Composition vs Superposition in Neural Networks

Composition builds complex features by combining simpler ones across layers; superposition packs multiple features into the same neurons or directions when they can be represented economically. Neural networks use both: depth supports feature construction, while distributed representations let limited activation space carry more information than a one-neuron-per-feature scheme would allow.

What distributed representations mean

A representation is the pattern of activations a network uses to encode information. In a distributed representation, a concept is not necessarily stored in one dedicated neuron. Instead, it may correspond to a direction or pattern across many units. A unit can contribute to several patterns, and a pattern can involve many units.

For example, a language model’s hidden state may contain information about a word’s grammatical role, topic, and the likely continuation of a sentence at the same time. Those properties need not occupy separate, cleanly isolated slots. They are encoded in a vector whose components jointly carry information.

The terms composition and superposition describe different aspects of this arrangement. Composition concerns how features are assembled into more complex ones. Superposition concerns how multiple features share representational dimensions. They are compatible, not competing explanations.

Composition: building features from features

Composition is the formation of higher-level representations from lower-level ones. In a layered network, early computations can detect simple patterns, while later computations combine their outputs into more complex patterns. In image recognition, for instance, a model may build from local edge-like responses to textures and parts, then to object-level evidence. This is a useful conceptual picture, not a guarantee that every network has a neat, human-readable hierarchy.

In language models, a token representation can be transformed through successive layers. Attention can bring information from other positions into a token’s current representation; nonlinear transformations then update it. Across layers, the resulting state can reflect relationships among words and broader context. A feature such as “the subject of this clause” can depend on several simpler cues and their interaction.

Composition does not require each intermediate feature to have a dedicated neuron. A feature can be distributed across many units, and a later layer can compute from that distributed pattern. Nor does “deeper” automatically mean “more abstract”: what a layer represents depends on the model, its training, and the input.

Superposition: sharing representational space

Superposition occurs when a model represents more features than it has dimensions available for clean, mutually independent storage. If each feature had its own orthogonal direction, a vector space of dimension d could hold at most d such directions. But a network can use directions that overlap, allowing more than d features to be represented approximately.

A simple analogy is a page with several faint drawings overlaid. Each drawing is recoverable when considered in context, but the page does not give every drawing its own separate area. In a neural network, the overlap is mathematical: feature vectors can point in related directions, and their activations add together. A downstream computation may still distinguish them when the patterns are sufficiently separable or the relevant context is available.

Sharing has a cost. Overlapping features can interfere: activating one may make another harder to recover, or a computation intended for one feature may respond to another. A model therefore faces a trade-off between representing many features compactly and keeping them easy to distinguish. The amount of interference depends on feature correlations, activation patterns, model architecture, and training.

How networks reuse and pack features

Networks reuse features because the same intermediate pattern can help with many later tasks. An edge-like visual response may contribute to recognizing different objects; a syntactic cue may help predict several kinds of words. This reuse is one reason learned representations can be more economical than storing a separate complete solution for every possible input.

Feature packing is not necessarily a conscious design choice. Training adjusts weights to reduce an objective, and shared or overlapping representations can emerge when they are useful under the model’s capacity constraints. In mechanistic interpretability, researchers often study this possibility by treating features as directions in activation space rather than assuming neurons correspond one-to-one with concepts.

• Composition: combines signals or features to compute a more complex feature.
• Reuse: allows an existing feature to support several contexts or downstream computations.
• Superposition: allows multiple features to share dimensions, often with some interference.

These can happen together. A layer can combine several overlapping features into a new representation, which later layers reuse or combine again.

Why neurons are not always concepts

A common misconception is that one neuron equals one meaningful feature. Some neurons do respond selectively to recognizable patterns, but this is not a general rule. A neuron is one coordinate in a chosen representation; changing the coordinate system can change which individual units look interpretable while leaving the overall represented information intact.

Conversely, finding a neuron that responds to a concept does not prove that the concept is stored only there. The information may also be distributed across other units, and the neuron’s response may depend on context. Researchers use methods such as activation interventions, feature visualization, and sparse autoencoders to investigate representations, but each method has assumptions and limitations. A discovered feature should not automatically be treated as a complete account of what the network computes.

What the comparison does and does not explain

Composition is mainly about computational structure: how one operation builds on the results of others. Superposition is mainly about representational capacity: how many patterns share a finite space. Composition can occur without strong superposition, and superposition can occur in a layer whose job is not obviously hierarchical.

The distinction also clarifies why model size alone does not tell us exactly how a network stores knowledge. More dimensions can provide room for more independent features, but training and data determine which features are useful and how they are organized. Likewise, a compact representation is not automatically robust or understandable. Overlap may save space yet make interpretation and reliable feature-specific control harder.

These ideas help explain observed behavior, but they are not a complete theory of neural networks. Real models use nonlinear transformations, attention, normalization, and other architectural details. Their features may be context-dependent, change across layers, and resist a single fixed decomposition.

Frequently Asked Questions

Is composition the same as superposition?
No. Composition combines features to form other features; superposition represents multiple features using overlapping directions or dimensions. A network can do both at once.

Does superposition mean information is lost?
Not necessarily. Overlap can make features harder to distinguish, but useful information may still be recoverable from the full activation pattern and context. Interference becomes a problem when the relevant computations cannot separate the signals well enough.

Can one neuron represent multiple features?
Yes. A unit can respond to several patterns, and different features can share its contribution. But a single neuron’s activity usually does not reveal the whole distributed representation.

Why not give every feature its own neuron?
That would require a separate dimension for every independent feature. Networks have finite capacity, and many features may be inactive in different contexts or share useful structure, making overlapping representations an efficient trade-off.

Does a deeper network always have more compositional representations?
Depth provides more stages for successive computation, but it does not ensure a clean hierarchy of interpretable features. The learned organization depends on architecture, training, and data.

Related

• How to Clean Gaming Headset Ear Pads Safely
• First Light Trophy List and How to Earn Them
• Thieves Dressed as Apple Employees: What Happened?