Self-Supervised Learning
Ishan discusses the innovative approach of using a vast collection of uncurated internet images—approximately a billion—to train large convolutional models in a self-supervised manner. This method challenges traditional datasets like ImageNet, raising questions about model performance and the inherent biases in image selection. The findings suggest that self-supervised learning can effectively harness diverse and random images, potentially leading to new insights in object recognition and machine learning scalability.In this clip
From this podcast
Lex Fridman Podcast
Ishan Misra: Self-Supervised Deep Learning in Computer Vision | Lex Fridman Podcast #206
Related Questions