Abstract:

As many of us know by now, Deep Learning has enabled tackling very challenging problems and applications that were previously almost impossible to solve with machine learning. However, for most of the tasks we want to solve with Deep Learning, we need large, if not huge, amounts of data and computing power. This is very limiting for many applications for which we do not have the necessary amounts of data or for practitioners who do not have access to enough computation power to train well-performing Deep Networks for their desired tasks.We hope to overcome these two limitations by leveraging the generality of already trained models through Transfer Learning or combining the information from multiple, perhaps relatively small, datasets with Multi-Task-Learning.In this project, we are investigating the generality of representations learned by Deep Networks. Today I would like to introduce one of the families of tools we use in this effort: Representational Similarity Analysis (RSA).I will present the methodology behind these tools and provide some insights into Deep Networks gained through their use. However, I would highlight some concerns to be aware of when using these tools and present some challenges that arise in practice. Considering these concerns, I will present a variant of these tools that solves some of the existing problems.Furthermore, I will shortly present a tool that we have developed to synthesize realistic image data, allowing us to systematically analyse which properties of the data are represented in Deep Networks.Finally, I want to mention our SCIoI cooperation with project 01 on “Scanpath Prediction in Dynamic Scenes using an end-to-end Deep Learning approach”.

Photo by Nina Ž. on Unsplash

This talk will take place in person at SCIoI.

***Want to attend one of our events? Sign up here.
To get regular updates, subscribe to our mailing list from this page.