Johanna Rock

Johanna Rock studied Computer Science at Technical University Graz with a focus on IT Security and Computational Intelligence. In 2022 she finished her PhD in the field of resource-efficient neural networks for radar signal processing.
She worked as an intern in the Arm Machine Learning Research Group in Cambridge and started a full time position at Tenstorrent Inc. as a Senior Engineer in Machine Learning in Fall 2022.


Session

04-15
13:00
45min
How to get by with AI: AI Tooling and how it’s changing everything we do.
Gabriel Schanner, Johanna Rock

Breakthroughs in AI frequently made the news in 2022. While “AI” as a Buzzword has been around for many years, it’s become a fast-evolving technology in recent times. AI models have not only become much more powerful, but they have also reached a level of maturity and ease of access that changed the user base from scientists and tech enthusiasts to people with a large variety of backgrounds. These tools offer services in a variety of fields and can be grouped by input and output modalities: text/language, audio, video, and images.

Arguably the most prominent ones in recent times have been the following:

  • Dalle 2.0 by OpenAI April 2022
  • Stable Diffusion by Stability.ai in September 2022
  • ChatGPT by OpenAI in November 2022
  • CoPilot by Github in 2021

The first part of this talk aims to give an introduction to these AI tools. With a structured overview, inspiring examples, and references to open-source alternatives, we want to illustrate how broad the AI tooling landscape has become and how fast it is evolving. Based on our trials, we show use cases of how to improve every day (work) life and discuss the potential as well as challenges and limitations when using these new tooling concepts.

The second part of this presentation focuses on the underlying AI concepts and technical background. Through one specific model, i.e. ChatGPT, we want to give the listener a good intuition about the inner workings of the model. We illustrate the model architecture and explain how training was done. We discuss the basic building blocks and then move on to the decisive breakthrough concepts that made modern generative AI so successful. Employing concrete examples we demonstrate the limitations and shortcomings of these models and point out directions of ongoing research.

With this talk, we hope to inspire new usage ideas of AI tools in the listener and give a good intuition about what AI tools are and how they work. It’s not black magic, but this doesn’t make it less fascinating.

HS i7