Sparse Autoencoder

The hidden layer contains more features than input dimensions, creating a higher-dimensional representation space. This overcompleteness allows the model to learn an expansive dictionary of monosemantic features that can disentangle the superimposed representations found in standard neural networks. Unlike undercomplete autoencoders that learn compressed representations, overcompleteness provides the capacity needed to separate overlapping concepts.

The loss function includes an L1 regularization term on the hidden activations, which drives most feature values to exactly zero for any given input. This enforces the core property: only a tiny fraction of available features activate simultaneously. The sparsity coefficient λ controls the trade-off between reconstruction fidelity and feature selectivity:

• High λ: Fewer active features, more interpretable but worse reconstruction
• Low λ: More active features, better reconstruction but less monosemanticity

The encoder and decoder weight matrices can be tied (transposes of each other) or untied (independently learned). Untied weights offer greater representational flexibility, allowing the encoder to detect features and the decoder to reconstruct them using different basis directions. In mechanistic interpretability, untied weights are standard because they enable the decoder vectors to serve as the interpretable feature directions visualized during analysis.

Each decoder weight vector corresponds to a single feature in the learned dictionary. By examining which input patterns maximally activate a given feature, researchers can identify the monosemantic concept it represents. Visualization techniques include:

• Maximally activating examples: Real inputs that trigger the feature most strongly
• Synthetic optimization: Generating inputs that maximize feature activation through gradient ascent
• Decoder vector inspection: Directly analyzing the weight pattern itself

The primary training signal is the mean squared error between the original input activations and the reconstructed output. This forces the sparse autoencoder to preserve the information content of the original representation while decomposing it into a sparse combination of interpretable features. The reconstruction fidelity serves as a direct measure of how completely the learned dictionary captures the model's internal representations.

After training, the sparse autoencoder serves as a diagnostic lens into the target model. Researchers analyze which features activate for specific inputs to understand how the model represents concepts internally. Key analyses include:

• Feature co-occurrence: Which concepts the model associates together
• Feature circuits: How features compose across layers to implement algorithms
• Ablation studies: What happens when specific features are zeroed out during inference

The theory that neural networks represent more independent features than they have dimensions by encoding them in overlapping, nearly orthogonal directions. This explains why individual neurons respond to multiple unrelated inputs. Sparse autoencoders directly address this by learning an overcomplete basis that disentangles these superimposed features into separate, interpretable directions.

A neuron that activates for multiple unrelated input patterns, complicating direct interpretation. For example, a single neuron might fire for both academic citations and URLs. This phenomenon motivated the development of sparse autoencoders, which decompose these mixed activations into distinct monosemantic features that each correspond to a single human-understandable concept.

A sparse coding approach that finds an overcomplete basis of interpretable feature directions from model activations. The technique learns a dictionary where:

• Each dictionary element represents a distinct concept
• Activations are reconstructed as sparse linear combinations
• The sparsity penalty forces the model to use few features per input This is the mathematical foundation upon which sparse autoencoders are built.

The property of a neuron or feature that activates exclusively for a single, human-interpretable concept. Achieving monosemanticity is the ideal decomposition goal of mechanistic interpretability. Sparse autoencoders are explicitly designed to produce monosemantic features from polysemantic neurons by enforcing a sparsity constraint that prevents features from activating for multiple unrelated inputs.

An optimization-based method that generates synthetic inputs to maximally activate a specific neuron, channel, or learned feature. When applied to sparse autoencoder features, this technique reveals what pattern each dictionary element has learned to detect. The resulting visualizations provide direct evidence that the decomposed features correspond to coherent, interpretable concepts rather than arbitrary directions.

The core data pathway in a transformer where each layer reads from and writes additive updates to a running hidden state. Sparse autoencoders are typically trained on residual stream activations to decompose the model's internal representations at specific layers. Understanding the residual stream is essential because it is the primary object of analysis for most model probing and decoding techniques.