The objective of "0th Berlin Symposium on Artificial Teacher Avatars" is the presentation, discussion and development of tools, models and methodologies leading to emergence of generative AI avatars usable and useful in domain of education.

Generative AI is reality and it's better to be its co-creator then its consumer.

Learning by doing & experimenting & creating critical stance.

Personalized AI tutors are important components of vaste majority of future educational systems.

Property "avatarized" is the property 23 on the roadmap to an ideal Bildungsinstrument I've been working on since some times.

It's about:

Artificial Intelligence in Education (AIED): Education about AI & Education with AI

generative artificial intelligence (notably large language models, text-to-speech / voice cloning, talking head generation)

multi-modal & collective avatars

open-source tools (text-webui-..., comfy ui), programming languages (python, mojo), models (Stable Diffusion, Mistral) & concrete methods (LORA, Direct Preferenc Optimization)

tweaking knobs & parameters

& maybe little bit of "prompt engineering"

This symposium is NOT about:

dysinformation

deep-fakes of living persons

corporate products

entertainment (e.g. humor / horror / porn / cognitive spam)

so-called discriminative AI (e.g. speech/face/object recognition etc.)

I-avatarizaion

Anything else?

During today's workshop, I will introduce You to two methods how You can fine-tune Your model to allow it to provide better inferences for the task You want to solve.

Lora training :: here You just have a bunch of data

Direct Preference Optimization :: here You have data which contains list of tasks (e.g. questions) and associated correct and incorrect answers

In the context of Artificial Intelligence in Education (AIED), teacher avatarization is the process of transforming the dataset produced by one concrete human teacher (or a precisely-defined group of human teachers) into a multi-modal generative artificial intelligence system with which future generations of learners can seemlessly interact.

 "descent of a Hindu deity to earth in an incarnate or tangible form," from Sanskrit avatarana "descent" (of a deity to the earth in incarnate form), from ava- "off, down" (from Proto-Indo-European root *au- (2) "off, away") + base of tarati "(he) crosses over" (from PIE root *tere- (2) "cross over, pass through, overcome").

The meaning "concrete embodiment of something abstract" is from 1815. In computer use, it seems to trace to the novel "Snowcrash" (1992) by Neal Stephenson.

Within the "idealist" philosophical tradition, it is not necessarily the physical person but rather a "pure idea" (e.g. protection, justice, wisdom, friendship etc.) which is to be "avatarized".

"I"-Avatarization is the process whereby a living human H consciously creates, develops, fine tunes and optimizes (his|her) own generative AI avatar datasets & models.

That is, using datasets (mails, chat transcripts etc.) to create a generative AI copy of one's self (an "I-Avatar") which could provide information in situation when H (her|him)self is not alive anymore.

Großes Sprachmodell (LLM), oder Large Language Model, ist ein Typ der künstlichen Intelligenz, der entwickelt wurde, um menschliche Sprache zu verstehen und zu generieren. Diese Modelle werden an riesigen Textmengen geschult, wodurch sie Grammatik, Wortschatz und Kontext verschiedener Sprachen lernen. Die Fähigkeit von LLM, zusammenhängende und grammatikalisch korrekte Texte zu generieren, ermöglicht es ihnen, verschiedene Aufgaben zu erfüllen, wie Fragen zu beantworten, Texte zu schreiben, Sprachen zu übersetzen oder Informationen zu zusammenzufassen, und das alles ohne explizite Programmierung spezifischer Regeln oder Logiken.

Neuronale Netze im Kontext der KI sind keine echten Nervenzellen, sondern softwarebasierte Modelle, deren Architektur von der Art und Weise inspiriert ist, wie das menschliche Gehirn Informationen verarbeitet. Diese Modelle bestehen aus "Schichten" von Datenstrukturen, die "Neuronen" genannt werden. Die Neuronen sind miteinander verbunden und ihre Verbindungen haben ein bestimmtes "Gewicht". Im Lernprozess passt das System diese Gewichte - auch "Parameter" genannt - allmählich an, um den Fehler zwischen seiner Vorhersage und dem, was vorhergesagt werden soll, zu verringern.

Tiefes Lernen oder Deep Learning nutzt neuronale Netze mit vielen Schichten (daher 'tief'), was dem Modell ermöglicht, komplexe Muster in Daten auf verschiedenen Abstraktionsebenen zu lernen. Deep Learning wird in vielen Bereichen angewendet, einschließlich Bild- und Spracherkennung, und ist eine Schlüsseltechnologie in der Entwicklung autonomer Systeme wie selbstfahrende Autos oder Spracherkennung.

Der Schlüssel zum Erfolg ist eine möglichst große Menge an Trainingsdaten.

GPU (Grafikverarbeitungseinheiten, NVIDIA) und TPU (Tensor-Verarbeitungseinheiten, Google) sind jene Hardware-Innovationen, die tiefgehendes Lernen in einem unvorstellbaren Umfang ermöglichen.

A trained neural network is stored in a file. This file is called "a model". When You want to use it - either for inferencing or training or both - You need to load the model from the disk into memory.

Lora is like a small "adapter" model which You can train as an addition to Your main "big" model. Training of a Lora is much faster than fine-tuning of the big model itself.

In machine learning and AI, we speak about

"training" when the AI is learning from the data we provide it

"inferencing" when the AI is using its current "knowledge" to solve new problems (e.g. problems which maybe weren't in the training data at all)