infinite-width limit

**The Infinite-Width Limit** is a **theoretical idealization in deep learning where the number of neurons in each hidden layer is taken to infinity — revealing that at this limit, randomly initialized neural networks become Gaussian processes, and gradient descent training becomes kernel regression in the Neural Tangent Kernel space — providing tractable mathematical models of neural network behavior that yield convergence guarantees, generalization bounds, and insights into scaling laws** — while simultaneously highlighting that practical neural networks operate away from this limit, relying on finite-width feature learning that the infinite-width regime cannot capture. **What Happens at Infinite Width?** - **Gaussian Process at Initialization**: As hidden layer width n → ∞ (with independent random parameter initialization), by the Central Limit Theorem, the pre-activation distribution at each layer becomes Gaussian — and the function computed by the network becomes a Gaussian Process (GP) with covariance determined by the activation function and architecture. - **NTK Freezes During Training**: As shown by NTK theory (Jacot et al., 2018), as width → ∞ trained with small learning rates, the Neural Tangent Kernel remains constant throughout training. Training dynamics simplify to linear kernel regression. - **No Bad Local Minima**: In the infinite-width limit with overparameterization, gradient descent converges to a global minimum — the loss landscape becomes convex in function space. - **No Feature Learning**: In the kernel regime, the network's internal representations do not change — only the output head weights (effectively) change. The network does not learn progressively better features; it performs fixed-basis function approximation. **Mathematical Framework** | Quantity | Finite Width | Infinite Width | |----------|-------------|----------------| | **Pre-activations** | Correlated (non-Gaussian) | Independent Gaussians (CLT) | | **Network at init** | Complex non-GP function | Exact Gaussian Process | | **Training dynamics** | Nonlinear ODE in weight space | Linear ODE in function space (kernel regression) | | **Feature representations** | Evolve (feature learning) | Fixed (no representation learning) | | **Generalization** | Complex, architecture-dependent | RKHS norm regularization (kernel theory) | **Practical Relevance and Limitations** **Where the limit helps**: - **Initialization Design**: Infinite-width analysis motivates proper weight initialization (e.g., He initialization for ReLU, LeCun for tanh) to ensure stable signal propagation and full-rank NTK at training start. - **Architecture Comparison**: Comparing infinite-width GP/NTK kernels of different architectures provides insight into their inductive biases before training. - **Neural Scaling Theory**: Infinite-width limit is the starting point for understanding how performance scales with width — corrections at finite width produce scaling law models. - **Bayesian Deep Learning**: Infinite-width GP correspondence enables exact posterior inference tractable for small datasets. **Where the limit fails**: - **Feature Learning**: Real transformer and CNN performance relies on learning increasingly abstract and task-relevant representations — absent at infinite width. - **Sparse Representations**: Finite-width networks develop sparse features; infinite-width representations are dense Gaussian. - **Generalization on Large Data**: Kernel methods (infinite-width equivalent) often underperform finite-width networks on large-scale tasks — evidence they lack the inductive biases arising from finite-width training dynamics. - **Emergent Capabilities**: The emergent capabilities of large language models (in-context learning, chain-of-thought reasoning) have no analog in the infinite-width regime. **Research Frontiers** - **Mean-Field Theory**: Studies the 1/n corrections to the infinite-width limit — capturing first-order feature learning effects. - **Tensor Programs (Greg Yang)**: A unified framework computing the limiting behavior of any architecture as width → ∞, enabling systematic analysis of Transformers, LSTMs, and normalization layers. - **Maximal Update Parameterization (muP)**: Derived from infinite-width analysis — enables training hyperparameters (learning rate, initialization) to transfer cleanly from small to large width, used in practice for scaling up LLMs efficiently. The Infinite-Width Limit is **the theoretical microscope for deep learning** — an idealized mathematical lens that, while not accurately describing production neural networks, reveals the structural principles governing convergence, generalization, and architectural inductive biases, grounding practical design decisions in rigorous theory.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account