Can AI Learn Hierarchies Like Humans Do?

​Enforcing structure in representation spaces is an important problem to solve in Computer Vision. In this talk, we will do a deep dive into two pivotal papers which introduce novel techniques for solving this. The first is ComFe, which learns interpretable "image prototypes", and then we will discuss Deep Taxonomic Networks, which learns a hierarchical representation space in an unsupervised way.

​About the speaker:
Harsha Bommana is an AI researcher at Lossfunk, working on compositional representation learning. He has eight years of experience in deep learning across multiple startups and runs Deep Learning Demystified, a page focused on explaining deep learning in a simple, accessible way.
LinkedIn: https://www.linkedin.com/in/harshabommana/
DLD IG: https://www.instagram.com/deeplearningdemystified/

​Ayush Nangia is an AI Researcher at Lossfunk who is working on kernel optimisation and diffusion language models.
Website: https://vitransformer.netlify.app/
X: https://x.com/vitransformer

​
To attend online:

• ​⁠Add to calendar: https://bit.ly/4jrPE9i

• ​Gmeet link: https://meet.google.com/kvv-zidk-rdx?hs=122&authuser=0

​Pre-read material:

• ​ComFe - https://arxiv.org/abs/2403.04125

• ​Deep Taxonomic Networks - https://arxiv.org/abs/2509.23602