Using Eigenvector Projections to Search Neural Network Architectures

Summary

The article describes a compact neural architecture search method for multilayer perceptrons, varying two settings: the number of hidden layers and the width shared by those layers. It constructs candidate networks across a two-dimensional grid, evaluates each on historical EURJPY four-hour data, and accumulates prediction deviation against a target. The benchmark matrix is normalized, its covariance matrix is decomposed into eigenvectors and eigenvalues, and projections are used to estimate preferable and weaker settings.

A reported run identifies six layers and size two as the estimated ideal settings, while nine layers and size four are estimated as worst. This is a single illustrative result, not evidence of general predictive or trading performance. The search leaves out other influential choices, including activation functions and initial weights and biases, because expanding the search would increase computational demands. The method's estimates depend on the selected data, benchmark, normalization, and interpretation of projected error.

Key ideas

• The search varies hidden-layer count and common layer width across a two-dimensional candidate grid.
• Candidate networks are benchmarked by accumulating absolute prediction deviations from target values.
• The benchmark matrix is normalized and analyzed through covariance eigenvectors and projections.
• The reported preferred architecture is specific to one EURJPY four-hour example and does not establish general trading performance.
• Activation functions and initial weights and biases are excluded from the search.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.