Sparse Autoencoders (SAEs) are typical autoencoders that enforce a
sparsity constraint on their hidden units, which are typically of
higher dimensionality than their input. This is motivated by the
superposition hypothesis: neurons hold information about multiple
concepts that are rarely simultaneously present. SAEs are frequently
used for mechanistic interpretability today.
↩
This 40x40 grid gives the average activation value of the 1600 hidden
units in this SAE, computed across a sample of the entire MNIST
dataset. Every two seconds, a new hidden unit is chosen.
↩
These four images displayed are the images which activated the current
hidden layer activation unit the most. It might be fun to see if there
is a common pattern among these images.
↩