TopK SAE
A class of autoencoders that learn sparse features by forcing each token's hidden-layer activation to retain at most k nonzero values, thereby avoiding the shrinkage bias of L1 penalties.
Timeline (last 90 days)
TopK SAE keeps the largest several pre-activation values on each token, and the margin of a feature is defined as the distance between its pre-activation value and the truncation threshold.
The researchers conducted experiments on five public TopK SAE pairs for three Pythia models (the largest being pythia-160m).
When a feature is set to zero, the proportion of next-layer features crossing the truncation threshold reaches as high as above 0.74, and these crossing events carry most of the energy of the response.
TopK SAE is a class of autoencoders that learn sparse features by forcing the hidden-layer activations of each token to retain at most k non-zero values, unlike traditional SAEs that use L1 regularization to encourage sparsity.
In the median case, the decoded change of features crossing the threshold is more than 1.24 times the norm of the model's own change, indicating that the contribution of these features to the output cannot be ignored.
All Facts (5)
TopK SAE keeps the largest several pre-activation values on each token, and the margin of a feature is defined as the distance between its pre-activation value and the truncation threshold.
50%UnverifiedThe researchers conducted experiments on five public TopK SAE pairs for three Pythia models (the largest being pythia-160m).
50%UnverifiedWhen a feature is set to zero, the proportion of next-layer features crossing the truncation threshold reaches as high as above 0.74, and these crossing events carry most of the energy of the response.
50%UnverifiedTopK SAE is a class of autoencoders that learn sparse features by forcing the hidden-layer activations of each token to retain at most k non-zero values, unlike traditional SAEs that use L1 regularization to encourage sparsity.
50%UnverifiedIn the median case, the decoded change of features crossing the threshold is more than 1.24 times the norm of the model's own change, indicating that the contribution of these features to the output cannot be ignored.
50%