II · THE IDEA · ARTIFICIAL INTELLIGENCE
Sparse Autoencoders for Dictionary Learning
▶ Listen · narrated
A single neuron often responds to several unrelated ideas at once. Dictionary learning with sparse autoencoders tries to recover the underlying features the network actually uses.
At a glance
- What it is
- Training a sparse autoencoder to rewrite activations as few active dictionary features
- Target
- Internal activations of language models
- Aim
- Features that are more monosemantic and human-interpretable than raw neurons
- Form
- Each activation approximated as a sparse linear combination of learned directions
Think of a busy shared drawer. Many tools are jammed into a few slots, so opening one slot shows a jumble. A sparse autoencoder is like building a larger rack of labelled hooks and insisting that any single job only takes a handful of hooks. The network’s activation is the jammed drawer; the learned hooks are dictionary features; the rule that only a few hooks may be used at once is sparsity. After training, you inspect which hooks light up for which sentences. Often a hook corresponds to something you can name — a kind of word, a pattern, a behaviour — more cleanly than the original jammed slot did.
In plain terms: the method copies an activation into a wider code, keeps most of that code near zero, and learns how to rebuild the activation from the few non-zero pieces. Those pieces are the features people then try to interpret.
Take activations x from a chosen site in a language model (for example a residual stream or MLP output). Train an autoencoder with encoder f and decoder g such that z = f(x) is high-dimensional and sparse, and x̂ = g(z) ≈ x. A common form uses a linear map plus nonlinearity for the encoder, an L1 or related penalty on z (or an explicit top-k), and a linear decoder whose columns are the dictionary atoms. The objective balances reconstruction error against sparsity of z.
The intended solution is an overcomplete dictionary that approximately undoes superposition: individual inputs activate few atoms, and many atoms align with human-interpretable concepts when inspected on data. Features live in activation space as decoder directions (and corresponding encoder read-ins), not as single original neurons. Failure modes include feature splitting, dead atoms, incomplete dictionaries, and residual polysemanticity. Hyperparameters — dictionary size, sparsity strength, site, and preprocessing such as centring — materially change what is recovered. Downstream use includes feature-level analysis, steering, and circuit description in the learned basis rather than in neuron indices.
Look closer
The bottleneck is deliberately overcomplete
The autoencoder’s hidden layer is usually wider than the activation vector it reconstructs. That extra room is the point: if many concepts are packed into a smaller space, an overcomplete dictionary gives them separate directions to occupy. Sparsity then pressures the model so that only a few of those directions turn on for any one input.
Reconstruction plus sparsity, not labels
Nothing in the training objective names the concepts. The encoder maps an activation to a sparse code; the decoder maps that code back; the loss asks for a faithful reconstruction while keeping most code entries near zero. Interpretability, when it appears, is a consequence of that pressure rather than a supervised teaching signal.
Features are directions, not neurons of the original net
A learned dictionary feature is a direction in activation space, recovered by the autoencoder, not necessarily a single unit in the language model. Reading the model through these features is a change of basis aimed at undoing superposition, not a simple inspection of existing weights.
The story
Language models do not store one clean concept per neuron. Empirical work on interpretability has repeatedly found units that fire for several unrelated ideas at once — a tangle often discussed under the names polysemanticity and superposition. If many features are squeezed into fewer dimensions than there are features, individual coordinates stop lining up with human-readable concepts. Looking at a neuron and hoping for a single meaning is then the wrong level of description.
Dictionary learning offers a different cut. The proposal is that an activation vector can be written, at least approximately, as a linear combination of a larger set of feature directions, with only a few coefficients large on any given input. Sparse autoencoders are one practical way to hunt for that dictionary. An encoder takes the activation, produces a wide code, and is trained so that most code entries stay near zero; a decoder rebuilds the activation from the surviving coefficients. Train this pair on many activations from a language model and the columns of the decoder become candidate features.
The editorial hope is modest and specific: those sparse features should be more interpretable building blocks than the raw neurons. A feature that lights up for a coherent pattern across prompts is easier to name, to track through a forward pass, and to test with interventions than a neuron that mixes several patterns. The method does not require a labelled concept dataset; the supervision is reconstruction and sparsity alone.
What the approach does not automatically guarantee is completeness or purity. A dictionary may recover many crisp features and still miss others, split one concept across several features, or leave residual polysemanticity in places. The quality of the story depends on the width of the dictionary, the sparsity pressure, the site in the network being decomposed, and how carefully the resulting features are checked against real prompts. The paper associated with this line of work reports that sparse autoencoders can find highly interpretable features in language models; that claim is an empirical finding about particular setups, not a theorem about every layer of every model.
Still, the move matters for how interpretability is done. Instead of reading the network only in its native coordinates, one learns a new coordinate system adapted to sparse structure in the activations. Circuits, probes, and editing methods can then be restated in terms of those features. The building blocks become linear directions that turn on sparsely — a description that matches the superposition picture more closely than a one-neuron-one-concept fantasy ever did.
Why it mattered then
By the time sparse autoencoders were applied seriously to language-model activations, it was already clear that raw units were a poor interface for explanation. Superposition suggested that useful features outnumbered available dimensions and would therefore share neurons. Dictionary learning was an old idea in signal processing; bringing it to residual streams and MLP activations was a way to test whether the features people believed were “in there” could be recovered as sparse linear components. The moment mattered because interpretability needed objects more stable and nameable than polysemantic neurons, and sparse autoencoders offered a training recipe rather than a hand-built probe for each concept.
Why it matters now
As models grow and are deployed in higher-stakes settings, the demand for internal accounts grows with them. Sparse dictionary features give researchers something concrete to monitor, ablate, and compare across checkpoints: a feature for a syntactic role, a factual cue, a behaviour one might want to damp. They also expose limits. If safety or reliability work depends on features that the autoencoder never found, the dictionary’s blind spots become operational risks. The method remains one tool among several, but it keeps shaping how people talk about what is linear, sparse, and readable inside these systems.
The surprising detail
The training signal never mentions meaning. The autoencoder only sees activation vectors and is judged on reconstruction and sparsity. When highly interpretable features fall out anyway, that is a fact about the structure of the activations — and about how well the dictionary matches it — not about concept labels being fed in. The interpretability is discovered, not supervised.
What is disputed
How fully sparse autoencoders recover the true features of a model remains an open empirical question. Dictionaries can be incomplete, fragmented, or still partly polysemantic depending on width, sparsity, layer choice, and evaluation method. Treat strong interpretability claims as tied to particular experiments, not as settled for language models in general.
Remember this
Sparse autoencoders re-express activations as a few dictionary directions at a time, aiming for features that are cleaner to read than polysemantic neurons.
Test yourself
Why might a sparse autoencoder need a hidden dimension larger than the activation it reconstructs, and what goes wrong if sparsity pressure is removed while keeping that width?
Width above the activation size gives the dictionary room to assign separate directions to features that were superimposed in the original space. Without sparsity pressure, the wide code can spread information thinly across many coordinates, and there is little force pushing each input toward a small set of active, potentially monosemantic features — so the overcomplete basis need not become an interpretable dictionary.
Go deeper
- [2301.04712] An analytic formula for entraining CAPE in mid-latitude storm environments · arxiv.org
- [2309.08600] Sparse Autoencoders Find Highly Interpretable Features in Language Models · arxiv.org
Image: Original diagram, The Daily Triptych. Licence: Original work. Source.