Explore an in-depth lecture on latent variable Energy-Based Models (EBMs) for structured prediction, delivered by renowned AI researcher Yann LeCun. Dive into various topics including training EBMs, contrastive and regularized methods, general margin loss, joint embedding architectures, and generative adversarial networks (GANs). Learn about cutting-edge techniques such as Wav2Vec 2.0, XLSR for multilingual speech recognition, and non-contrastive methods like BYOL and SwAV. Discover practical applications of latent variable models, including DETR for object detection. Gain insights into structured prediction, factor graphs, and the Viterbi algorithm through detailed explanations and whiteboard demonstrations. Conclude with an exploration of graph transformer networks and their compositions, providing a comprehensive overview of advanced machine learning concepts and techniques.
Latent Variable EBMs for Structured Prediction
Overview
Syllabus
– Welcome to class
– Training of an EBM
– Contrastive vs. regularised / architectural methods
– General margin loss
– List of loss functions
– Generalised additive margin loss
– Joint embedding architectures
– Wav2Vec 2.0
– XLSR: multilingual speech recognition
– Generative adversarial networks GANs
– Mode collapse
– Non-contrastive methods
– BYOL: bootstrap your own latent
– SwAV
– Barlow twins
– SEER
– Latent variable models in practice
– DETR
– Structured prediction
– Factor graph
– Viterbi algorithm whiteboard time
– Graph transformer networks
– Graph composition, transducers
– Final remarks
Taught by
Alfredo Canziani
