cs.LGNov 18, 2026

Bias-Only Training Dynamics in ReLU Networks

Authors: Shreyas Chaudhari, Jose M. F. Moura

Organizations: Department of Electrical and Computer Engineering, Carnegie Mellon University

Abstract

Neural networks with randomly initialized, fixed weights and trainable biases are universal approximators, yet the dynamics of bias-only training remain largely unexplored. We study these dynamics in one-hidden-layer ReLU networks through activation codes, the binary patterns of active neurons that partition the input space into convex polyhedral regions. We show that the gradients with respect to the biases depend on each input only through its activation code, and hence the training dynamics can be characterized with these codes. In particular, as long as the activation codes remain fixed, training reduces to a linear least-squares problem with the codes as features. Training can then be understood as a sequence of least-squares problems, with transitions occurring when neurons change their activation states. We show that the fraction of switching neurons vanishes as the network widens, and the training dynamics approach those of linear least squares on the initial activation codes in the infinite-width limit. Our experiments support these theoretical findings, showing that while activation switches can noticeably affect training in narrower networks, their relative frequency and the associated changes in the dynamics diminish as the width increases.

Figures & tables

Explore similar work

Jul 3, 2026

Implicit Bias of SGD in Multivariate ReLU Networks: Effective Width Collapse

We study the implicit bias of noisy stochastic gradient descent in training wide two-layer ReLU networks for multivariate regression. In a mean-field regime, the training dynamics are approximated by a Wasserstein gradient flow that converges to a unique stationary measure. We characterize the structure of this stationary measure and the predictor it represents. We show that, despite the network being infinitely overparameterized, the learned predictor admits an effectively finite representation: the input weights and biases align along finitely many directions, leading to an effective width collapse. In particular, the solution function is continuous piecewise affine, with affine regions determined by the cells of a finite hyperplane arrangement. The number of learned directions, and hence hyperplanes, is bounded above by 2P−12\mathcal{P}-1, where P\mathcal{P} denotes the number of linear dichotomies realizable on the training inputs. We further establish a non-redundancy property of the learned representation by proving that each learned direction induces a unique ternary activation pattern on the training data. Consequently, the complexity of the learned predictor is governed by the combinatorial geometry of the training data.
Apr 9, 2026

Mathematical analysis of one-layer neural network with fixed biases, a new activation function and other observations

We analyze a simple one-hidden-layer neural network with ReLU activation functions and fixed biases, with one-dimensional input and output. We study both continuous and discrete versions of the model, and we rigorously prove the convergence of the learning process with the L2L^2 squared loss function and the gradient descent procedure. We also prove the spectral bias property for this learning process. Several conclusions of this analysis are discussed; in particular, regarding the structure and properties that activation functions should possess, as well as the relationships between the spectrum of certain operators and the learning process. Based on this, we also propose an alternative activation function, the full-wave rectified exponential function (FReX), and we discuss the convergence of the gradient descent with this alternative activation function.
Feb 22, 2026

Why ReLU? A Bit-Model Dichotomy for Deep Network Training

Theoretical analyses of Empirical Risk Minimization (ERM) are standardly framed within the Real-RAM model of computation. In this setting, training even simple neural networks is known to be ∃R\exists \mathbb{R}-complete -- a complexity class believed to be harder than NP, that characterizes the difficulty of solving systems of polynomial inequalities over the real numbers. However, this algebraic framework diverges from the reality of digital computation with finite-precision hardware. In this work, we analyze the theoretical complexity of ERM under a realistic bit-level model (ERMbit\mathsf{ERM}_{\text{bit}}), where network parameters and inputs are constrained to be rational numbers with polynomially bounded bit-lengths. Under this model, we reveal a sharp dichotomy in tractability governed by the network's activation function. We prove that for deep networks with {\em any} polynomial activations with rational coefficients and degree at least 22, the bit-complexity of training is severe: deciding ERMbit\mathsf{ERM}_{\text{bit}} is #P\#P-Hard, hence believed to be strictly harder than NP-complete problems. Furthermore, we show that determining the sign of a single partial derivative of the empirical loss function is intractable (unlikely in BPP), and deciding a specific bit in the gradient is #P\#P-Hard. This provides a complexity-theoretic perspective for the phenomenon of exploding and vanishing gradients. In contrast, we show that for piecewise-linear activations such as ReLU, the precision requirements remain manageable: ERMbit\mathsf{ERM}_{\text{bit}} is contained within NP (specifically NP-complete), and standard backpropagation runs in polynomial time. Our results demonstrate that finite-precision constraints are not merely implementation details but fundamental determinants of learnability.