Tri Dao, Stanford: FlashAttention and sparsity, quantization, and efficient inference
Generally Intelligent
00:00
Understanding Data Augmentation and its Mathematical Link to Feature Regularization
Data augmentation is a way to encode knowledge into a model by modifying the data directly./nUsing data augmentation in training contrastive learning models has been successful in learning good representations./nThere is a mathematical link between data augmentation and feature regularization, which helps enforce invariance in the features that the model learns./nManipulating features instead of manipulating images can also be used to translate knowledge about images.
Play episode from 08:29
Transcript
