Neural Architecture Search by Estimation of Network Structure Distributions

TL;DR

This paper introduces a probabilistic approach to neural architecture search that models entire network structures, enabling the discovery of irregular, high-performance architectures beyond traditional block-based methods.

Contribution

It proposes a novel probabilistic representation for neural networks and an estimation of distribution algorithm to efficiently explore diverse, irregular architectures.

Findings

Discovered non-regular architectures with competitive accuracy.

Achieved efficient search without complex dataflows or training techniques.

Demonstrated interpretability and extensibility of the probabilistic model.

Abstract

The influence of deep learning is continuously expanding across different domains, and its new applications are ubiquitous. The question of neural network design thus increases in importance, as traditional empirical approaches are reaching their limits. Manual design of network architectures from scratch relies heavily on trial and error, while using existing pretrained models can introduce redundancies or vulnerabilities. Automated neural architecture design is able to overcome these problems, but the most successful algorithms operate on significantly constrained design spaces, assuming the target network to consist of identical repeating blocks. While such approach allows for faster search, it does so at the cost of expressivity. We instead propose an alternative probabilistic representation of a whole neural network structure under the assumption of independence between layer types. Our matrix of probabilities is equivalent to the population of models, but allows for discovery of structural irregularities, while being simple to interpret and analyze. We construct an architecture search algorithm, inspired by the estimation of distribution algorithms, to take advantage of this representation. The probability matrix is tuned towards generating high-performance models by repeatedly sampling the architectures and evaluating the corresponding networks, while gradually increasing the model depth. Our algorithm is shown to discover non-regular models which cannot be expressed via blocks, but are competitive both in accuracy and computational cost, while not utilizing complex dataflows or advanced training techniques, as well as remaining conceptually simple and highly extensible.