AI-Introduction

• Presenter: Nils Seipel

Premises and Terms

• The goal of artifical intelligence is to simulate human intelligence
• This achieved by simulation functions of signal neurons.
• Problem: Human intelligence is not well defined
• Artificial intelligence > Machine Learning > Deep Learning
• Artificial intelligence(knowledge-based systems)
• StrongAI: General AI, solves Problems of its own (Holy Grail/Singularity)
• WeakAI(narrowAI): helps with specific problems, not capable of knowledge transfer

Algorithm:

• An algorithm is like a recipe, defined input and expected output are clear
• The problem is: algorithms are difficult to adept to reality we need to cover every possible case

Artificial neural networks (ANN)

• Solve the problem with algorithms
• Can adapt to different inputs
• Have to be trained by Humans or another Artificial neural network
• Training data must be complex and we need a lot. And this data needs to be labled.
• Training is really crucial, bad training creates bad ai.
• Biased data will create a biased ai, diverse data is important to remove bias. This can become dangerous for societies that relay on AI. It has the potential to perpetuate prejudice.
• AI learns similar to learning child by pointing on things
• try yourself: GOOGLETOOL
• A shallow neural networks is build out of input layer, processing layer and output layer.
• Deep neural networks means: lots of processing layers of “neurons”.
• After training, the artificial neural network is fixed, it doesnt learn any more.
• All ANN work based on probability.

Use-cases:

• Cancer diagnosis, histological images are rated by an AI that supports the doctor.
• Self-driving-cars
• Learn sign-language assisted by ai
• DALL-E
• Deep-Fake Videos
• openAI-Playground
• Universal Translation

Bias in neural networks:

• Really important issue, if we give important tasks/decisions to ai
• Reality is sometimes not diverse enough to provide sufficiant data
• Currently to focused on western white males, due to limits in training data

Big data is required for training

• 45 petabyte pure text as training data for ChatGPT
• Example: In order to classify a handwritten 0 as a zero, with 95% likelyhood of correct classification you need 50.000 handwritten samples of this singel digit.

Future of AI

• WeakAI will be better than humans in every discipline.
• AI will not drive us into unemployment, but more that we’ll be interacting with AI. Simlar to a conductor and the orchestra working together.
• https://makerspace-giessen.de/ki/ further reading and nice resources
• AI for Start-Ups

Assist me, AI: The pun-ishingly good future of productivity

• presenter: Johannes Hammp
• Why this talk: Plagiarism detector software didn’t recognize plagiarims.
• The whole talk is available as pdf

Quillbot

• Paraphraser-Tool is often used by students for homework
• Quality is often lacking
• Problem: PlagAware (anti-plagiarism) Software does not recognize anything

Semantic Scholar

• Alternative to Google Scholar
• Creates references for you
• Has integrated “unpaywall”
• AI-feature TLDR (Beta). Shorter then abstracts, oftern better content summary

ChatGPT

• does everything what all the other tools do
• higher usability through the chat like interface
• keywords to circumvent blocks: hypothetical,educational
• Uses “However” to often in its answers
• Context can also be a primer into a fault direction
• Awesome prompts, great prompts for chat gpt (https://prompts.chat/)

Use-cases:

• Essay-writing
• Code-writing/Debugging
• Summariser
• Assistant to create fundamentals of a research paper.
  • Advantage: new way of writing, less writing blockade. not necessarily less time consuming.

Problem: Teachers are confronted with content which is often difficult to veryfi The border between human and AI generated content blurs.

• Enhance your own creativity with suggestions from ChatGPT

Transparency Issues

• not everything labeled AI is real AI
• we dont know how it works
• it has some quality issues, e.g. creating eloquent bullshit
• reproducibility