Talk, IBM Research, New York, USA
Presented our work on using attention guidance, in which we use intuitive priors to modify self-attention heads in Transformers to get faster convergence and better performance. [Slides]
Talk, Research Group, Princeton/IITM
ELECTRA - This work introduces a new pre-training objective called replaced-token detection which is more efficient than the traditional masked language modeling objective. [Slides]Plug and Play Language Models - This work introduces simple approach that can control the text generation based on attributes. [Slides]Abstractive and Extractive Summarization - Presented two key papers for text summarization using neural networks. [Slides]A Laplacian Framework for Option Discovery in Reinforcement Learning - Presented usefulness of Proto-Value Functions and the paper’s laplacian framework for finding useful options. [Slides]Dataless Classification - Presented the idea of using semantic information in label names and how this idea dovetails with co-training. [Slides]Children search for information as efficiently as adults, but seek additional confirmatory evidence - Presented a cognitive approach to explaining how children and adults search for information equally efficiently but differ in the implementation of the stopping rule along with a primer on ANOVA and Bonferroni corrected multiple comparisons. [Slides]