Dataset Information

Locally linear attributes of ReLU neural networks.

ABSTRACT: A ReLU neural network functions as a continuous piecewise linear map from an input space to an output space. The weights in the neural network determine a partitioning of the input space into convex polytopes, where each polytope is associated with a distinct affine mapping. The structure of this partitioning, together with the affine map attached to each polytope, can be analyzed to investigate the behavior of the associated neural network. We investigate simple problems to build intuition on how these regions act and both how they can potentially be reduced in number and how similar structures occur across different networks. To validate these intuitions, we apply them to networks trained on MNIST to demonstrate similarity between those networks and the potential for them to be reduced in complexity.

SUBMITTER: Sattelberg B

PROVIDER: S-EPMC10701692 | biostudies-literature | 2023

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Locally linear attributes of ReLU neural networks.

Sattelberg Ben B Cavalieri Renzo R Kirby Michael M Peterson Chris C Beveridge Ross R

Frontiers in artificial intelligence 20231123

A ReLU neural network functions as a continuous piecewise linear map from an input space to an output space. The weights in the neural network determine a partitioning of the input space into convex polytopes, where each polytope is associated with a distinct affine mapping. The structure of this partitioning, together with the affine map attached to each polytope, can be analyzed to investigate the behavior of the associated neural network. We investigate simple problems to build intuition on h ...[more]

PMID: 38075385

Similar Datasets

Project description:Understanding the learning dynamics and inductive bias of neural networks (NNs) is hindered by the opacity of the relationship between NN parameters and the function represented. Partially, this is due to symmetries inherent within the NN parameterization, allowing multiple different parameter settings to result in an identical output function, resulting in both an unclear relationship and redundant degrees of freedom. The NN parameterization is invariant under two symmetries: permutation of the neurons and a continuous family of transformations of the scale of weight and bias parameters. We propose taking a quotient with respect to the second symmetry group and reparametrizing ReLU NNs as continuous piecewise linear splines. Using this spline lens, we study learning dynamics in shallow univariate ReLU NNs, finding unexpected insights and explanations for several perplexing phenomena. We develop a surprisingly simple and transparent view of the structure of the loss surface, including its critical and fixed points, Hessian, and Hessian spectrum. We also show that standard weight initializations yield very flat initial functions, and that this flatness, together with overparametrization and the initial weight scale, is responsible for the strength and type of implicit regularization, consistent with previous work. Our implicit regularization results are complementary to recent work, showing that initialization scale critically controls implicit regularization via a kernel-based argument. Overall, removing the weight scale symmetry enables us to prove these results more simply and enables us to prove new results and gain new insights while offering a far more transparent and intuitive picture. Looking forward, our quotiented spline-based approach will extend naturally to the multivariate and deep settings, and alongside the kernel-based view, we believe it will play a foundational role in efforts to understand neural networks. Videos of learning dynamics using a spline-based visualization are available at http://shorturl.at/tFWZ2.

Project description:Spiking neural networks (SNNs) are a computational tool in which the information is coded into spikes, as in some parts of the brain, differently from conventional neural networks (NNs) that compute over real-numbers. Therefore, SNNs can implement intelligent information extraction in real-time at the edge of data acquisition and correspond to a complementary solution to conventional NNs working for cloud-computing. Both NN classes face hardware constraints due to limited computing parallelism and separation of logic and memory. Emerging memory devices, like resistive switching memories, phase change memories, or memristive devices in general are strong candidates to remove these hurdles for NN applications. The well-established training procedures of conventional NNs helped in defining the desiderata for memristive device dynamics implementing synaptic units. The generally agreed requirements are a linear evolution of memristive conductance upon stimulation with train of identical pulses and a symmetric conductance change for conductance increase and decrease. Conversely, little work has been done to understand the main properties of memristive devices supporting efficient SNN operation. The reason lies in the lack of a background theory for their training. As a consequence, requirements for NNs have been taken as a reference to develop memristive devices for SNNs. In the present work, we show that, for efficient CMOS/memristive SNNs, the requirements for synaptic memristive dynamics are very different from the needs of a conventional NN. System-level simulations of a SNN trained to classify hand-written digit images through a spike timing dependent plasticity protocol are performed considering various linear and non-linear plausible synaptic memristive dynamics. We consider memristive dynamics bounded by artificial hard conductance values and limited by the natural dynamics evolution toward asymptotic values (soft-boundaries). We quantitatively analyze the impact of resolution and non-linearity properties of the synapses on the network training and classification performance. Finally, we demonstrate that the non-linear synapses with hard boundary values enable higher classification performance and realize the best trade-off between classification accuracy and required training time. With reference to the obtained results, we discuss how memristive devices with non-linear dynamics constitute a technologically convenient solution for the development of on-line SNN training.

Dataset Information

Locally linear attributes of ReLU neural networks.

Publications

Locally linear attributes of ReLU neural networks.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets