%20i%2B%2B%29%20%7B%0A%09%09normalizedTrainInputArray.push%28%5B%28trainInputArray%5Bi%5D%5B0%5D%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28trainInputArray%5Bi%5D%5B1%5D%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%7D%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20validInputArray.length%3B%20i%2B%2B%29%20%7B%0A%09%09normalizedValidInputArray.push%28%5B%28validInputArray%5Bi%5D%5B0%5D%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28validInputArray%5Bi%5D%5B1%5D%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%7D%0A%0A%09%2F%2F%20generate%20training%20input%20and%20output%20tensors%0A%09trainInput%20%3D%20tf.tensor2d%28normalizedTrainInputArray%2C%20%5BnormalizedTrainInputArray.length%2C%202%5D%29%3B%0A%09trainOutput%20%3D%20tf.tensor2d%28trainOutputArray%2C%20%5BtrainOutputArray.length%2C%201%5D%29%3B%0A%09%2F%2F%20let%20trainOutput%20%3D%20tf.oneHot%28tf.tensor1d%28trainOutputArray%29.toInt%28%29%2C%202%29%3B%0A%09%2F%2F%20trainInput.print%28%29%3B%20%2F%2F%20Print%20Tensor%0A%0A%09%2F%2F%20generate%20validation%20input%20and%20output%20tensors%0A%09validInput%20%3D%20tf.tensor2d%28normalizedValidInputArray%2C%20%5BnormalizedValidInputArray.length%2C%202%5D%29%3B%0A%09validOutput%20%3D%20tf.tensor2d%28validOutputArray%2C%20%5BvalidOutputArray.length%2C%201%5D%29%3B%0A%09%2F%2F%20let%20validOutput%20%3D%20tf.oneHot%28tf.tensor1d%28validOutputArray%29.toInt%28%29%2C%202%29%3B%0A%0A%09%2F%2F%20calculate%20normalized%20testing%20inputs%20for%20visualisation%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20i%20%3D%20i%20%2B%201%29%20%7B%0A%09%09for%20%28j%20%3D%200%3B%20j%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20j%20%3D%20j%20%2B%201%29%20%7B%0A%09%09%09normalizedTestInputArray.push%28%5B%28i%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28j%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%09%7D%0A%09%7D%0A%0A%09%2F%2F%20generate%20testing%20input%20tensor%20for%20visualisation%0A%09testInput%20%3D%20tf.tensor2d%28normalizedTestInputArray%2C%20%5BnormalizedTestInputArray.length%2C%202%5D%29%3B%0A%09%2F%2F%09testInput.print%28%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20build%20neural%20network%20model%20and%20define%20training%0A%2F%2F%0Afunction%20compileModel%28%29%20%7B%0A%09%2F%2F%20neural%20network%20structure%0A%09model%20%3D%20tf.sequential%28%7B%0A%09%09layers%3A%20%5B%0A%09%09%09tf.layers.dense%28%7B%0A%09%09%09%09name%3A%20%27HiddenLayer1%27%2C%0A%09%09%09%09inputShape%3A%20%5B2%5D%2C%0A%09%09%09%09units%3A%20ANZAHL_HIDDEN_NEURONS%2C%0A%09%09%09%09activation%3A%20%27tanh%27%0A%09%09%09%7D%29%2C%0A%09%09%09tf.layers.dense%28%7B%0A%09%09%09%09name%3A%20%27OutputLayer%27%2C%0A%09%09%09%09units%3A%201%2C%0A%09%09%09%09activation%3A%20%27sigmoid%27%0A%09%09%09%7D%29%0A%09%09%5D%0A%09%7D%29%3B%0A%0A%09%2F%2F%20neural%20network%20training%0A%09model.compile%28%7B%0A%09%09optimizer%3A%20tf.train.sgd%28LERNRATE%29%2C%0A%09%09loss%3A%20tf.losses.meanSquaredError%2C%0A%09%09metrics%3A%20%5B%27mse%27%5D%2C%0A%09%7D%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20train%20neural%20network%0A%2F%2F%0Aasync%20function%20trainModel%28%29%20%7B%0A%09await%20model.fit%28%0A%09%09trainInput%2C%20trainOutput%2C%20%7B%0A%09%09%09epochs%3A%20ANZAHL_EPOCHEN%2C%0A%09%09%09shuffle%3A%20true%2C%0A%09%09%09callbacks%3A%20%5B%7B%0A%09%09%09%09onEpochEnd%3A%20async%20%28epoch%2C%20logs%29%20%3D%3E%20%7B%0A%09%09%09%09%09%2F%2F%20remember%20epoch%20number%20if%20interrupted%0A%09%09%09%09%09actualEpoch%20%3D%20epoch%20%2B%201%3B%0A%09%09%09%09%09actualLogs%20%3D%20logs%3B%0A%09%09%09%09%09%2F%2F%20initiate%20graphic%20update%0A%09%09%09%09%09if%20%28%28%28actualEpoch%29%20%25%20EPOCHEN_FUER_GRAFIK_UPDATE%20%3D%3D%3D%200%29%20%7C%7C%20%28epoch%20%3D%3D%3D%200%29%29%20%7B%0A%09%09%09%09%09%09drawPrediction%28actualEpoch%2C%20logs.loss%29%3B%0A%09%09%09%09%09%7D%0A%09%09%09%09%7D%0A%09%09%09%7D%2C%20%5D%0A%09%09%7D%0A%09%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20calculate%20mean%20squared%20error%0A%2F%2F%0Afunction%20calculateMSE%28predictedOutput%2C%20trueOutput%29%20%7B%0A%09let%20mse%20%3D%200.0%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20predictedOutput.length%3B%20i%2B%2B%29%20%7B%0A%09%09mse%20%3D%20mse%20%2B%20%28predictedOutput%5Bi%5D%20-%20trueOutput%5Bi%5D%29%20%2a%2a%202%3B%0A%09%7D%0A%09return%20mse%20%2F%20predictedOutput.length%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20calculate%20percentage%20of%20correct%20classifications%0A%2F%2F%0Afunction%20calculatePercentageCorrect%28predictedOutput%2C%20trueOutput%29%20%7B%0A%09let%20correct%20%3D%200%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20predictedOutput.length%3B%20i%2B%2B%29%20%7B%0A%09%09if%20%28%28trueOutput%5Bi%5D%20%3D%3D%3D%201%20%26%26%20predictedOutput%5Bi%5D%20%3E%3D%200.5%29%20%7C%7C%20%28trueOutput%5Bi%5D%20%3D%3D%3D%200%20%26%26%20predictedOutput%5Bi%5D%20%3C%200.5%29%29%20%7B%0A%09%09%09correct%20%3D%20correct%20%2B%201%3B%0A%09%09%7D%0A%09%7D%0A%09return%20%7B%0A%09%09accuracy%3A%20Math.round%28correct%20%2a%20100%20%2F%20predictedOutput.length%29%2C%0A%09%09wrong%3A%20predictedOutput.length%20-%20correct%0A%09%7D%3B%0A%7D%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20FUNCTIONS%20FOR%20VISUALISATION%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20decision%20boundary%20and%20network%20output%0A%2F%2F%0Afunction%20drawPrediction%28epoch%2C%20mseTrain%29%20%7B%0A%09%2F%2F%20make%20predictions%20with%20neural%20network%0A%09let%20training%20%3D%20model.predict%28trainInput%29.arraySync%28%29%3B%0A%09let%20validation%20%3D%20model.predict%28validInput%29.arraySync%28%29%3B%0A%09let%20prediction%20%3D%20model.predict%28testInput%29.arraySync%28%29%3B%0A%0A%09%2F%2F%20calculate%20mse%20and%20percentages%20of%20correct%20classifications%0A%09let%20mseValid%20%3D%20calculateMSE%28validation%2C%20validOutputArray%29%3B%0A%09let%20trainCorrect%20%3D%20calculatePercentageCorrect%28training%2C%20trainOutputArray%29%3B%0A%09let%20validCorrect%20%3D%20calculatePercentageCorrect%28validation%2C%20validOutputArray%29%3B%0A%0A%09%2F%2F%20draw%20decision%20boundary%0A%09decisionBuffer.noStroke%28%29%3B%0A%09let%20element%20%3D%200%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20i%20%3D%20i%20%2B%201%29%20%7B%0A%09%09for%20%28j%20%3D%200%3B%20j%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20j%20%3D%20j%20%2B%201%29%20%7B%0A%09%09%09%2F%2F%20transform%20neural%20network%20output%20into%20a%20color%0A%09%09%09if%20%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28Math.round%28%281%20-%20prediction%5Belement%5D%5B0%5D%29%20%2a%20100%20%2B%20200.5%29%2C%20255%2C%20Math.round%28%281%20-%20prediction%5Belement%5D%5B0%5D%29%20%2a%20100%20%2B%20200.5%29%29%3B%0A%09%09%09%7D%20else%20%7B%0A%09%09%09%09decisionBuffer.fill%28Math.round%28prediction%5Belement%5D%5B0%5D%20%2a%20100%20%2B%20200.5%29%2C%20Math.round%28prediction%5Belement%5D%5B0%5D%20%2a%20100%20%2B%20200.5%29%2C%20255%29%3B%0A%09%09%09%7D%0A%09%09%09%2F%2F%20check%20for%20decision%20boundary%20and%20alter%20color%0A%09%09%09if%20%28%28element%20%25%20%28NUM_DATA_PIXEL%20%2B%201%29%29%20%21%3D%3D%200%20%26%26%20%28%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%20%26%26%20prediction%5Belement%20-%201%5D%5B0%5D%20%3C%3D%200.5%29%20%7C%7C%20%28prediction%5Belement%5D%5B0%5D%20%3C%3D%200.5%20%26%26%20prediction%5Belement%20-%201%5D%5B0%5D%20%3E%200.5%29%29%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28%27red%27%29%3B%0A%09%09%09%7D%0A%09%09%09if%20%28element%20%3E%3D%20NUM_DATA_PIXEL%20%2B%201%20%26%26%20%28%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%20%26%26%20prediction%5Belement%20-%20%28NUM_DATA_PIXEL%20%2B%201%29%5D%5B0%5D%20%3C%3D%200.5%29%20%7C%7C%20%28prediction%5Belement%5D%5B0%5D%20%3C%3D%200.5%20%26%26%20prediction%5Belement%20-%20%28NUM_DATA_PIXEL%20%2B%201%29%5D%5B0%5D%20%3E%200.5%29%29%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28%27red%27%29%3B%0A%09%09%09%7D%0A%09%09%09decisionBuffer.rect%28DATA_SCALE%20%2a%20i%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20PIXEL_SIZE%20%2F%202%2C%20DATA_SCALE%20%2a%20%2810%20-%20j%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%29%20-%20PIXEL_SIZE%20%2F%202%2C%20PIXEL_SIZE%2C%20PIXEL_SIZE%29%3B%0A%09%09%09element%20%3D%20element%20%2B%201%3B%0A%09%09%7D%0A%09%7D%0A%09image%28decisionBuffer%2C%200%2C%200%29%3B%0A%09image%28decisionBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20draw%20training%20and%20validation%20samples%0A%09image%28trainBuffer%2C%200%2C%200%29%3B%0A%09image%28validBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20add%20labels%20to%20training%20and%20validation%20feature%20space%20%2F%2F%20wrong%0A%09noStroke%28%29%3B%0A%09text%28%22Training%3A%20%22%20%2B%20trainCorrect.accuracy%20%2B%20%22%25%20%E2%9C%94%EF%B8%8F%20%28%22%20%2B%20trainCorrect.wrong%20%2B%20%22%20falsch%29%22%2C%202%20%2a%20PIXEL_SIZE%2C%204%20%2a%20PIXEL_SIZE%29%3B%0A%09text%28%22Validation%3A%20%22%20%2B%20validCorrect.accuracy%20%2B%20%22%25%20%E2%9C%94%EF%B8%8F%28%22%20%2B%20validCorrect.wrong%20%2B%20%22%20falsch%29%22%2C%20DATA_WINDOW_SIZE%20%2B%202%20%2a%20PIXEL_SIZE%2C%204%20%2a%20PIXEL_SIZE%29%3B%0A%0A%09%2F%2F%20print%20training%20and%20validation%20results%20to%20console%0A%09console.log%28%22%7C%7C%20%22%20%2B%20epoch.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%20%7C%7C%20%22%20%2B%20trainCorrect.accuracy.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%25%20%7C%20%22%20%2B%20mseTrain.toFixed%286%29%20%2B%20%22%20%7C%7C%20%22%20%2B%20validCorrect.accuracy.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%25%20%7C%20%22%20%2B%20mseValid.toFixed%286%29%20%2B%20%22%20%7C%7C%22%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20draw%20data%20in%20graphic%20buffer%0A%2F%2F%0Afunction%20drawDataInBuffer%28inputData%2C%20outputData%2C%20buffer%29%20%7B%0A%09buffer.noFill%28%29%3B%0A%09buffer.strokeWeight%28PIXEL_SIZE%20%2F%202%29%3B%0A%09buffer.rect%280%2C%200%2C%20DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20inputData.length%3B%20i%2B%2B%29%20%7B%0A%09%09if%20%28outputData%5Bi%5D%20%3D%3D%3D%201%29%20%7B%0A%09%09%09%2F%2F%20green%20circle%20for%20output%20label%20%3D%3D%3D%201%0A%09%09%09buffer.stroke%28%22darkgreen%22%29%3B%0A%09%09%09buffer.circle%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%2C%202%20%2a%20PIXEL_SIZE%29%3B%0A%09%09%7D%20else%20%7B%0A%09%09%09%2F%2F%20blue%20cross%20for%20output%20label%20%3D%3D%3D%200%0A%09%09%09buffer.stroke%28%22blue%22%29%3B%0A%09%09%09buffer.line%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20-%20PIXEL_SIZE%29%3B%0A%09%09%09buffer.line%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20%2B%20PIXEL_SIZE%29%3B%0A%09%09%7D%0A%09%7D%0A%7D%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20MAIN%20PROGRAM%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20p5.js%20setup%28%29%0A%2F%2F%0Aasync%20function%20setup%28%29%20%7B%20%2F%2Fnoprotect%20%28for%20large%20arrays%29%0A%09%2F%2F%20initiate%20console%20output%0A%09console.log%28%22%7C%7C%20%20%20%20%20%20%20%20%7C%7C%20%20%20%20%20%20Training%20%20%20%20%20%20%7C%7C%20%20%20%20%20Validation%20%20%20%20%20%7C%7C%22%29%3B%0A%09console.log%28%22%7C%7C%20Epoche%20%7C%7C%20korrekt%20%7C%20%20%20%20%20%20mse%20%7C%7C%20korrekt%20%7C%20%20%20%20%20%20mse%20%7C%7C%22%29%3B%0A%0A%09%2F%2F%20create%20canvas%0A%09createCanvas%282%20%2a%20DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09textSize%28PIXEL_SIZE%20%2a%203%29%3B%0A%0A%09%2F%2F%20create%20all%20grafic%20buffers%0A%09trainBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09validBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09decisionBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%0A%09%2F%2F%20draw%20initial%20training%20and%20validation%20buffer%0A%09drawDataInBuffer%28trainInputArray%2C%20trainOutputArray%2C%20trainBuffer%29%3B%0A%09drawDataInBuffer%28validInputArray%2C%20validOutputArray%2C%20validBuffer%29%3B%0A%09image%28trainBuffer%2C%200%2C%200%29%3B%0A%09image%28validBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20data%20preprocessing%0A%09generateTensors%28%29%3B%0A%0A%09%2F%2F%20define%20model%20and%20training%0A%09compileModel%28%29%3B%0A%0A%09%2F%2F%20train%20and%20visualise%20neural%20network%0A%09await%20trainModel%28%29%3B%0A%0A%09%2F%2F%20end%20of%20training%0A%09console.log%28%22Das%20Training%20ist%20beendet%20%3A-%29%22%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20p5.js%20keyPressed%28%29%20ESC%20%3D%20emergency%20stop%0A%2F%2F%0Afunction%20keyPressed%28%29%20%7B%0A%09if%20%28keyCode%20%3D%3D%3D%2027%29%20%7B%20%2F%2F%20ESC%20key%0A%09%09%2F%2F%20initiate%20stop%20of%20training%20after%20next%20epoch%0A%09%09model.stopTraining%20%3D%20true%3B%0A%09%09%2F%2F%20response%20to%20user%0A%09%09console.log%28%22Training%20wurde%20Abgebrochen%21%20Finales%20Resultat%3A%22%29%3B%0A%09%09%2F%2F%20draw%20final%20Prediction%0A%09%09drawPrediction%28actualEpoch%2C%20actualLogs.loss%29%3B%0A%09%7D%0A%7D&css=undefined&html=%3Cscript%20src%3D%22https%3A%2F%2Fapp.exorciser.ch%2Flib%2Fp5.js%22%3E%3C%2Fscript%3E%0A%3Cscript%20src%3D%22https%3A%2F%2Fcdn.jsdelivr.net%2Fnpm%2F%40tensorflow%2Ftfjs%40latest%2Fdist%2Ftf.min.js%22%3E%3C%2Fscript%3E&autorun=off&height=400px#Fehlende_Daten}} | {{exorciser/jspg?javascript=%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%2F%2F%20Diese%20Trainings-%20und%20Updateparameter%20d%C3%BCrfen%20ge%C3%A4ndert%20werden.%0Aconst%20LERNRATE%20%3D%201%3B%20%2F%2F%20Dezimalzahl%20%3E%200%0Aconst%20ANZAHL_HIDDEN_NEURONS%20%3D%203%3B%20%2F%2F%20ganzzahliger%20Wert%20%3E%200%0Aconst%20ANZAHL_EPOCHEN%20%3D%201000%3B%20%2F%2F%20ganzzahliger%20Wert%20%3E%200%0Aconst%20EPOCHEN_FUER_GRAFIK_UPDATE%20%3D%20100%3B%20%2F%2F%20ganzzahlig%20Wert%20%3E%200%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20CONSTANTS%20AND%20GLOBAL%20VARIABLES%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%20constants%0Aconst%20DATA_RANGE%20%3D%2010%3B%0Aconst%20DATA_WINDOW_SIZE%20%3D%20200%3B%0Aconst%20DATA_SCALE%20%3D%20Math.round%28DATA_WINDOW_SIZE%20%2F%20DATA_RANGE%29%3B%0Aconst%20NUM_DATA_PIXEL%20%3D%2050%3B%0Aconst%20PIXEL_SIZE%20%3D%20Math.round%28DATA_WINDOW_SIZE%20%2F%20NUM_DATA_PIXEL%29%3B%0A%0A%2F%2F%20training%20data%0Aconst%20trainInputArray%20%3D%20%5B%0A%09%5B2.5%2C%206.0%5D%2C%0A%09%5B3.5%2C%201.5%5D%2C%0A%09%5B3.5%2C%204.0%5D%2C%0A%09%5B4.0%2C%206.0%5D%2C%0A%09%5B4.0%2C%207.0%5D%2C%0A%09%5B1.5%2C%203.0%5D%2C%0A%09%5B1.5%2C%204.5%5D%2C%0A%09%5B1.5%2C%207.0%5D%2C%0A%09%5B2.0%2C%201.5%5D%2C%0A%09%5B3.0%2C%203.0%5D%2C%0A%09%5B3.0%2C%205.5%5D%2C%0A%09%5B3.0%2C%206.5%5D%2C%0A%09%5B4.0%2C%208.5%5D%2C%0A%5D%3B%0A%0Aconst%20trainOutputArray%20%3D%20%5B0%2C%200%2C%200%2C%200%2C%200%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%5D%3B%0Aconst%20trainInputMax%20%3D%20%5B4.0%2C%208.5%5D%3B%0Aconst%20trainInputMin%20%3D%20%5B1.5%2C%201.5%5D%3B%0A%0A%2F%2F%20validation%20data%0Aconst%20validInputArray%20%3D%20%5B%0A%09%5B2.5%2C%207.0%5D%2C%0A%09%5B3.5%2C%201.0%5D%2C%0A%09%5B3.5%2C%207.0%5D%2C%0A%09%5B4.0%2C%203.5%5D%2C%0A%09%5B4.5%2C%206.5%5D%2C%0A%09%5B4.5%2C%202.5%5D%2C%0A%09%5B4.5%2C%205.0%5D%2C%0A%09%5B6.0%2C%202.5%5D%2C%0A%09%5B6.0%2C%204.5%5D%2C%0A%09%5B6.5%2C%205.5%5D%2C%0A%09%5B7.5%2C%204.5%5D%2C%0A%09%5B8.5%2C%205.0%5D%2C%0A%09%5B1.5%2C%205.0%5D%2C%0A%09%5B1.5%2C%206.5%5D%2C%0A%09%5B2.0%2C%201.5%5D%2C%0A%09%5B2.5%2C%203.5%5D%2C%0A%09%5B2.5%2C%206.0%5D%2C%0A%09%5B3.5%2C%206.0%5D%2C%0A%09%5B4.0%2C%208.0%5D%2C%0A%09%5B4.0%2C%207.0%5D%2C%0A%09%5B5.5%2C%208.5%5D%2C%0A%09%5B6.5%2C%207.0%5D%2C%0A%09%5B7.5%2C%208.5%5D%2C%0A%09%5B8.5%2C%207.0%5D%0A%5D%3B%0Aconst%20validOutputArray%20%3D%20%5B0%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%200%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%5D%3B%0A%0A%2F%2F%20global%20variables%0Alet%20model%2C%20trainInput%2C%20trainOutput%2C%20validInput%2C%20validOutput%2C%20testInput%2C%20actualEpoch%2C%20actualLogs%2C%20trainBuffer%2C%20validBuffer%2C%20decisionBuffer%3B%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20NEURAL%20NETWORK%20FUNCTIONS%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20generate%20data%20tensors%0A%2F%2F%0Afunction%20generateTensors%28%29%20%7B%0A%09let%20normalizedTrainInputArray%20%3D%20%5B%5D%3B%0A%09let%20normalizedValidInputArray%20%3D%20%5B%5D%3B%0A%09let%20normalizedTestInputArray%20%3D%20%5B%5D%3B%0A%0A%09%2F%2F%20normalize%20training%20and%20validation%20inputs%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20trainInputArray.length%3B%20i%2B%2B%29%20%7B%0A%09%09normalizedTrainInputArray.push%28%5B%28trainInputArray%5Bi%5D%5B0%5D%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28trainInputArray%5Bi%5D%5B1%5D%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%7D%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20validInputArray.length%3B%20i%2B%2B%29%20%7B%0A%09%09normalizedValidInputArray.push%28%5B%28validInputArray%5Bi%5D%5B0%5D%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28validInputArray%5Bi%5D%5B1%5D%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%7D%0A%0A%09%2F%2F%20generate%20training%20input%20and%20output%20tensors%0A%09trainInput%20%3D%20tf.tensor2d%28normalizedTrainInputArray%2C%20%5BnormalizedTrainInputArray.length%2C%202%5D%29%3B%0A%09trainOutput%20%3D%20tf.tensor2d%28trainOutputArray%2C%20%5BtrainOutputArray.length%2C%201%5D%29%3B%0A%09%2F%2F%20let%20trainOutput%20%3D%20tf.oneHot%28tf.tensor1d%28trainOutputArray%29.toInt%28%29%2C%202%29%3B%0A%09%2F%2F%20trainInput.print%28%29%3B%20%2F%2F%20Print%20Tensor%0A%0A%09%2F%2F%20generate%20validation%20input%20and%20output%20tensors%0A%09validInput%20%3D%20tf.tensor2d%28normalizedValidInputArray%2C%20%5BnormalizedValidInputArray.length%2C%202%5D%29%3B%0A%09validOutput%20%3D%20tf.tensor2d%28validOutputArray%2C%20%5BvalidOutputArray.length%2C%201%5D%29%3B%0A%09%2F%2F%20let%20validOutput%20%3D%20tf.oneHot%28tf.tensor1d%28validOutputArray%29.toInt%28%29%2C%202%29%3B%0A%0A%09%2F%2F%20calculate%20normalized%20testing%20inputs%20for%20visualisation%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20i%20%3D%20i%20%2B%201%29%20%7B%0A%09%09for%20%28j%20%3D%200%3B%20j%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20j%20%3D%20j%20%2B%201%29%20%7B%0A%09%09%09normalizedTestInputArray.push%28%5B%28i%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28j%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%09%7D%0A%09%7D%0A%0A%09%2F%2F%20generate%20testing%20input%20tensor%20for%20visualisation%0A%09testInput%20%3D%20tf.tensor2d%28normalizedTestInputArray%2C%20%5BnormalizedTestInputArray.length%2C%202%5D%29%3B%0A%09%2F%2F%09testInput.print%28%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20build%20neural%20network%20model%20and%20define%20training%0A%2F%2F%0Afunction%20compileModel%28%29%20%7B%0A%09%2F%2F%20neural%20network%20structure%0A%09model%20%3D%20tf.sequential%28%7B%0A%09%09layers%3A%20%5B%0A%09%09%09tf.layers.dense%28%7B%0A%09%09%09%09name%3A%20%27HiddenLayer1%27%2C%0A%09%09%09%09inputShape%3A%20%5B2%5D%2C%0A%09%09%09%09units%3A%20ANZAHL_HIDDEN_NEURONS%2C%0A%09%09%09%09activation%3A%20%27tanh%27%0A%09%09%09%7D%29%2C%0A%09%09%09tf.layers.dense%28%7B%0A%09%09%09%09name%3A%20%27OutputLayer%27%2C%0A%09%09%09%09units%3A%201%2C%0A%09%09%09%09activation%3A%20%27sigmoid%27%0A%09%09%09%7D%29%0A%09%09%5D%0A%09%7D%29%3B%0A%0A%09%2F%2F%20neural%20network%20training%0A%09model.compile%28%7B%0A%09%09optimizer%3A%20tf.train.sgd%28LERNRATE%29%2C%0A%09%09loss%3A%20tf.losses.meanSquaredError%2C%0A%09%09metrics%3A%20%5B%27mse%27%5D%2C%0A%09%7D%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20train%20neural%20network%0A%2F%2F%0Aasync%20function%20trainModel%28%29%20%7B%0A%09await%20model.fit%28%0A%09%09trainInput%2C%20trainOutput%2C%20%7B%0A%09%09%09epochs%3A%20ANZAHL_EPOCHEN%2C%0A%09%09%09shuffle%3A%20true%2C%0A%09%09%09callbacks%3A%20%5B%7B%0A%09%09%09%09onEpochEnd%3A%20async%20%28epoch%2C%20logs%29%20%3D%3E%20%7B%0A%09%09%09%09%09%2F%2F%20remember%20epoch%20number%20if%20interrupted%0A%09%09%09%09%09actualEpoch%20%3D%20epoch%20%2B%201%3B%0A%09%09%09%09%09actualLogs%20%3D%20logs%3B%0A%09%09%09%09%09%2F%2F%20initiate%20graphic%20update%0A%09%09%09%09%09if%20%28%28%28actualEpoch%29%20%25%20EPOCHEN_FUER_GRAFIK_UPDATE%20%3D%3D%3D%200%29%20%7C%7C%20%28epoch%20%3D%3D%3D%200%29%29%20%7B%0A%09%09%09%09%09%09drawPrediction%28actualEpoch%2C%20logs.loss%29%3B%0A%09%09%09%09%09%7D%0A%09%09%09%09%7D%0A%09%09%09%7D%2C%20%5D%0A%09%09%7D%0A%09%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20calculate%20mean%20squared%20error%0A%2F%2F%0Afunction%20calculateMSE%28predictedOutput%2C%20trueOutput%29%20%7B%0A%09let%20mse%20%3D%200.0%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20predictedOutput.length%3B%20i%2B%2B%29%20%7B%0A%09%09mse%20%3D%20mse%20%2B%20%28predictedOutput%5Bi%5D%20-%20trueOutput%5Bi%5D%29%20%2a%2a%202%3B%0A%09%7D%0A%09return%20mse%20%2F%20predictedOutput.length%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20calculate%20percentage%20of%20correct%20classifications%0A%2F%2F%0Afunction%20calculatePercentageCorrect%28predictedOutput%2C%20trueOutput%29%20%7B%0A%09let%20correct%20%3D%200%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20predictedOutput.length%3B%20i%2B%2B%29%20%7B%0A%09%09if%20%28%28trueOutput%5Bi%5D%20%3D%3D%3D%201%20%26%26%20predictedOutput%5Bi%5D%20%3E%3D%200.5%29%20%7C%7C%20%28trueOutput%5Bi%5D%20%3D%3D%3D%200%20%26%26%20predictedOutput%5Bi%5D%20%3C%200.5%29%29%20%7B%0A%09%09%09correct%20%3D%20correct%20%2B%201%3B%0A%09%09%7D%0A%09%7D%0A%09return%20%7B%0A%09%09accuracy%3A%20Math.round%28correct%20%2a%20100%20%2F%20predictedOutput.length%29%2C%0A%09%09wrong%3A%20predictedOutput.length%20-%20correct%0A%09%7D%3B%0A%7D%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20FUNCTIONS%20FOR%20VISUALISATION%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20decision%20boundary%20and%20network%20output%0A%2F%2F%0Afunction%20drawPrediction%28epoch%2C%20mseTrain%29%20%7B%0A%09%2F%2F%20make%20predictions%20with%20neural%20network%0A%09let%20training%20%3D%20model.predict%28trainInput%29.arraySync%28%29%3B%0A%09let%20validation%20%3D%20model.predict%28validInput%29.arraySync%28%29%3B%0A%09let%20prediction%20%3D%20model.predict%28testInput%29.arraySync%28%29%3B%0A%0A%09%2F%2F%20calculate%20mse%20and%20percentages%20of%20correct%20classifications%0A%09let%20mseValid%20%3D%20calculateMSE%28validation%2C%20validOutputArray%29%3B%0A%09let%20trainCorrect%20%3D%20calculatePercentageCorrect%28training%2C%20trainOutputArray%29%3B%0A%09let%20validCorrect%20%3D%20calculatePercentageCorrect%28validation%2C%20validOutputArray%29%3B%0A%0A%09%2F%2F%20draw%20decision%20boundary%0A%09decisionBuffer.noStroke%28%29%3B%0A%09let%20element%20%3D%200%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20i%20%3D%20i%20%2B%201%29%20%7B%0A%09%09for%20%28j%20%3D%200%3B%20j%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20j%20%3D%20j%20%2B%201%29%20%7B%0A%09%09%09%2F%2F%20transform%20neural%20network%20output%20into%20a%20color%0A%09%09%09if%20%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28Math.round%28%281%20-%20prediction%5Belement%5D%5B0%5D%29%20%2a%20100%20%2B%20200.5%29%2C%20255%2C%20Math.round%28%281%20-%20prediction%5Belement%5D%5B0%5D%29%20%2a%20100%20%2B%20200.5%29%29%3B%0A%09%09%09%7D%20else%20%7B%0A%09%09%09%09decisionBuffer.fill%28Math.round%28prediction%5Belement%5D%5B0%5D%20%2a%20100%20%2B%20200.5%29%2C%20Math.round%28prediction%5Belement%5D%5B0%5D%20%2a%20100%20%2B%20200.5%29%2C%20255%29%3B%0A%09%09%09%7D%0A%09%09%09%2F%2F%20check%20for%20decision%20boundary%20and%20alter%20color%0A%09%09%09if%20%28%28element%20%25%20%28NUM_DATA_PIXEL%20%2B%201%29%29%20%21%3D%3D%200%20%26%26%20%28%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%20%26%26%20prediction%5Belement%20-%201%5D%5B0%5D%20%3C%3D%200.5%29%20%7C%7C%20%28prediction%5Belement%5D%5B0%5D%20%3C%3D%200.5%20%26%26%20prediction%5Belement%20-%201%5D%5B0%5D%20%3E%200.5%29%29%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28%27red%27%29%3B%0A%09%09%09%7D%0A%09%09%09if%20%28element%20%3E%3D%20NUM_DATA_PIXEL%20%2B%201%20%26%26%20%28%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%20%26%26%20prediction%5Belement%20-%20%28NUM_DATA_PIXEL%20%2B%201%29%5D%5B0%5D%20%3C%3D%200.5%29%20%7C%7C%20%28prediction%5Belement%5D%5B0%5D%20%3C%3D%200.5%20%26%26%20prediction%5Belement%20-%20%28NUM_DATA_PIXEL%20%2B%201%29%5D%5B0%5D%20%3E%200.5%29%29%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28%27red%27%29%3B%0A%09%09%09%7D%0A%09%09%09decisionBuffer.rect%28DATA_SCALE%20%2a%20i%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20PIXEL_SIZE%20%2F%202%2C%20DATA_SCALE%20%2a%20%2810%20-%20j%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%29%20-%20PIXEL_SIZE%20%2F%202%2C%20PIXEL_SIZE%2C%20PIXEL_SIZE%29%3B%0A%09%09%09element%20%3D%20element%20%2B%201%3B%0A%09%09%7D%0A%09%7D%0A%09image%28decisionBuffer%2C%200%2C%200%29%3B%0A%09image%28decisionBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20draw%20training%20and%20validation%20samples%0A%09image%28trainBuffer%2C%200%2C%200%29%3B%0A%09image%28validBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20add%20labels%20to%20training%20and%20validation%20feature%20space%20%2F%2F%20wrong%0A%09noStroke%28%29%3B%0A%09text%28%22Training%3A%20%22%20%2B%20trainCorrect.accuracy%20%2B%20%22%25%20%E2%9C%94%EF%B8%8F%20%28%22%20%2B%20trainCorrect.wrong%20%2B%20%22%20falsch%29%22%2C%202%20%2a%20PIXEL_SIZE%2C%204%20%2a%20PIXEL_SIZE%29%3B%0A%09text%28%22Validation%3A%20%22%20%2B%20validCorrect.accuracy%20%2B%20%22%25%20%E2%9C%94%EF%B8%8F%28%22%20%2B%20validCorrect.wrong%20%2B%20%22%20falsch%29%22%2C%20DATA_WINDOW_SIZE%20%2B%202%20%2a%20PIXEL_SIZE%2C%204%20%2a%20PIXEL_SIZE%29%3B%0A%0A%09%2F%2F%20print%20training%20and%20validation%20results%20to%20console%0A%09console.log%28%22%7C%7C%20%22%20%2B%20epoch.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%20%7C%7C%20%22%20%2B%20trainCorrect.accuracy.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%25%20%7C%20%22%20%2B%20mseTrain.toFixed%286%29%20%2B%20%22%20%7C%7C%20%22%20%2B%20validCorrect.accuracy.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%25%20%7C%20%22%20%2B%20mseValid.toFixed%286%29%20%2B%20%22%20%7C%7C%22%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20draw%20data%20in%20graphic%20buffer%0A%2F%2F%0Afunction%20drawDataInBuffer%28inputData%2C%20outputData%2C%20buffer%29%20%7B%0A%09buffer.noFill%28%29%3B%0A%09buffer.strokeWeight%28PIXEL_SIZE%20%2F%202%29%3B%0A%09buffer.rect%280%2C%200%2C%20DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20inputData.length%3B%20i%2B%2B%29%20%7B%0A%09%09if%20%28outputData%5Bi%5D%20%3D%3D%3D%201%29%20%7B%0A%09%09%09%2F%2F%20green%20circle%20for%20output%20label%20%3D%3D%3D%201%0A%09%09%09buffer.stroke%28%22darkgreen%22%29%3B%0A%09%09%09buffer.circle%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%2C%202%20%2a%20PIXEL_SIZE%29%3B%0A%09%09%7D%20else%20%7B%0A%09%09%09%2F%2F%20blue%20cross%20for%20output%20label%20%3D%3D%3D%200%0A%09%09%09buffer.stroke%28%22blue%22%29%3B%0A%09%09%09buffer.line%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20-%20PIXEL_SIZE%29%3B%0A%09%09%09buffer.line%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20%2B%20PIXEL_SIZE%29%3B%0A%09%09%7D%0A%09%7D%0A%7D%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20MAIN%20PROGRAM%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20p5.js%20setup%28%29%0A%2F%2F%0Aasync%20function%20setup%28%29%20%7B%20%2F%2Fnoprotect%20%28for%20large%20arrays%29%0A%09%2F%2F%20initiate%20console%20output%0A%09console.log%28%22%7C%7C%20%20%20%20%20%20%20%20%7C%7C%20%20%20%20%20%20Training%20%20%20%20%20%20%7C%7C%20%20%20%20%20Validation%20%20%20%20%20%7C%7C%22%29%3B%0A%09console.log%28%22%7C%7C%20Epoche%20%7C%7C%20korrekt%20%7C%20%20%20%20%20%20mse%20%7C%7C%20korrekt%20%7C%20%20%20%20%20%20mse%20%7C%7C%22%29%3B%0A%0A%09%2F%2F%20create%20canvas%0A%09createCanvas%282%20%2a%20DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09textSize%28PIXEL_SIZE%20%2a%203%29%3B%0A%0A%09%2F%2F%20create%20all%20grafic%20buffers%0A%09trainBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09validBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09decisionBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%0A%09%2F%2F%20draw%20initial%20training%20and%20validation%20buffer%0A%09drawDataInBuffer%28trainInputArray%2C%20trainOutputArray%2C%20trainBuffer%29%3B%0A%09drawDataInBuffer%28validInputArray%2C%20validOutputArray%2C%20validBuffer%29%3B%0A%09image%28trainBuffer%2C%200%2C%200%29%3B%0A%09image%28validBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20data%20preprocessing%0A%09generateTensors%28%29%3B%0A%0A%09%2F%2F%20define%20model%20and%20training%0A%09compileModel%28%29%3B%0A%0A%09%2F%2F%20train%20and%20visualise%20neural%20network%0A%09await%20trainModel%28%29%3B%0A%0A%09%2F%2F%20end%20of%20training%0A%09console.log%28%22Das%20Training%20ist%20beendet%20%3A-%29%22%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20p5.js%20keyPressed%28%29%20ESC%20%3D%20emergency%20stop%0A%2F%2F%0Afunction%20keyPressed%28%29%20%7B%0A%09if%20%28keyCode%20%3D%3D%3D%2027%29%20%7B%20%2F%2F%20ESC%20key%0A%09%09%2F%2F%20initiate%20stop%20of%20training%20after%20next%20epoch%0A%09%09model.stopTraining%20%3D%20true%3B%0A%09%09%2F%2F%20response%20to%20user%0A%09%09console.log%28%22Training%20wurde%20Abgebrochen%21%20Finales%20Resultat%3A%22%29%3B%0A%09%09%2F%2F%20draw%20final%20Prediction%0A%09%09drawPrediction%28actualEpoch%2C%20actualLogs.loss%29%3B%0A%09%7D%0A%7D&css=undefined&html=%3Cscript%20src%3D%22https%3A%2F%2Fapp.exorciser.ch%2Flib%2Fp5.js%22%3E%3C%2Fscript%3E%0A%3Cscript%20src%3D%22https%3A%2F%2Fcdn.jsdelivr.net%2Fnpm%2F%40tensorflow%2Ftfjs%40latest%2Fdist%2Ftf.min.js%22%3E%3C%2Fscript%3E&autorun=off&height=400px#Fehlende_Daten}} |
| </WRAP> | </WRAP> |
| | \\ |
| |
| ==== - Eingangsgrössen finden ==== | ==== - Eingangsgrössen finden ==== |
| [{{ :p:pasted:fischfeatures.png?249|Mögliche Eingangsgrössen Fisch-NN | <figure>{{:p:pasted:fischfeatures.png?249}}<caption>Mögliche Eingangsgrössen Fisch-NN |
| ((Eigene Darstellung [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]], der Fisch ist übernommen: Lodde [Fb78, Public Domain][[https://commons.wikimedia.org/wiki/File:Mallotus_villosus.gif|Mallotus villosus]] by | ((Eigene Darstellung [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]], der Fisch ist übernommen: Lodde [Fb78, Public Domain][[https://commons.wikimedia.org/wiki/File:Mallotus_villosus.gif|Mallotus villosus]] by |
| [[https://commons.wikimedia.org/wiki/User:Fb78|Fb78]] on wikimedia.))}}] | [[https://commons.wikimedia.org/wiki/User:Fb78|Fb78]] on wikimedia.))</caption></figure> |
| |
| ; 🐟 Fisch-NN\\ \\ | ; 🐟 Fisch-NN\\ \\ |
| - Der folgende Text besteht aus dem ersten Token im Vocabular von GPT-4o und dem letzten Token im Vocabular von GPT-4o. Gib den Text in Tokenizer für GPT-4o ein und lasse dir die Token-IDs anzeigen. Dann siehst du, wie gross das Vokabular von GPT-4o ist.\\ ''! cocos'' | - Der folgende Text besteht aus dem ersten Token im Vocabular von GPT-4o und dem letzten Token im Vocabular von GPT-4o. Gib den Text in Tokenizer für GPT-4o ein und lasse dir die Token-IDs anzeigen. Dann siehst du, wie gross das Vokabular von GPT-4o ist.\\ ''! cocos'' |
| </WRAP> | </WRAP> |
| | \\ |
| ==== - Fehlerfunktion bestimmen und interpretieren ==== | ==== - Fehlerfunktion bestimmen und interpretieren ==== |
| [{{ :p:pasted:fischnnresultat.png?200px|Fisch-NN Fischklassifikation ((Eigene Darstellung, [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]])) }}] | <figure right>{{:p:pasted:fischnnresultat.png?200px}}<caption>Fisch-NN Fischklassifikation ((Eigene Darstellung, [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]])) </caption></figure> |
| |
| ; 🐟 Fisch-NN\\ \\ | ; 🐟 Fisch-NN\\ \\ |
| ; 💬 KI-Chatbot\\ \\ | ; 💬 KI-Chatbot\\ \\ |
| : Bei einem KI-Chatbot wird beurteilt, wie gut dieser Texte verstehen und Aufgaben richtig lösen kann. Trainiert wird dieser jedoch hauptsächlich mit anderen Fehlerfunktionen.\\ \\ | : Bei einem KI-Chatbot wird beurteilt, wie gut dieser Texte verstehen und Aufgaben richtig lösen kann. Trainiert wird dieser jedoch hauptsächlich mit anderen Fehlerfunktionen.\\ \\ |
| : Die grossen KI-Chatbot-Hersteller testen und vergleichen ihre und andere KI-Chatbots mithilfe von verschiedenen "Benchmarks" d.h. Sammlungen von Aufgaben, welche ein KI-Chatbot lösen muss. Es gibt Benchmarks, die aus Multiple-Choice Aufgaben aus verschiedenen Wissenschaften bestehen (z.B. der [[https://huggingface.co/datasets/cais/mmlu|MMLU-Benchmark]]), solche, die Text-Aufgaben aus der Mathematik umfassen (z.B. der [[https://huggingface.co/datasets/gsm8k|GSM8K-Benchmark]]) oder andere, die das Textverständnis verschiedener Texte prüfen (z.B. der [[https://leaderboard.allenai.org/drop/submissions/about|DROP-Benchmark]]). Da diese Benchmarks jedoch mittlerweile fast schon zu einfach sind, für die immer besser werdenden KIs, gilt (aktuell) das [[https://agi.safe.ai/|Humanity's Last Exam]] als ultimative Herausforderung. Dieser Benchmark besteht ausschliesslich aus Fragen, an welchen sich auch menschliche Experten schnell einmal die Zähne ausbeissen.\\ \\ Beliebt ist auch das [[https://lmarena.ai?leaderboard|Chatbot Arena Leaderboard]], welche eine stets aktuelle "Hitparade" der KI-Chatbots aufgrund von Nutzerwertungen zeigt. Wenn du selber zum Leaderboard beitragen möchtest, gehe in die [[https://lmarena.ai/|Chatbot Arena]] und gib einen Prompt ein. Dieser wird dann von zwei zufällig gewählten KI-Chatbots beantwortet. Danach wählst aus, welche Antwort du besser findest. Am Ende wird aufgelöst, welche zwei KI-Chatbots du befragt hattest.\\ \\ | : Die grossen KI-Chatbot-Hersteller testen und vergleichen ihre und andere KI-Chatbots mithilfe von verschiedenen "Benchmarks" d.h. Sammlungen von Aufgaben, welche ein KI-Chatbot lösen muss. Es gibt Benchmarks, die aus Multiple-Choice Aufgaben aus verschiedenen Wissenschaften bestehen (z.B. der [[https://huggingface.co/datasets/cais/mmlu|MMLU-Benchmark]]), solche, die Text-Aufgaben aus der Mathematik umfassen (z.B. der [[https://huggingface.co/datasets/gsm8k|GSM8K-Benchmark]]) oder andere, die das Textverständnis verschiedener Texte prüfen (z.B. der [[https://leaderboard.allenai.org/drop/submissions/about|DROP-Benchmark]]). Da diese Benchmarks jedoch mittlerweile fast schon zu einfach sind, für die immer besser werdenden KIs, gilt (aktuell) das [[https://agi.safe.ai/|Humanity's Last Exam]] als ultimative Herausforderung. Dieser Benchmark besteht ausschliesslich aus Fragen, an welchen sich auch menschliche Experten schnell einmal die Zähne ausbeissen.\\ \\ 🤔 Das Problem mit diesen Benchmarks: Eine falsche Antwort oder die Antwort "keine Ahnung" ergeben Null Punkte. Daher ist es aussichtsreicher für die KI's zu raten, als zu sagen, "Keine Ahnung". Wer eine KI erstellen möchte, welche möglichst gut abschneidet, fördert wildes Raten (d.h. Halluzinieren) statt ehrlichen Antworten (z.B. "keine Ahnung").\\ \\ Beliebt ist auch das [[https://lmarena.ai?leaderboard|Chatbot Arena Leaderboard]], welche eine stets aktuelle "Hitparade" der KI-Chatbots aufgrund von Nutzerwertungen zeigt. Wenn du selber zum Leaderboard beitragen möchtest, gehe in die [[https://lmarena.ai/|Chatbot Arena]] und gib einen Prompt ein. Dieser wird dann von zwei zufällig gewählten KI-Chatbots beantwortet. Danach wählst aus, welche Antwort du besser findest. Am Ende wird aufgelöst, welche zwei KI-Chatbots du befragt hattest.\\ \\ 🤔 Das Problem mit diesen Benchmarks: Je lieber, wohlwollender oder gar lobend die Antworten sind, desto eher gefallen sie den Personen, welche für die Chatbot Arena abstimmen. Wer eine KI erstellen möchte, welche möglichst gut abschneidet, fördert Lob und unkritische Rückmeldungen (auch wenn es vielleicht nicht gerechtfertigt ist).\\ \\ |
| : ++Details zu den verwendeten Fehlerfunktionen (bei Interesse anklicken)|\\ \\ Trainiert werden die KI-Chatbots mit verschiedenen Fehlerfunktionen. Im Pre-Training geht es darum, das nächsten Token in einem Text vorherzusagen. Dazu wird für vom zugrundeliegenden neuronalen Netz für jedes Token im Token-Wörterbuch eine Zahl ausgegeben, die sogenannte Auswahl-Wahrscheinlichkeit. Die verwendete Fehlerfunktion ist so konstruiert, dass das neuronale Netz lernt, dem tatsächlich als Nächstes im Text vorkommenden Token eine möglichst grosse Auswahl-Wahrscheinlichkeit zu geben und alle anderen Tokens eine möglichst kleine. Dieses Vorgehen führt in der Praxis jedoch noch nicht zu den gewünschten Textantworten. Im darauf folgenden Fine-Tuning werden komplette, von KI-Chatbot erzeugte Textantworten mit einer zweiten KI beurteilt und der KI-Chatbot mithilfe einer komplexen Fehlerfunktion so nachtrainiert, dass dieser Texte mit einer möglichst hohen Beurteilung produziert. Obwohl diese Form von Fine-Tuning der Beurteilung mit Benchmarks schon relativ nahe kommt, kann damit das Pre-Training nicht ersetzt werden. Das scheitert nur schon daran, dass dafür ungleich mehr handverlesene Texte benötigt werden würden, als irgendwie zur Verfügung gestellt werden könnten.++ | : ++Details zu den verwendeten Fehlerfunktionen (bei Interesse anklicken)|\\ \\ Trainiert werden die KI-Chatbots mit verschiedenen Fehlerfunktionen. Im Pre-Training geht es darum, das nächsten Token in einem Text vorherzusagen. Dazu wird für vom zugrundeliegenden neuronalen Netz für jedes Token im Token-Wörterbuch eine Zahl ausgegeben, die sogenannte Auswahl-Wahrscheinlichkeit. Die verwendete Fehlerfunktion ist so konstruiert, dass das neuronale Netz lernt, dem tatsächlich als Nächstes im Text vorkommenden Token eine möglichst grosse Auswahl-Wahrscheinlichkeit zu geben und alle anderen Tokens eine möglichst kleine. Dieses Vorgehen führt in der Praxis jedoch noch nicht zu den gewünschten Textantworten. Im darauf folgenden Fine-Tuning werden komplette, von KI-Chatbot erzeugte Textantworten mit einer zweiten KI beurteilt und der KI-Chatbot mithilfe einer komplexen Fehlerfunktion so nachtrainiert, dass dieser Texte mit einer möglichst hohen Beurteilung produziert. Obwohl diese Form von Fine-Tuning der Beurteilung mit Benchmarks schon relativ nahe kommt, kann damit das Pre-Training nicht ersetzt werden. Das scheitert nur schon daran, dass dafür ungleich mehr handverlesene Texte benötigt werden würden, als irgendwie zur Verfügung gestellt werden könnten.++ |
| |
| - Trage selbst zur Chatbot Arena bei, indem du [[https://lmarena.ai|hier (du musst etwas nach unten scrollen zum Eingabefenster)]] zwei zufälligen KI-Chatbots eine Frage stellst und die bessere der beiden Antworten auswählst. | - Trage selbst zur Chatbot Arena bei, indem du [[https://lmarena.ai|hier (du musst etwas nach unten scrollen zum Eingabefenster)]] zwei zufälligen KI-Chatbots eine Frage stellst und die bessere der beiden Antworten auswählst. |
| </WRAP> | </WRAP> |
| | \\ |
| ==== - Modell wählen und trainieren ==== | ==== - Modell wählen und trainieren ==== |
| [{{ :p:pasted:fischnnsmall.png?240px|FischNN ((Eigene Darstellung, [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]])) }}] | <figure right>{{:p:pasted:fischnnsmall.png?240px}}<caption>FischNN ((Eigene Darstellung, [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]])) </caption></figure> |
| |
| ; 🐟 Fisch-NN\\ \\ | ; 🐟 Fisch-NN\\ \\ |
| 💡 Es scheint Limitationen für KIs zu geben, welche nicht von der Anzahl der verfügbaren Daten, Rechenpower oder der verfügbaren Zeit abhängen. In den folgenden beiden Unterkapiteln wird auf zwei davon genauer eingegangen. | 💡 Es scheint Limitationen für KIs zu geben, welche nicht von der Anzahl der verfügbaren Daten, Rechenpower oder der verfügbaren Zeit abhängen. In den folgenden beiden Unterkapiteln wird auf zwei davon genauer eingegangen. |
| |
| | \\ |
| ==== - Nicht Fehlerfrei ==== | ==== - Nicht Fehlerfrei ==== |
| [{{ p:pasted:classoverlap.png?185px|Überlappende Klassen((eigene Darstellung, [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]])) }}] | <figure right>{{p:pasted:classoverlap.png?185px}}<caption>Überlappende Klassen((eigene Darstellung, [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]])) </caption></figure> |
| |
| ; 🤖 KI im Allgemeinen\\ \\ | ; 🤖 KI im Allgemeinen\\ \\ |
| |
| ; 💬 KI-Chatbot\\ \\ | ; 💬 KI-Chatbot\\ \\ |
| : Texte zu Themen, zu welchen verschiedene Meinungen vorherrschen (z.B. zu "KI Fluch oder Segen"), könnten als "überlappende Muster" und somit als Herausforderung für das Training von KI-Chatbots angesehen werden. Es kann durchaus sein, dass seltene Aussagen oder Meinungen im Training unter gehen und somit nicht von einem KI-Chatbot ausgegeben werden können. | : Texte zu Themen, zu welchen verschiedene Meinungen vorherrschen (z.B. zu "KI Fluch oder Segen"), könnten als "überlappende Muster" und somit als Herausforderung für das Training von KI-Chatbots angesehen werden. Es kann durchaus sein, dass seltene Aussagen oder Meinungen im Training unter gehen und somit nicht von einem KI-Chatbot ausgegeben werden können.\\ \\ Weiter wird versucht in den gängigen Benchmarks und der Chatobt Arena möglichst gut abzuschneiden, um sich von der Konkrrenz abzuheben. Das führt einerseits dazu, dass im Training wildes Raten als wertvoller bewertet wird, als die Antwort "keine Ahnung" und somit Halluzinieren, d.h. Falschaussagen gefördert werden. Und andererseits, dass die KI-Chatbots dahin getrimmt werden eher zu unkritisch und zu lobend zu antworten, was ebenfalls eine Fehlerquelle darstellen kann. |
| |
| <WRAP center round box > | <WRAP center round box > |
| {{exorciser/jspg?javascript=%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%2F%2F%20Diese%20Trainings-%20und%20Updateparameter%20d%C3%BCrfen%20ge%C3%A4ndert%20werden.%0Aconst%20LERNRATE%20%3D%201%3B%20%2F%2F%20Dezimalzahl%20%3E%200%0Aconst%20ANZAHL_HIDDEN_NEURONS%20%3D%203%3B%20%2F%2F%20ganzzahliger%20Wert%20%3E%200%0Aconst%20ANZAHL_EPOCHEN%20%3D%201000%3B%20%2F%2F%20ganzzahliger%20Wert%20%3E%200%0Aconst%20EPOCHEN_FUER_GRAFIK_UPDATE%20%3D%20100%3B%20%2F%2F%20ganzzahlig%20Wert%20%3E%200%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20CONSTANTS%20AND%20GLOBAL%20VARIABLES%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%20constants%0Aconst%20DATA_RANGE%20%3D%2010%3B%0Aconst%20DATA_WINDOW_SIZE%20%3D%20200%3B%0Aconst%20DATA_SCALE%20%3D%20Math.round%28DATA_WINDOW_SIZE%20%2F%20DATA_RANGE%29%3B%0Aconst%20NUM_DATA_PIXEL%20%3D%2050%3B%0Aconst%20PIXEL_SIZE%20%3D%20Math.round%28DATA_WINDOW_SIZE%20%2F%20NUM_DATA_PIXEL%29%3B%0A%0A%2F%2F%20training%20data%0Aconst%20trainInputArray%20%3D%20%5B%0A%09%5B2.5%2C%206.0%5D%2C%0A%09%5B1.5%2C%203.0%5D%2C%0A%09%5B1.5%2C%204.5%5D%2C%0A%09%5B1.5%2C%207.0%5D%2C%0A%09%5B2.0%2C%201.5%5D%2C%0A%09%5B3.0%2C%203.0%5D%2C%0A%09%5B3.0%2C%205.5%5D%2C%0A%09%5B3.0%2C%206.5%5D%2C%0A%09%5B4.0%2C%208.5%5D%2C%0A%09%5B4.5%2C%206.5%5D%2C%0A%09%5B6.5%2C%207.5%5D%2C%0A%09%5B8.0%2C%206.5%5D%2C%0A%09%5B8.0%2C%208.5%5D%0A%5D%3B%0Aconst%20trainOutputArray%20%3D%20%5B0%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%5D%3B%0Aconst%20trainInputMax%20%3D%20%5B9.0%2C%208.5%5D%3B%0Aconst%20trainInputMin%20%3D%20%5B1.5%2C%201.5%5D%3B%0A%0A%2F%2F%20validation%20data%0Aconst%20validInputArray%20%3D%20%5B%0A%09%5B2.5%2C%207.0%5D%2C%0A%09%5B1.5%2C%205.0%5D%2C%0A%09%5B1.5%2C%206.5%5D%2C%0A%09%5B2.0%2C%201.5%5D%2C%0A%09%5B2.5%2C%203.5%5D%2C%0A%09%5B2.5%2C%206.0%5D%2C%0A%09%5B3.5%2C%206.0%5D%2C%0A%09%5B4.0%2C%208.0%5D%2C%0A%09%5B4.0%2C%207.0%5D%2C%0A%09%5B5.5%2C%208.5%5D%2C%0A%09%5B6.5%2C%207.0%5D%2C%0A%09%5B7.5%2C%208.5%5D%2C%0A%09%5B8.5%2C%207.0%5D%0A%5D%3B%0Aconst%20validOutputArray%20%3D%20%5B0%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%5D%3B%0A%0A%2F%2F%20global%20variables%0Alet%20model%2C%20trainInput%2C%20trainOutput%2C%20validInput%2C%20validOutput%2C%20testInput%2C%20actualEpoch%2C%20actualLogs%2C%20trainBuffer%2C%20validBuffer%2C%20decisionBuffer%3B%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20NEURAL%20NETWORK%20FUNCTIONS%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20generate%20data%20tensors%0A%2F%2F%0Afunction%20generateTensors%28%29%20%7B%0A%09let%20normalizedTrainInputArray%20%3D%20%5B%5D%3B%0A%09let%20normalizedValidInputArray%20%3D%20%5B%5D%3B%0A%09let%20normalizedTestInputArray%20%3D%20%5B%5D%3B%0A%0A%09%2F%2F%20normalize%20training%20and%20validation%20inputs%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20trainInputArray.length%3B%20i%2B%2B%29%20%7B%0A%09%09normalizedTrainInputArray.push%28%5B%28trainInputArray%5Bi%5D%5B0%5D%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28trainInputArray%5Bi%5D%5B1%5D%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%09normalizedValidInputArray.push%28%5B%28validInputArray%5Bi%5D%5B0%5D%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28validInputArray%5Bi%5D%5B1%5D%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%7D%0A%0A%09%2F%2F%20generate%20training%20input%20and%20output%20tensors%0A%09trainInput%20%3D%20tf.tensor2d%28normalizedTrainInputArray%2C%20%5BnormalizedTrainInputArray.length%2C%202%5D%29%3B%0A%09trainOutput%20%3D%20tf.tensor2d%28trainOutputArray%2C%20%5BtrainOutputArray.length%2C%201%5D%29%3B%0A%09%2F%2F%20let%20trainOutput%20%3D%20tf.oneHot%28tf.tensor1d%28trainOutputArray%29.toInt%28%29%2C%202%29%3B%0A%09%2F%2F%20trainInput.print%28%29%3B%20%2F%2F%20Print%20Tensor%0A%0A%09%2F%2F%20generate%20validation%20input%20and%20output%20tensors%0A%09validInput%20%3D%20tf.tensor2d%28normalizedValidInputArray%2C%20%5BnormalizedValidInputArray.length%2C%202%5D%29%3B%0A%09validOutput%20%3D%20tf.tensor2d%28validOutputArray%2C%20%5BvalidOutputArray.length%2C%201%5D%29%3B%0A%09%2F%2F%20let%20validOutput%20%3D%20tf.oneHot%28tf.tensor1d%28validOutputArray%29.toInt%28%29%2C%202%29%3B%0A%0A%09%2F%2F%20calculate%20normalized%20testing%20inputs%20for%20visualisation%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20i%20%3D%20i%20%2B%201%29%20%7B%0A%09%09for%20%28j%20%3D%200%3B%20j%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20j%20%3D%20j%20%2B%201%29%20%7B%0A%09%09%09normalizedTestInputArray.push%28%5B%28i%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28j%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%09%7D%0A%09%7D%0A%0A%09%2F%2F%20generate%20testing%20input%20tensor%20for%20visualisation%0A%09testInput%20%3D%20tf.tensor2d%28normalizedTestInputArray%2C%20%5BnormalizedTestInputArray.length%2C%202%5D%29%3B%0A%09%2F%2F%09testInput.print%28%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20build%20neural%20network%20model%20and%20define%20training%0A%2F%2F%0Afunction%20compileModel%28%29%20%7B%0A%09%2F%2F%20neural%20network%20structure%0A%09model%20%3D%20tf.sequential%28%7B%0A%09%09layers%3A%20%5B%0A%09%09%09tf.layers.dense%28%7B%0A%09%09%09%09name%3A%20%27HiddenLayer1%27%2C%0A%09%09%09%09inputShape%3A%20%5B2%5D%2C%0A%09%09%09%09units%3A%20ANZAHL_HIDDEN_NEURONS%2C%0A%09%09%09%09activation%3A%20%27tanh%27%0A%09%09%09%7D%29%2C%0A%09%09%09tf.layers.dense%28%7B%0A%09%09%09%09name%3A%20%27OutputLayer%27%2C%0A%09%09%09%09units%3A%201%2C%0A%09%09%09%09activation%3A%20%27sigmoid%27%0A%09%09%09%7D%29%0A%09%09%5D%0A%09%7D%29%3B%0A%0A%09%2F%2F%20neural%20network%20training%0A%09model.compile%28%7B%0A%09%09optimizer%3A%20tf.train.sgd%28LERNRATE%29%2C%0A%09%09loss%3A%20tf.losses.meanSquaredError%2C%0A%09%09metrics%3A%20%5B%27mse%27%5D%2C%0A%09%7D%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20train%20neural%20network%0A%2F%2F%0Aasync%20function%20trainModel%28%29%20%7B%0A%09await%20model.fit%28%0A%09%09trainInput%2C%20trainOutput%2C%20%7B%0A%09%09%09epochs%3A%20ANZAHL_EPOCHEN%2C%0A%09%09%09shuffle%3A%20true%2C%0A%09%09%09callbacks%3A%20%5B%7B%0A%09%09%09%09onEpochEnd%3A%20async%28epoch%2C%20logs%29%20%3D%3E%20%7B%0A%09%09%09%09%09%2F%2F%20remember%20epoch%20number%20if%20interrupted%0A%09%09%09%09%09actualEpoch%20%3D%20epoch%20%2B%201%3B%0A%09%09%09%09%09actualLogs%20%3D%20logs%3B%0A%09%09%09%09%09%2F%2F%20initiate%20graphic%20update%0A%09%09%09%09%09if%20%28%28%28actualEpoch%29%20%25%20EPOCHEN_FUER_GRAFIK_UPDATE%20%3D%3D%3D%200%29%20%7C%7C%20%28epoch%20%3D%3D%3D%200%29%29%20%7B%0A%09%09%09%09%09%09drawPrediction%28actualEpoch%2C%20logs.loss%29%3B%0A%09%09%09%09%09%7D%0A%09%09%09%09%7D%0A%09%09%09%7D%2C%20%5D%0A%09%09%7D%0A%09%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20calculate%20mean%20squared%20error%0A%2F%2F%0Afunction%20calculateMSE%28predictedOutput%2C%20trueOutput%29%20%7B%0A%09let%20mse%20%3D%200.0%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20predictedOutput.length%3B%20i%2B%2B%29%20%7B%0A%09%09mse%20%3D%20mse%20%2B%20%28predictedOutput%5Bi%5D%20-%20trueOutput%5Bi%5D%29%20%2a%2a%202%3B%0A%09%7D%0A%09return%20mse%20%2F%20predictedOutput.length%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20calculate%20percentage%20of%20correct%20classifications%0A%2F%2F%0Afunction%20calculatePercentageCorrect%28predictedOutput%2C%20trueOutput%29%20%7B%0A%09let%20correct%20%3D%200%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20predictedOutput.length%3B%20i%2B%2B%29%20%7B%0A%09%09if%20%28%28trueOutput%5Bi%5D%20%3D%3D%3D%201%20%26%26%20predictedOutput%5Bi%5D%20%3E%3D%200.5%29%20%7C%7C%20%28trueOutput%5Bi%5D%20%3D%3D%3D%200%20%26%26%20predictedOutput%5Bi%5D%20%3C%200.5%29%29%20%7B%0A%09%09%09correct%20%3D%20correct%20%2B%201%3B%0A%09%09%7D%0A%09%7D%0A%09return%20%7B%0A%09%09accuracy%3A%20Math.round%28correct%20%2a%20100%20%2F%20predictedOutput.length%29%2C%0A%09%09wrong%3A%20predictedOutput.length%20-%20correct%0A%09%7D%3B%0A%7D%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20FUNCTIONS%20FOR%20VISUALISATION%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20decision%20boundary%20and%20network%20output%0A%2F%2F%0Afunction%20drawPrediction%28epoch%2C%20mseTrain%29%20%7B%0A%09%2F%2F%20make%20predictions%20with%20neural%20network%0A%09let%20training%20%3D%20model.predict%28trainInput%29.arraySync%28%29%3B%0A%09let%20validation%20%3D%20model.predict%28validInput%29.arraySync%28%29%3B%0A%09let%20prediction%20%3D%20model.predict%28testInput%29.arraySync%28%29%3B%0A%0A%09%2F%2F%20calculate%20mse%20and%20percentages%20of%20correct%20classifications%0A%09let%20mseValid%20%3D%20calculateMSE%28validation%2C%20validOutputArray%29%3B%0A%09let%20trainCorrect%20%3D%20calculatePercentageCorrect%28training%2C%20trainOutputArray%29%3B%0A%09let%20validCorrect%20%3D%20calculatePercentageCorrect%28validation%2C%20validOutputArray%29%3B%0A%0A%09%2F%2F%20draw%20decision%20boundary%0A%09decisionBuffer.noStroke%28%29%3B%0A%09let%20element%20%3D%200%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20i%20%3D%20i%20%2B%201%29%20%7B%0A%09%09for%20%28j%20%3D%200%3B%20j%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20j%20%3D%20j%20%2B%201%29%20%7B%0A%09%09%09%2F%2F%20transform%20neural%20network%20output%20into%20a%20color%0A%09%09%09if%20%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28Math.round%28%281%20-%20prediction%5Belement%5D%5B0%5D%29%20%2a%20100%20%2B%20200.5%29%2C%20255%2C%20Math.round%28%281%20-%20prediction%5Belement%5D%5B0%5D%29%20%2a%20100%20%2B%20200.5%29%29%3B%0A%09%09%09%7D%20else%20%7B%0A%09%09%09%09decisionBuffer.fill%28Math.round%28prediction%5Belement%5D%5B0%5D%20%2a%20100%20%2B%20200.5%29%2C%20Math.round%28prediction%5Belement%5D%5B0%5D%20%2a%20100%20%2B%20200.5%29%2C%20255%29%3B%0A%09%09%09%7D%0A%09%09%09%2F%2F%20check%20for%20decision%20boundary%20and%20alter%20color%0A%09%09%09if%20%28%28element%20%25%20%28NUM_DATA_PIXEL%20%2B%201%29%29%20%21%3D%3D%200%20%26%26%20%28%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%20%26%26%20prediction%5Belement%20-%201%5D%5B0%5D%20%3C%3D%200.5%29%20%7C%7C%20%28prediction%5Belement%5D%5B0%5D%20%3C%3D%200.5%20%26%26%20prediction%5Belement%20-%201%5D%5B0%5D%20%3E%200.5%29%29%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28%27red%27%29%3B%0A%09%09%09%7D%0A%09%09%09if%20%28element%20%3E%3D%20NUM_DATA_PIXEL%20%2B%201%20%26%26%20%28%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%20%26%26%20prediction%5Belement%20-%20%28NUM_DATA_PIXEL%20%2B%201%29%5D%5B0%5D%20%3C%3D%200.5%29%20%7C%7C%20%28prediction%5Belement%5D%5B0%5D%20%3C%3D%200.5%20%26%26%20prediction%5Belement%20-%20%28NUM_DATA_PIXEL%20%2B%201%29%5D%5B0%5D%20%3E%200.5%29%29%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28%27red%27%29%3B%0A%09%09%09%7D%0A%09%09%09decisionBuffer.rect%28DATA_SCALE%20%2a%20i%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20PIXEL_SIZE%20%2F%202%2C%20DATA_SCALE%20%2a%20%2810%20-%20j%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%29%20-%20PIXEL_SIZE%20%2F%202%2C%20PIXEL_SIZE%2C%20PIXEL_SIZE%29%3B%0A%09%09%09element%20%3D%20element%20%2B%201%3B%0A%09%09%7D%0A%09%7D%0A%09image%28decisionBuffer%2C%200%2C%200%29%3B%0A%09image%28decisionBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20draw%20training%20and%20validation%20samples%0A%09image%28trainBuffer%2C%200%2C%200%29%3B%0A%09image%28validBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20add%20labels%20to%20training%20and%20validation%20feature%20space%20%2F%2F%20wrong%0A%09noStroke%28%29%3B%0A%09text%28%22Training%3A%20%22%20%2B%20trainCorrect.accuracy%20%2B%20%22%25%20%E2%9C%94%EF%B8%8F%20%28%22%20%2B%20trainCorrect.wrong%20%2B%20%22%20falsch%29%22%2C%202%20%2a%20PIXEL_SIZE%2C%204%20%2a%20PIXEL_SIZE%29%3B%0A%09text%28%22Validation%3A%20%22%20%2B%20validCorrect.accuracy%20%2B%20%22%25%20%E2%9C%94%EF%B8%8F%28%22%20%2B%20validCorrect.wrong%20%2B%20%22%20falsch%29%22%2C%20DATA_WINDOW_SIZE%20%2B%202%20%2a%20PIXEL_SIZE%2C%204%20%2a%20PIXEL_SIZE%29%3B%0A%0A%09%2F%2F%20print%20training%20and%20validation%20results%20to%20console%0A%09console.log%28%22%7C%7C%20%22%20%2B%20epoch.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%20%7C%7C%20%22%20%2B%20trainCorrect.accuracy.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%25%20%7C%20%22%20%2B%20mseTrain.toFixed%286%29%20%2B%20%22%20%7C%7C%20%22%20%2B%20validCorrect.accuracy.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%25%20%7C%20%22%20%2B%20mseValid.toFixed%286%29%20%2B%20%22%20%7C%7C%22%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20draw%20data%20in%20graphic%20buffer%0A%2F%2F%0Afunction%20drawDataInBuffer%28inputData%2C%20outputData%2C%20buffer%29%20%7B%0A%09buffer.noFill%28%29%3B%0A%09buffer.strokeWeight%28PIXEL_SIZE%20%2F%202%29%3B%0A%09buffer.rect%280%2C%200%2C%20DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20inputData.length%3B%20i%2B%2B%29%20%7B%0A%09%09if%20%28outputData%5Bi%5D%20%3D%3D%3D%201%29%20%7B%0A%09%09%09%2F%2F%20green%20circle%20for%20output%20label%20%3D%3D%3D%201%0A%09%09%09buffer.stroke%28%22darkgreen%22%29%3B%0A%09%09%09buffer.circle%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%2C%202%20%2a%20PIXEL_SIZE%29%3B%0A%09%09%7D%20else%20%7B%0A%09%09%09%2F%2F%20blue%20cross%20for%20output%20label%20%3D%3D%3D%200%0A%09%09%09buffer.stroke%28%22blue%22%29%3B%0A%09%09%09buffer.line%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20-%20PIXEL_SIZE%29%3B%0A%09%09%09buffer.line%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20%2B%20PIXEL_SIZE%29%3B%0A%09%09%7D%0A%09%7D%0A%7D%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20MAIN%20PROGRAM%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20p5.js%20setup%28%29%0A%2F%2F%0Aasync%20function%20setup%28%29%20%7B%20%2F%2Fnoprotect%20%28for%20large%20arrays%29%0A%09%2F%2F%20initiate%20console%20output%0A%09console.log%28%22%7C%7C%20%20%20%20%20%20%20%20%7C%7C%20%20%20%20%20%20Training%20%20%20%20%20%20%7C%7C%20%20%20%20%20Validation%20%20%20%20%20%7C%7C%22%29%3B%0A%09console.log%28%22%7C%7C%20Epoche%20%7C%7C%20korrekt%20%7C%20%20%20%20%20%20mse%20%7C%7C%20korrekt%20%7C%20%20%20%20%20%20mse%20%7C%7C%22%29%3B%0A%0A%09%2F%2F%20create%20canvas%0A%09createCanvas%282%20%2a%20DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09textSize%28PIXEL_SIZE%20%2a%203%29%3B%0A%0A%09%2F%2F%20create%20all%20grafic%20buffers%0A%09trainBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09validBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09decisionBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%0A%09%2F%2F%20draw%20initial%20training%20and%20validation%20buffer%0A%09drawDataInBuffer%28trainInputArray%2C%20trainOutputArray%2C%20trainBuffer%29%3B%0A%09drawDataInBuffer%28validInputArray%2C%20validOutputArray%2C%20validBuffer%29%3B%0A%09image%28trainBuffer%2C%200%2C%200%29%3B%0A%09image%28validBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20data%20preprocessing%0A%09generateTensors%28%29%3B%0A%0A%09%2F%2F%20define%20model%20and%20training%0A%09compileModel%28%29%3B%0A%0A%09%2F%2F%20train%20and%20visualise%20neural%20network%0A%09await%20trainModel%28%29%3B%0A%0A%09%2F%2F%20end%20of%20training%0A%09console.log%28%22Das%20Training%20ist%20beendet%20%3A-%29%22%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20p5.js%20keyPressed%28%29%20ESC%20%3D%20emergency%20stop%0A%2F%2F%0Afunction%20keyPressed%28%29%20%7B%0A%09if%20%28keyCode%20%3D%3D%3D%2027%29%20%7B%20%2F%2F%20ESC%20key%0A%09%09%2F%2F%20initiate%20stop%20of%20training%20after%20next%20epoch%0A%09%09model.stopTraining%20%3D%20true%3B%0A%09%09%2F%2F%20response%20to%20user%0A%09%09console.log%28%22Training%20wurde%20Abgebrochen%21%20Finales%20Resultat%3A%22%29%3B%0A%09%09%2F%2F%20draw%20final%20Prediction%0A%09%09drawPrediction%28actualEpoch%2C%20actualLogs.loss%29%3B%0A%09%7D%0A%7D&css=undefined&html=%3Cscript%20src%3D%22https%3A%2F%2Fapp.exorciser.ch%2Flib%2Fp5.js%22%3E%3C%2Fscript%3E%0A%3Cscript%20src%3D%22https%3A%2F%2Fcdn.jsdelivr.net%2Fnpm%2F%40tensorflow%2Ftfjs%40latest%2Fdist%2Ftf.min.js%22%3E%3C%2Fscript%3E&autorun=off&height=400px#SelteneKlasse}} | {{exorciser/jspg?javascript=%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%2F%2F%20Diese%20Trainings-%20und%20Updateparameter%20d%C3%BCrfen%20ge%C3%A4ndert%20werden.%0Aconst%20LERNRATE%20%3D%201%3B%20%2F%2F%20Dezimalzahl%20%3E%200%0Aconst%20ANZAHL_HIDDEN_NEURONS%20%3D%203%3B%20%2F%2F%20ganzzahliger%20Wert%20%3E%200%0Aconst%20ANZAHL_EPOCHEN%20%3D%201000%3B%20%2F%2F%20ganzzahliger%20Wert%20%3E%200%0Aconst%20EPOCHEN_FUER_GRAFIK_UPDATE%20%3D%20100%3B%20%2F%2F%20ganzzahlig%20Wert%20%3E%200%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20CONSTANTS%20AND%20GLOBAL%20VARIABLES%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%20constants%0Aconst%20DATA_RANGE%20%3D%2010%3B%0Aconst%20DATA_WINDOW_SIZE%20%3D%20200%3B%0Aconst%20DATA_SCALE%20%3D%20Math.round%28DATA_WINDOW_SIZE%20%2F%20DATA_RANGE%29%3B%0Aconst%20NUM_DATA_PIXEL%20%3D%2050%3B%0Aconst%20PIXEL_SIZE%20%3D%20Math.round%28DATA_WINDOW_SIZE%20%2F%20NUM_DATA_PIXEL%29%3B%0A%0A%2F%2F%20training%20data%0Aconst%20trainInputArray%20%3D%20%5B%0A%09%5B2.5%2C%206.0%5D%2C%0A%09%5B1.5%2C%203.0%5D%2C%0A%09%5B1.5%2C%204.5%5D%2C%0A%09%5B1.5%2C%207.0%5D%2C%0A%09%5B2.0%2C%201.5%5D%2C%0A%09%5B3.0%2C%203.0%5D%2C%0A%09%5B3.0%2C%205.5%5D%2C%0A%09%5B3.0%2C%206.5%5D%2C%0A%09%5B4.0%2C%208.5%5D%2C%0A%09%5B4.5%2C%206.5%5D%2C%0A%09%5B6.5%2C%207.5%5D%2C%0A%09%5B8.0%2C%206.5%5D%2C%0A%09%5B8.0%2C%208.5%5D%0A%5D%3B%0Aconst%20trainOutputArray%20%3D%20%5B0%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%5D%3B%0Aconst%20trainInputMax%20%3D%20%5B9.0%2C%208.5%5D%3B%0Aconst%20trainInputMin%20%3D%20%5B1.5%2C%201.5%5D%3B%0A%0A%2F%2F%20validation%20data%0Aconst%20validInputArray%20%3D%20%5B%0A%09%5B2.5%2C%207.0%5D%2C%0A%09%5B1.5%2C%205.0%5D%2C%0A%09%5B1.5%2C%206.5%5D%2C%0A%09%5B2.0%2C%201.5%5D%2C%0A%09%5B2.5%2C%203.5%5D%2C%0A%09%5B2.5%2C%206.0%5D%2C%0A%09%5B3.5%2C%206.0%5D%2C%0A%09%5B4.0%2C%208.0%5D%2C%0A%09%5B4.0%2C%207.0%5D%2C%0A%09%5B5.5%2C%208.5%5D%2C%0A%09%5B6.5%2C%207.0%5D%2C%0A%09%5B7.5%2C%208.5%5D%2C%0A%09%5B8.5%2C%207.0%5D%0A%5D%3B%0Aconst%20validOutputArray%20%3D%20%5B0%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%2C%201%5D%3B%0A%0A%2F%2F%20global%20variables%0Alet%20model%2C%20trainInput%2C%20trainOutput%2C%20validInput%2C%20validOutput%2C%20testInput%2C%20actualEpoch%2C%20actualLogs%2C%20trainBuffer%2C%20validBuffer%2C%20decisionBuffer%3B%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20NEURAL%20NETWORK%20FUNCTIONS%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20generate%20data%20tensors%0A%2F%2F%0Afunction%20generateTensors%28%29%20%7B%0A%09let%20normalizedTrainInputArray%20%3D%20%5B%5D%3B%0A%09let%20normalizedValidInputArray%20%3D%20%5B%5D%3B%0A%09let%20normalizedTestInputArray%20%3D%20%5B%5D%3B%0A%0A%09%2F%2F%20normalize%20training%20and%20validation%20inputs%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20trainInputArray.length%3B%20i%2B%2B%29%20%7B%0A%09%09normalizedTrainInputArray.push%28%5B%28trainInputArray%5Bi%5D%5B0%5D%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28trainInputArray%5Bi%5D%5B1%5D%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%09normalizedValidInputArray.push%28%5B%28validInputArray%5Bi%5D%5B0%5D%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28validInputArray%5Bi%5D%5B1%5D%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%7D%0A%0A%09%2F%2F%20generate%20training%20input%20and%20output%20tensors%0A%09trainInput%20%3D%20tf.tensor2d%28normalizedTrainInputArray%2C%20%5BnormalizedTrainInputArray.length%2C%202%5D%29%3B%0A%09trainOutput%20%3D%20tf.tensor2d%28trainOutputArray%2C%20%5BtrainOutputArray.length%2C%201%5D%29%3B%0A%09%2F%2F%20let%20trainOutput%20%3D%20tf.oneHot%28tf.tensor1d%28trainOutputArray%29.toInt%28%29%2C%202%29%3B%0A%09%2F%2F%20trainInput.print%28%29%3B%20%2F%2F%20Print%20Tensor%0A%0A%09%2F%2F%20generate%20validation%20input%20and%20output%20tensors%0A%09validInput%20%3D%20tf.tensor2d%28normalizedValidInputArray%2C%20%5BnormalizedValidInputArray.length%2C%202%5D%29%3B%0A%09validOutput%20%3D%20tf.tensor2d%28validOutputArray%2C%20%5BvalidOutputArray.length%2C%201%5D%29%3B%0A%09%2F%2F%20let%20validOutput%20%3D%20tf.oneHot%28tf.tensor1d%28validOutputArray%29.toInt%28%29%2C%202%29%3B%0A%0A%09%2F%2F%20calculate%20normalized%20testing%20inputs%20for%20visualisation%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20i%20%3D%20i%20%2B%201%29%20%7B%0A%09%09for%20%28j%20%3D%200%3B%20j%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20j%20%3D%20j%20%2B%201%29%20%7B%0A%09%09%09normalizedTestInputArray.push%28%5B%28i%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20trainInputMin%5B0%5D%29%20%2F%20%28trainInputMax%5B0%5D%20-%20trainInputMin%5B0%5D%29%2C%20%28j%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20trainInputMin%5B1%5D%29%20%2F%20%28trainInputMax%5B1%5D%20-%20trainInputMin%5B1%5D%29%5D%29%3B%0A%09%09%7D%0A%09%7D%0A%0A%09%2F%2F%20generate%20testing%20input%20tensor%20for%20visualisation%0A%09testInput%20%3D%20tf.tensor2d%28normalizedTestInputArray%2C%20%5BnormalizedTestInputArray.length%2C%202%5D%29%3B%0A%09%2F%2F%09testInput.print%28%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20build%20neural%20network%20model%20and%20define%20training%0A%2F%2F%0Afunction%20compileModel%28%29%20%7B%0A%09%2F%2F%20neural%20network%20structure%0A%09model%20%3D%20tf.sequential%28%7B%0A%09%09layers%3A%20%5B%0A%09%09%09tf.layers.dense%28%7B%0A%09%09%09%09name%3A%20%27HiddenLayer1%27%2C%0A%09%09%09%09inputShape%3A%20%5B2%5D%2C%0A%09%09%09%09units%3A%20ANZAHL_HIDDEN_NEURONS%2C%0A%09%09%09%09activation%3A%20%27tanh%27%0A%09%09%09%7D%29%2C%0A%09%09%09tf.layers.dense%28%7B%0A%09%09%09%09name%3A%20%27OutputLayer%27%2C%0A%09%09%09%09units%3A%201%2C%0A%09%09%09%09activation%3A%20%27sigmoid%27%0A%09%09%09%7D%29%0A%09%09%5D%0A%09%7D%29%3B%0A%0A%09%2F%2F%20neural%20network%20training%0A%09model.compile%28%7B%0A%09%09optimizer%3A%20tf.train.sgd%28LERNRATE%29%2C%0A%09%09loss%3A%20tf.losses.meanSquaredError%2C%0A%09%09metrics%3A%20%5B%27mse%27%5D%2C%0A%09%7D%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20train%20neural%20network%0A%2F%2F%0Aasync%20function%20trainModel%28%29%20%7B%0A%09await%20model.fit%28%0A%09%09trainInput%2C%20trainOutput%2C%20%7B%0A%09%09%09epochs%3A%20ANZAHL_EPOCHEN%2C%0A%09%09%09shuffle%3A%20true%2C%0A%09%09%09callbacks%3A%20%5B%7B%0A%09%09%09%09onEpochEnd%3A%20async%28epoch%2C%20logs%29%20%3D%3E%20%7B%0A%09%09%09%09%09%2F%2F%20remember%20epoch%20number%20if%20interrupted%0A%09%09%09%09%09actualEpoch%20%3D%20epoch%20%2B%201%3B%0A%09%09%09%09%09actualLogs%20%3D%20logs%3B%0A%09%09%09%09%09%2F%2F%20initiate%20graphic%20update%0A%09%09%09%09%09if%20%28%28%28actualEpoch%29%20%25%20EPOCHEN_FUER_GRAFIK_UPDATE%20%3D%3D%3D%200%29%20%7C%7C%20%28epoch%20%3D%3D%3D%200%29%29%20%7B%0A%09%09%09%09%09%09drawPrediction%28actualEpoch%2C%20logs.loss%29%3B%0A%09%09%09%09%09%7D%0A%09%09%09%09%7D%0A%09%09%09%7D%2C%20%5D%0A%09%09%7D%0A%09%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20calculate%20mean%20squared%20error%0A%2F%2F%0Afunction%20calculateMSE%28predictedOutput%2C%20trueOutput%29%20%7B%0A%09let%20mse%20%3D%200.0%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20predictedOutput.length%3B%20i%2B%2B%29%20%7B%0A%09%09mse%20%3D%20mse%20%2B%20%28predictedOutput%5Bi%5D%20-%20trueOutput%5Bi%5D%29%20%2a%2a%202%3B%0A%09%7D%0A%09return%20mse%20%2F%20predictedOutput.length%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20calculate%20percentage%20of%20correct%20classifications%0A%2F%2F%0Afunction%20calculatePercentageCorrect%28predictedOutput%2C%20trueOutput%29%20%7B%0A%09let%20correct%20%3D%200%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20predictedOutput.length%3B%20i%2B%2B%29%20%7B%0A%09%09if%20%28%28trueOutput%5Bi%5D%20%3D%3D%3D%201%20%26%26%20predictedOutput%5Bi%5D%20%3E%3D%200.5%29%20%7C%7C%20%28trueOutput%5Bi%5D%20%3D%3D%3D%200%20%26%26%20predictedOutput%5Bi%5D%20%3C%200.5%29%29%20%7B%0A%09%09%09correct%20%3D%20correct%20%2B%201%3B%0A%09%09%7D%0A%09%7D%0A%09return%20%7B%0A%09%09accuracy%3A%20Math.round%28correct%20%2a%20100%20%2F%20predictedOutput.length%29%2C%0A%09%09wrong%3A%20predictedOutput.length%20-%20correct%0A%09%7D%3B%0A%7D%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20FUNCTIONS%20FOR%20VISUALISATION%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20decision%20boundary%20and%20network%20output%0A%2F%2F%0Afunction%20drawPrediction%28epoch%2C%20mseTrain%29%20%7B%0A%09%2F%2F%20make%20predictions%20with%20neural%20network%0A%09let%20training%20%3D%20model.predict%28trainInput%29.arraySync%28%29%3B%0A%09let%20validation%20%3D%20model.predict%28validInput%29.arraySync%28%29%3B%0A%09let%20prediction%20%3D%20model.predict%28testInput%29.arraySync%28%29%3B%0A%0A%09%2F%2F%20calculate%20mse%20and%20percentages%20of%20correct%20classifications%0A%09let%20mseValid%20%3D%20calculateMSE%28validation%2C%20validOutputArray%29%3B%0A%09let%20trainCorrect%20%3D%20calculatePercentageCorrect%28training%2C%20trainOutputArray%29%3B%0A%09let%20validCorrect%20%3D%20calculatePercentageCorrect%28validation%2C%20validOutputArray%29%3B%0A%0A%09%2F%2F%20draw%20decision%20boundary%0A%09decisionBuffer.noStroke%28%29%3B%0A%09let%20element%20%3D%200%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20i%20%3D%20i%20%2B%201%29%20%7B%0A%09%09for%20%28j%20%3D%200%3B%20j%20%3C%20NUM_DATA_PIXEL%20%2B%201%3B%20j%20%3D%20j%20%2B%201%29%20%7B%0A%09%09%09%2F%2F%20transform%20neural%20network%20output%20into%20a%20color%0A%09%09%09if%20%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28Math.round%28%281%20-%20prediction%5Belement%5D%5B0%5D%29%20%2a%20100%20%2B%20200.5%29%2C%20255%2C%20Math.round%28%281%20-%20prediction%5Belement%5D%5B0%5D%29%20%2a%20100%20%2B%20200.5%29%29%3B%0A%09%09%09%7D%20else%20%7B%0A%09%09%09%09decisionBuffer.fill%28Math.round%28prediction%5Belement%5D%5B0%5D%20%2a%20100%20%2B%20200.5%29%2C%20Math.round%28prediction%5Belement%5D%5B0%5D%20%2a%20100%20%2B%20200.5%29%2C%20255%29%3B%0A%09%09%09%7D%0A%09%09%09%2F%2F%20check%20for%20decision%20boundary%20and%20alter%20color%0A%09%09%09if%20%28%28element%20%25%20%28NUM_DATA_PIXEL%20%2B%201%29%29%20%21%3D%3D%200%20%26%26%20%28%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%20%26%26%20prediction%5Belement%20-%201%5D%5B0%5D%20%3C%3D%200.5%29%20%7C%7C%20%28prediction%5Belement%5D%5B0%5D%20%3C%3D%200.5%20%26%26%20prediction%5Belement%20-%201%5D%5B0%5D%20%3E%200.5%29%29%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28%27red%27%29%3B%0A%09%09%09%7D%0A%09%09%09if%20%28element%20%3E%3D%20NUM_DATA_PIXEL%20%2B%201%20%26%26%20%28%28prediction%5Belement%5D%5B0%5D%20%3E%200.5%20%26%26%20prediction%5Belement%20-%20%28NUM_DATA_PIXEL%20%2B%201%29%5D%5B0%5D%20%3C%3D%200.5%29%20%7C%7C%20%28prediction%5Belement%5D%5B0%5D%20%3C%3D%200.5%20%26%26%20prediction%5Belement%20-%20%28NUM_DATA_PIXEL%20%2B%201%29%5D%5B0%5D%20%3E%200.5%29%29%29%20%7B%0A%09%09%09%09decisionBuffer.fill%28%27red%27%29%3B%0A%09%09%09%7D%0A%09%09%09decisionBuffer.rect%28DATA_SCALE%20%2a%20i%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%20-%20PIXEL_SIZE%20%2F%202%2C%20DATA_SCALE%20%2a%20%2810%20-%20j%20%2F%20NUM_DATA_PIXEL%20%2a%20DATA_RANGE%29%20-%20PIXEL_SIZE%20%2F%202%2C%20PIXEL_SIZE%2C%20PIXEL_SIZE%29%3B%0A%09%09%09element%20%3D%20element%20%2B%201%3B%0A%09%09%7D%0A%09%7D%0A%09image%28decisionBuffer%2C%200%2C%200%29%3B%0A%09image%28decisionBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20draw%20training%20and%20validation%20samples%0A%09image%28trainBuffer%2C%200%2C%200%29%3B%0A%09image%28validBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20add%20labels%20to%20training%20and%20validation%20feature%20space%20%2F%2F%20wrong%0A%09noStroke%28%29%3B%0A%09text%28%22Training%3A%20%22%20%2B%20trainCorrect.accuracy%20%2B%20%22%25%20%E2%9C%94%EF%B8%8F%20%28%22%20%2B%20trainCorrect.wrong%20%2B%20%22%20falsch%29%22%2C%202%20%2a%20PIXEL_SIZE%2C%204%20%2a%20PIXEL_SIZE%29%3B%0A%09text%28%22Validation%3A%20%22%20%2B%20validCorrect.accuracy%20%2B%20%22%25%20%E2%9C%94%EF%B8%8F%28%22%20%2B%20validCorrect.wrong%20%2B%20%22%20falsch%29%22%2C%20DATA_WINDOW_SIZE%20%2B%202%20%2a%20PIXEL_SIZE%2C%204%20%2a%20PIXEL_SIZE%29%3B%0A%0A%09%2F%2F%20print%20training%20and%20validation%20results%20to%20console%0A%09console.log%28%22%7C%7C%20%22%20%2B%20epoch.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%20%7C%7C%20%22%20%2B%20trainCorrect.accuracy.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%25%20%7C%20%22%20%2B%20mseTrain.toFixed%286%29%20%2B%20%22%20%7C%7C%20%22%20%2B%20validCorrect.accuracy.toString%28%29.padStart%286%2C%20%22%20%22%29%20%2B%20%22%25%20%7C%20%22%20%2B%20mseValid.toFixed%286%29%20%2B%20%22%20%7C%7C%22%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20draw%20data%20in%20graphic%20buffer%0A%2F%2F%0Afunction%20drawDataInBuffer%28inputData%2C%20outputData%2C%20buffer%29%20%7B%0A%09buffer.noFill%28%29%3B%0A%09buffer.strokeWeight%28PIXEL_SIZE%20%2F%202%29%3B%0A%09buffer.rect%280%2C%200%2C%20DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09for%20%28i%20%3D%200%3B%20i%20%3C%20inputData.length%3B%20i%2B%2B%29%20%7B%0A%09%09if%20%28outputData%5Bi%5D%20%3D%3D%3D%201%29%20%7B%0A%09%09%09%2F%2F%20green%20circle%20for%20output%20label%20%3D%3D%3D%201%0A%09%09%09buffer.stroke%28%22darkgreen%22%29%3B%0A%09%09%09buffer.circle%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%2C%202%20%2a%20PIXEL_SIZE%29%3B%0A%09%09%7D%20else%20%7B%0A%09%09%09%2F%2F%20blue%20cross%20for%20output%20label%20%3D%3D%3D%200%0A%09%09%09buffer.stroke%28%22blue%22%29%3B%0A%09%09%09buffer.line%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20-%20PIXEL_SIZE%29%3B%0A%09%09%09buffer.line%28DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20%2B%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20inputData%5Bi%5D%5B0%5D%20-%20PIXEL_SIZE%2C%20DATA_SCALE%20%2a%20%2810%20-%20inputData%5Bi%5D%5B1%5D%29%20%2B%20PIXEL_SIZE%29%3B%0A%09%09%7D%0A%09%7D%0A%7D%0A%0A%0A%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%2F%2F%20MAIN%20PROGRAM%0A%2F%2F%20%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%2a%0A%0A%0A%2F%2F%0A%2F%2F%20p5.js%20setup%28%29%0A%2F%2F%0Aasync%20function%20setup%28%29%20%7B%20%2F%2Fnoprotect%20%28for%20large%20arrays%29%0A%09%2F%2F%20initiate%20console%20output%0A%09console.log%28%22%7C%7C%20%20%20%20%20%20%20%20%7C%7C%20%20%20%20%20%20Training%20%20%20%20%20%20%7C%7C%20%20%20%20%20Validation%20%20%20%20%20%7C%7C%22%29%3B%0A%09console.log%28%22%7C%7C%20Epoche%20%7C%7C%20korrekt%20%7C%20%20%20%20%20%20mse%20%7C%7C%20korrekt%20%7C%20%20%20%20%20%20mse%20%7C%7C%22%29%3B%0A%0A%09%2F%2F%20create%20canvas%0A%09createCanvas%282%20%2a%20DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09textSize%28PIXEL_SIZE%20%2a%203%29%3B%0A%0A%09%2F%2F%20create%20all%20grafic%20buffers%0A%09trainBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09validBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%09decisionBuffer%20%3D%20createGraphics%28DATA_WINDOW_SIZE%2C%20DATA_WINDOW_SIZE%29%3B%0A%0A%09%2F%2F%20draw%20initial%20training%20and%20validation%20buffer%0A%09drawDataInBuffer%28trainInputArray%2C%20trainOutputArray%2C%20trainBuffer%29%3B%0A%09drawDataInBuffer%28validInputArray%2C%20validOutputArray%2C%20validBuffer%29%3B%0A%09image%28trainBuffer%2C%200%2C%200%29%3B%0A%09image%28validBuffer%2C%20DATA_WINDOW_SIZE%2C%200%29%3B%0A%0A%09%2F%2F%20data%20preprocessing%0A%09generateTensors%28%29%3B%0A%0A%09%2F%2F%20define%20model%20and%20training%0A%09compileModel%28%29%3B%0A%0A%09%2F%2F%20train%20and%20visualise%20neural%20network%0A%09await%20trainModel%28%29%3B%0A%0A%09%2F%2F%20end%20of%20training%0A%09console.log%28%22Das%20Training%20ist%20beendet%20%3A-%29%22%29%3B%0A%7D%0A%0A%0A%2F%2F%0A%2F%2F%20p5.js%20keyPressed%28%29%20ESC%20%3D%20emergency%20stop%0A%2F%2F%0Afunction%20keyPressed%28%29%20%7B%0A%09if%20%28keyCode%20%3D%3D%3D%2027%29%20%7B%20%2F%2F%20ESC%20key%0A%09%09%2F%2F%20initiate%20stop%20of%20training%20after%20next%20epoch%0A%09%09model.stopTraining%20%3D%20true%3B%0A%09%09%2F%2F%20response%20to%20user%0A%09%09console.log%28%22Training%20wurde%20Abgebrochen%21%20Finales%20Resultat%3A%22%29%3B%0A%09%09%2F%2F%20draw%20final%20Prediction%0A%09%09drawPrediction%28actualEpoch%2C%20actualLogs.loss%29%3B%0A%09%7D%0A%7D&css=undefined&html=%3Cscript%20src%3D%22https%3A%2F%2Fapp.exorciser.ch%2Flib%2Fp5.js%22%3E%3C%2Fscript%3E%0A%3Cscript%20src%3D%22https%3A%2F%2Fcdn.jsdelivr.net%2Fnpm%2F%40tensorflow%2Ftfjs%40latest%2Fdist%2Ftf.min.js%22%3E%3C%2Fscript%3E&autorun=off&height=400px#SelteneKlasse}} |
| </WRAP> | </WRAP> |
| | |
| | \\ |
| | ==== - Bias ==== |
| | ; 🤖 KI im Allgemeinen\\ \\ |
| | : Ein neuronales Netz kann nur so gut sein, wie die für das Training verwendeten Daten. Sind die Daten einseitig, wir auch das resultierende neuronale Netz einseitig antworten. Solch ein einseitiges Antworten wird als **Bias** bezeichnet. |
| | |
| | ; 🐟 Fisch-NN\\ \\ |
| | : Würde das Fisch-NN vorwiegend mit Lodde und kaum mit Hering trainiert werden, so würde das Fisch-NN die meisten Heringe als Lodde klassifizieren. |
| | |
| | ; 💬 KI-Chatbot\\ \\ |
| | : KI-Chatbots werden vorwiegend mit Texten aus dem Internet trainiert, in welchen westliche Philosophien vorherrschen, sehr viele einfache Sichtweisen und Vorurteile vorhanden sind und Minderheiten häufig schlecht dargestellt werden. Daher ist es kaum zu vermeiden, dass KI-Chatbots diese Eigenheiten wiedergeben. KI-Chatbots unterliegen daher immer einer Bias und können nicht als "neutral" bezeichnet werden. |
| | |
| | <WRAP center round box > |
| | == ✍ Auftrag == |
| | In diesem Auftrag recherchierst du nach konkreten Beispielen für eine KI-Bias. |
| | - Recherchiere nach drei verschiedenen KI-Bias-Beispielen. Überlege dir, wer davon betroffen ist und welche Auswirkungen das für die Betroffenen hat. |
| | - Schreibe deine Beispiele und Überlegungen in das Textfeld. {{gem/plain?0=N4XyA#094093a86002bb1a}} |
| | </WRAP> |
| | \\ |
| | |
| ==== - Black Box ==== | ==== - Black Box ==== |
| [{{ p:pasted:blackbox.png?400px|Neuronales Netz als Black Box ((eigene Darstellung, [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]])) }}] | <figure right>{{p:pasted:blackbox.png?400px}}<caption>Neuronales Netz als Black Box ((eigene Darstellung, [[https://creativecommons.org/publicdomain/zero/1.0/deed.de|CC0 1.0]])) </caption></figure> |
| |
| ; 🤖 KI im Allgemeinen\\ \\ | ; 🤖 KI im Allgemeinen\\ \\ |
| : Das Wissen eines neuronalen Netzes ist in seinen Gewichten gespeichert. Diese Gewichte sind im Wesentlichen mehrere, mehr oder weniger miteinander verbundene Zahlen. Damit lässt sich zwar gut aus einer Eingabe eine Ausgabe berechnen, jedoch lässt sich aufgrund dieser Gewichte kaum herausfinden, warum ein neuronales Netz eine bestimmte Ausgabe ausgibt. Nur schon deswegen sind neuronale Netze nicht für jeden Anwendungsbereich geeignet. Es gibt zwar andere KI-Verfahren, wie z.B. statistische Verfahren, welche besser interpretiert werden können, jedoch zeigen diese in einigen Anwendungsgebieten deutlich schlechtere Resultate. | : Das Wissen eines neuronalen Netzes ist in seinen Gewichten gespeichert. Diese Gewichte sind im Wesentlichen mehrere, mehr oder weniger miteinander verbundene Zahlen. Damit lässt sich zwar gut aus einer Eingabe eine Ausgabe berechnen, jedoch lässt sich aufgrund dieser Gewichte kaum herausfinden, warum ein neuronales Netz eine bestimmte Ausgabe ausgibt - in diesem Sinne ist ein neuronales Netz eine **Black Box**. Nur schon deswegen sind neuronale Netze nicht für jeden Anwendungsbereich geeignet. Es gibt zwar andere KI-Verfahren, wie z.B. statistische Verfahren, welche besser interpretiert werden können, jedoch zeigen diese in einigen Anwendungsgebieten deutlich schlechtere Resultate. |
| |
| ; 🐟 Fisch-NN\\ \\ | ; 🐟 Fisch-NN\\ \\ |
| |
| \\ | \\ |
| |
| ===== - Fazit ===== | ===== - Fazit ===== |
| <WRAP center round tip 80%> | <WRAP center round tip 80%> |
| Wir Menschen wählen das KI-Modell, die Fehlerfunktion und die Daten für das Training aus. Somit sind wir auch verantwortlich dafür, was eine KI macht. Aber egal wie viel Mühe wir uns dabei auch geben und egal wie gut die KI am Ende funktioniert, eine KI wird vermutlich immer dann und wann Fehler produzieren. Es stellt sich die Frage, wie wir damit umgehen wollen. | Wir Menschen wählen das KI-Modell, die Fehlerfunktion und die Daten für das Training aus. Somit sind wir auch verantwortlich dafür, was eine KI macht. Aber egal wie viel Mühe wir uns dabei auch geben und egal wie gut die KI am Ende funktioniert, eine KI wird vermutlich immer dann und wann Fehler produzieren und eine gewisse Bias aufweisen. Es stellt sich die Frage, wie wir damit umgehen wollen. |
| </WRAP> | </WRAP> |
| |
