Unsupervised Representation Learning

Explore unsupervised representation learning in this comprehensive lecture by Yann LeCun from New York University. Delve into deep learning concepts, focusing on hierarchical representations and instance segmentation with Mask R-CNN. Examine the nature of common sense and its role in machine learning. Investigate techniques for training actors with optimized action sequences and augmenting neural networks with memory modules. Discover energy-based unsupervised learning strategies, including methods to shape energy functions and maintain constant low-energy volumes. Analyze approaches like sparse coding, auto-encoders, and predictive sparse decomposition. Tackle the challenges of prediction under uncertainty, exploring invariant prediction and video frame forecasting. Gain insights into cutting-edge machine learning research and applications in this hour-long Simons Institute presentation.

Intro
Deep Learning = Learning Hierarchical Representations
Mask R-CNN: instance segmentation
What is Common Sense?
Common Sense is the ability to fill in the blanks
How Much Information Does the Machine Need to
Training the Actor with Optimized Action Sequences
Augmenting Neural Nets with a Memory Module
Memory/Stack-Augmented Recurrent Nets
Entity Recurrent Neural Net
Energy-Based Unsupervised Learning
Seven Strategies to Shape the Energy Function
constant volume of low energy Energy surface for PCA and K-means 1. build the machine so that the volume of low energy stuff is constant
use a regularizer that limits do the volume of space that has low energy Sparse coding, sparse auto-encoder, Predictive Sparse Decomposition
The Hard Part: Prediction Under Uncertainty Invariant prediction: The training samples are merely representatives of a whole set of possible outputs (eg, a manifold of outputs).
Video Prediction: predicting 5 frames