DoC student presents research poster at prestigious AAAI-23 conference

The AAAI Conference on Artificial Intelligence is a prestigious annual international conference in AI.

Sudarshan Sreeram, final Year MEng Computing at Imperial College London, attended the prestigious AAAI-23 conference at the Walter E. Washington Convention Center to present a poster of his accepted student abstract on "Exploring The Relative Value of Collaborative Optimisation Pathways."

Sudarshan's work lays the groundwork for questioning the merit of a full-scale model compression pipeline involving pruning, clustering and quantisation-aware training (PCQAT) as opposed to skipping either or both pruning and clustering, which would reduce re-training costs and could yield better metrics. This initial case study was on a small keyword spotting model, and he's excited to extend this work based on the feedback he's received during the poster session.

Sudarshan said: "I found it an insightful experience to interact with fellow students, experienced academics, and industry professionals, especially at the poster sessions." Furthermore, Prof. Tenenbaum's lecture on "Learning to see the human way" was delightful and eye-opening; it was certainly interesting to see ideas from a cognitive science perspective. The notion of robust world modelling to capture and grow intelligence in the ways a human being does is an understated perspective. I loved his examples throughout the lecture, especially the one on checking for emergent world models in LLMs, where he asked ChatGPT to devise a plan to retrieve a ball from underneath a couch (attached)."

He conducted the work presented in his abstract during his industrial placement at Arm last year and also said: "I'm grateful to Dr Elena Zhelezina and Dr Anton Kachatkou from Arm for their continued support, constructive comments, and insightful discussions. In addition, I would also like to thank Imperial College Computing for partially sponsoring the expenses for this wonderful trip."