Ishan Misra: Self-Supervised Deep Learning in Computer Vision | Lex Fridman Podcast #206
Topics covered
Popular Clips
Episode Highlights
Contrastive Learning
Contrastive learning is a powerful paradigm in machine learning that enhances model training by distinguishing between related and unrelated data samples. explains that this method involves creating an embedding space where similar items are pulled together, while dissimilar ones are pushed apart. This approach is not limited to self-supervised learning but is also prevalent in supervised learning contexts 1.
The idea is that you have a sample, you have another sample that's related to it, so that's called the positive. And you have another sample that's not related to it. So that's negative.
---
Energy-based models further elucidate this concept by using energy functions to describe the relationships between data points, revealing commonalities across various learning models 2.
Active Learning
Active learning is a strategic approach in machine learning that optimizes data usage by intelligently selecting the most informative samples for training. highlights its efficiency in learning by asking targeted questions about data, which helps models learn faster and with fewer samples 3. This method involves understanding what the model knows and doesn't know, allowing it to focus on areas that need improvement.
The idea was basically you would train an agent that would ask a question about the image, it would get an answer and basically then it would update itself.
---
Misra also notes its potential in data labeling, where models predict similarities and dissimilarities to refine their understanding of concepts 4.