PRISMΔDB
Training Center

Start New Experiment

Latent expansion factor (e.g. 4x)

Top-K active neurons

Auxiliary Loss Coefficient

Cancel
Hyperparameters

Configure the architecture and training constraints for the Sparse Autoencoder.

  • Expansion Factor Determines the size of the latent space relative to the input (e.g. 4x means latent_dim = input_dim * 4).
  • K-Sparsity (Top-K) Enforces strict sparsity by keeping only the top K most active neurons per sample.
  • Alpha Aux Coefficient for the auxiliary loss, used to revive "dead" neurons during training.