The International Conference on Learning Representations (ICLR) is a world leading event which brings together experts in the field of deep learning, and is globally renowned for being at the cutting-edge of research. Deep learning has applications in a variety of fields including AI, statistics and data science, computational biology, gaming, and robotics.
Lisa Bonheme presented research on deep generative models at the conference this month, and featured amongst a select group of speakers. Deep generative models, like ChatGPT or Midjourney, are deep learning models that can create new data, whether that’s a conversation about cats or images of cats. However, sometimes it goes wrong; ChatGPT becomes nonsensical and Midjourney generates something closer to a nightmare than to a cat. But why does it happen?
“My goal is to explore the inner workings of generative models, especially how the concepts they have learned can explain the image (or text) they generate. I hope my research will help to better understand, train and improve these models.”
Lisa represented Kent at the ICLR as part of a paper track which aims to increase participation from underrepresented groups.