Sparse Autoencoder

The defining mechanism that forces only a tiny fraction of hidden neurons to activate for any given input. This is typically enforced through an L1 penalty on the hidden layer activations added to the reconstruction loss. The sparsity objective pushes the model to learn a distributed, factorial representation where each feature captures a single, independent concept rather than a dense, entangled mixture. The target sparsity level is a critical hyperparameter: too sparse and the model fails to reconstruct; too dense and features remain polysemantic.

A two-component structure consisting of an encoder that maps the input activation vector x to a higher-dimensional hidden representation h via learned weights W_enc and a bias, followed by a ReLU nonlinearity: h = ReLU(W_enc * x + b_enc). The decoder then reconstructs the original activation from this sparse code: x̂ = W_dec * h + b_dec. The decoder weight matrix columns (or encoder rows) serve as the feature dictionary, with each direction in activation space corresponding to a distinct, interpretable feature.

The primary goal of training an SAE on model activations is to disentangle superposed features into individually interpretable components. In the superposition hypothesis, a model represents more concepts than it has dimensions by encoding them in overlapping directions. A trained SAE's hidden neurons become monosemantic — each neuron fires exclusively for a single, human-understandable concept (e.g., 'Arabic text', 'DNA sequences', 'praise language'). This transforms an opaque activation vector into an explicit, inspectable feature vector.

The SAE is trained to minimize the mean squared error between the original activation x and its reconstruction x̂, while simultaneously minimizing the sparsity penalty. The combined loss function is: L(x) = ||x - x̂||²₂ + λ * ||h||₁, where λ controls the trade-off between fidelity and sparsity. A well-trained SAE achieves high reconstruction fidelity with very few active features, demonstrating that the model's internal representations can be compressed into a sparse, interpretable code without significant information loss.

The decoder weight matrix W_dec functions as an overcomplete dictionary of feature vectors. Because the hidden dimension is typically 4x to 16x larger than the input dimension, the SAE learns a rich, redundant set of basis vectors. This overcompleteness is essential for capturing the full diversity of features present in the model's activations. The learned dictionary can be directly inspected by examining which input tokens or patterns maximally activate each dictionary element, enabling feature visualization and manual interpretation.

A common failure mode where some hidden neurons never activate across the entire training dataset, effectively becoming dead features. This occurs when the sparsity penalty is too aggressive or when the ReLU threshold pushes neurons permanently to zero. Monitoring the fraction of dead neurons is a critical diagnostic metric during training. Techniques like neuron resampling — reinitializing dead neurons to the encoder vectors of highly active examples — are used to revive them and ensure the dictionary's capacity is fully utilized.