In our first two Meetups in December last year, we had a look at OpenAIs GPT models and the underlying Transformer model.
- In the first Meetup, we were looking at the special characteristics of the Transformer model introduced in 2017 by the paper titled Attention Is Al You Need. In order to do so, we were discussing the video Transformer Neural Networks - EXPLAINED!, which brilliantly explains the basics of the Transformer in just 13 minutes. In a further video, we were looking at the advantages of using a Transformer-based approaches as the ones used in the BERT or the GPT-3 model by discussing the video NLP with Neural Networks & Transformers.
For further details on the Transformer model, we also highly recommend Jay Alammars blog post The Illustrated Transformer.
- The second meetup started with a short presentation by Henrik Horst, course guide of the Natural Language Processing course at opencampus.sh, on the technical communalities and differences of the GPT-2 and GPT-3 models published by OpenAI.
Thereafter, we had a presentation by Vladimir Alexeev, also known as "Merzmensch". Vladimir is one of the few "GPT-3 Ambassadors" who serve as point of contact between the GPT-3 community and OpenAI, and he has been working extensively with the GPT-3 model. In his presentation, he was sharing the extraordinary capabilities of GPT-3 (see, for example, this short story and his presentation here) in conversation, writing poems, creating short stories, and other fields of creativity.
Our last two Meetups then focused on the current capabilities of Generative Adversarial Networks (GANs) and their application.
- In our first Meetup on GANs we were happy to welcome Vladimir Alexeev again, this time sharing his work in the field of image and video generation based on GANs (see his presentation here), which he also summarized in his blog post Creative Artificial Intelligence. Index.
- In the second Meetup on GANs we then had a look at other applications of GANs that were given in a presentation by Ian Goodfellow (the inventor of the GAN) in September last year, in which he showed their application for making dental crowns or extending training datasets. In a further presentation by Anmima Anandkumar (Machine Learning Director at NVIDIA) from the same event, we had a look at the current research issues in the field of GANs.
The end of the Meetup was then marked by a presentation of Ali Farooq, who presented his final project from last semester's course Machine Learning With TensorFlow at opencampus.sh, in which he together with Felix Facklam and Dieter Hoogestraat were generating CT scans of COVID-19 infected lungs using GAN.