Predicting classification performance of convolutional neural networks

Abstract

In recent years, neural networks have garnered significant interest due to their capacity to learn from data examples pertaining to various tasks, subsequently demonstrating proficiency in addressing these challenges. The efficacy of a neural network not only hinges on the caliber and volume of the data samples but also crucially on the architecture of the network itself. To date, a concrete theoretical approach to discern the optimal structure of neural networks remains elusive. As a result, much of the architectural design relies heavily on the expertise and intuition of the designer. Furthermore, empirical validation is indispensable to ensure the network operates as anticipated. This design paradigm often culminates in extended experimental durations. In an endeavor to mitigate this challenge, we explore the feasibility of predicting image classification accuracy post-training by leveraging information gleaned from the neural network's initial state, specifically for a designated image classification task. Employing multiple regression analyses, grounded in our antecedent knowledge, we aim to project the image classification accuracy subsequent to 150 training epochs. Our empirical findings attest to the viability of this approach, elucidating that the classification accuracy at the terminal training epochs can be forecasted with minimal margin of error.