Talks
All the talks so far.
Training LLMs to Self-Correct via RL
Adhilsha AnsadPublished: at 03:45 PMThis talk will discuss the training of large language models (LLMs) to self-correct their predictions using reinforcement learning (RL).
GNNX-BENCH - Perturbation-based GNN Explainers
Rishi Raj SahooPublished: at 02:00 PMThis work presents a benchmark for Graph Neural Network Explainability with perturbation. This involves some metrics like Sufficiency, Necessity, Stability, and Reproducibility to observe which method performs better. Finally, the work provides recommendations for choosing methods for particular graph tasks.
Balancing Explainability-Accuracy of Complex Models
Poushali SenguptaPublished: at 03:45 PMThis talk covers the Explainability through Correlation Impact Ratio (ExCIR) method, which balances explainability and accuracy in complex machine learning models, especially with independent features. ExCIR creates a streamlined data space from the original dataset, requiring fewer input samples while maintaining model accuracy. The Correlation Impact Ratio (CIR) quantifies each feature’s contribution to the model’s output, accounting for feature uncertainty regarding its presence or absence. Using the CAUEEG dataset related to dementia, we demonstrate that ExCIR maintains accuracy while enhancing explainability through feature impact scores. The method reliably ranks feature importance in both original and simplified models, validating its consistency.
End-to-End Task-Oriented Dialogue Systems
Nalin KumarPublished: at 03:45 PMThis talk covers recent developments in dialogue systems, focusing on task-oriented systems that help users accomplish specific goals, like booking services or providing customer support. It begins by distinguishing between open-domain dialogue systems, which handle general conversations, and task-oriented systems, which are goal-driven. The talk provides an overview of traditional approaches, including statistical and neural network-based methods, and highlights the shift toward end-to-end architectures, along with existing challenges in system evaluation. Despite advancements, task-oriented systems often generate dull, misaligned responses. Moreover, the speaker presents recent work, accepted at NAACL 2024, that improves the naturalness and alignment of responses in task-oriented settings without compromising task success.
PIDformer - Transformer Meets Control Theory
Pankaj KumarPublished: at 11:00 AMThis talk discusses a paper that addresses two key flaws of Transformer architectures - input corruption and rank collapse in output representation, by framing self-attention as a state-space model and revealing its tendency towards lower-rank outputs and sensitivity to input perturbations and introducing a Proportional-Integral-Derivative (PID) feedback control system to enhance robustness and representation capacity.
Online Dynamics Learning for Predictive Control with an Application to Aerial Robots
Rishi Raj SahooPublished: at 11:00 AMThis work presents an online dynamics learning framework to improve the accuracy of model predictive control (MPC) during deployment. Using knowledge-based neural ODEs (KNODEs) and transfer learning techniques, the model continually adapts to disturbances, demonstrated through simulations and quadrotor experiments. Results show improved trajectory tracking under varying conditions.