Manifold Learning and the Intrinsic Geometry of Neural Representations

The assumption that neural networks learn high-dimensional representations is fundamentally misleading—they learn low-dimensional structures embedded in high-dimensional spaces, and this distinction reshapes everything we think we know about how learning actually works.

When we observe a trained neural network, we see activations scattered across thousands or millions of dimensions. The natural instinct is to treat this as the network's true representational space. But the data doesn't live there. The meaningful structure—the patterns that matter for computation—concentrates on lower-dimensional manifolds. A network trained on images doesn't actually need ten thousand dimensions to represent visual concepts. It carves out a surface, a fold, a twisted geometry in that vast space, and all the computation happens on that surface. Everything else is noise, padding, redundancy.

This is where most theoretical accounts go wrong. They describe neural computation as if it occurs in the ambient space—the full dimensionality we observe. But the actual geometry is intrinsic. The distances that matter, the neighborhoods that define similarity, the transformations that preserve meaning: these are all defined on the manifold itself, not in the surrounding space. A network learning to classify handwritten digits doesn't learn a function from ℝ²⁸⁸ to ten classes. It learns a function from a low-dimensional manifold (perhaps twenty or thirty dimensions) to ten classes. The high-dimensional input is just how we happen to present the data.

The implications are severe. First, it explains why neural networks generalize at all. If representations truly occupied high-dimensional space densely, interpolation between training examples would be meaningless. But manifolds are sparse. They have structure. Generalization becomes possible because the network learns the geometry of the manifold, not a lookup table in ambient space. When you show it a new point on the manifold, it knows how to respond because it understands the manifold's topology, not because it memorized nearby examples.

Second, it reframes what we mean by "learning." We typically think of learning as fitting a function. But from the manifold perspective, learning is discovering the intrinsic geometry of the data distribution. The network is performing dimensionality reduction and function approximation simultaneously. These aren't separate processes. The hidden layers progressively reveal the manifold structure. Early layers might work in higher dimensions; later layers operate on increasingly refined, lower-dimensional representations. The network is peeling away irrelevant dimensions, exposing the true structure underneath.

Third, and most consequentially, it changes how we should think about interpretability and abstraction. If we want to understand what a network has learned, we cannot simply examine weight matrices or activation patterns in isolation. We must understand the manifold geometry. What is the intrinsic dimensionality at each layer? How do the manifolds at different layers relate to one another? What topological features—holes, twists, self-intersections—characterize the learned representations? These are geometric questions, not algebraic ones.

This perspective also illuminates why certain architectural choices work. Attention mechanisms, for instance, can be understood as adaptive metric learning on manifolds. They compute which points on the manifold are relevant to one another, dynamically reshaping the geometry based on context. Residual connections preserve manifold structure across layers, allowing information to flow along the intrinsic geometry rather than forcing it through bottlenecks.

The hard problem remains unsolved: we lack principled methods to characterize manifold geometry in real networks. We have tools from differential geometry and topological data analysis, but applying them to high-dimensional neural representations remains technically difficult. We can measure intrinsic dimensionality. We can estimate local curvature. But we cannot yet read off the full topological story.

Yet the conceptual shift is already valuable. It reorients research toward the right questions. Instead of asking "what do these activations represent," we should ask "what is the geometry of this representation space." Instead of treating dimensionality reduction as a post-hoc analysis tool, we should recognize it as central to understanding learning itself. The network is not computing in ten thousand dimensions. It is computing on a manifold, and that manifold is where the real work happens.