Recent Talks

This talk provides a comprehensive overview of ferret, a novel python library for benchmarking interpretability techniques on …

All recent language technologies, such as BERT, GPT3, or ChatGPT, rely on complex neural networks called Transformers. But, what …

In the blooming age of language technologies, the impact of social biases encoded in our models has surged to unprecedented levels. In this half-talk, half-discussion, we go through fundamental definitions of social bias in computer systems, specifically language models and social harm.

In this talk, I present a chronological overview of modeling solutions for time series and natural language. I go through seminal papers on the applications of recurrent models to tasks like language modeling, neural machine translation, image captioning, or multi-model text generation.

List of papers I presented at reading groups (latest first). Slides are free to use upon attribution. Sharing and mixing under CC BY-SA 4.0. Toolformer: Language Models Can Teach Themselves to Use Tools Training language models to follow instructions with human feedback The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks How Gender Debiasing Affects Internal Model Representations, and Why It Matters Multitask Prompted Training Enables Zero-Shot Task Generalization How does the pre-training objective affect what large language models learn about linguistic properties Vision-and-Language or Vision-for-Language?