Deep Learning for Vision: Tricks of the Trade

Explore the evolution and practical applications of deep learning in computer vision through this 42-minute talk by Marc'Aurelio Ranzato, a research scientist at Meta. Gain insights into the historical development of deep learning over the past two decades, understanding the reasons behind its recent success and learning practical techniques for implementing these methods in common vision applications. Discover various approaches, from unsupervised feature learning algorithms to popular object recognition systems, while examining the challenges faced by the field and potential future breakthroughs. Delve into topics such as ideal feature extraction, learning non-linear features, convolutional neural networks, and strategies for optimizing and improving generalization in deep learning models. Learn from Ranzato's extensive experience in machine learning, computer vision, and artificial intelligence as he shares valuable insights and practical advice for implementing deep learning techniques in vision-related tasks.

Intro
Ideal Features
The Manifold of Natural Images
Ideal Feature Extraction
Learning Non-Linear Features
Linear Combination prediction of class
A Potential Problem with Deep Learning
Deep Learning in Practice
KEY IDEAS: WHY DEEP LEARNING
Buzz Words
(My) Definition
ConvNets: today
Deep Gated MRF
Sampling High-Resolution Images
Sampling After Training on Face Images
Cons
CONV NETS: TYPICAL ARCHITECTURE
CONV NETS: EXAMPLES
CHOOSING THE ARCHITECTURE
HOW TO OPTIMIZE
HOW TO IMPROVE GENERALIZATION
OTHER THINGS GOOD TO KNOW
WHAT IF IT DOES NOT WORK?