GS 3 Day 2 (8-April-2021)
3.1 A general ML session bt Geoffrey Hinton 9 AM
Prof. Geoffrey gave a talk on a non-working system called GLOM. It tries to solve the question “How can a neural network with a fixed architecture parse an image into a part-whole hierarchy?” Geoffrey believes that if GLOM can be made to work in neural networks and transformers, it will significantly improve interpretability.
3.2 A AI4SG session by Pradeep Varakantham 10:30 AM
Pradeep talked about a non-trivial solution to an optimal supply-demand ecosystem that is not working based on greedy supply to the immediate demands (e.g., assigning the closest Uber taxi to a customer). He proposes a Resource-constrained RL model for the problem where several constraints specific to a problem are supplied to an RL algorithm. He called it ReCo-RL (Resource-constrained RL problem). Another applied approach is the Deep Q network. Pradeep is trying to solve problems related to Taxi fleets, Emergency Response, Traffic and Security Patrols, and, Bike-sharing systems.
3.3 A General ML Session by M Pawan and K Dvijotham, 2 PM
The talk covers various techniques to prevent adversarial attacks on image classifiers. These techniques involve the verification of the neural networks trained on the datasets.
3.3.1 Additional resources
This was a close to the common sessions, from tomorrow onwards, special topic sessions are delivered. I am in the core-ML track