Semantic and Visual Similarities for Efficient Knowledge Transfer in CNN Training

TL;DR

This paper proposes a method to improve transfer learning in CNNs by leveraging semantic and visual similarities between classes, resulting in faster training and better performance with limited data.

Contribution

It introduces a novel approach that combines semantic and visual similarities to transfer supplementary weights, enhancing CNN fine-tuning efficiency.

Findings

Faster training convergence with similarity-based weight transfer.

Improved initial performance in limited data scenarios.

Superior long-term accuracy when training data is scarce.

Abstract

In recent years, representation learning approaches have disrupted many multimedia computing tasks. Among those approaches, deep convolutional neural networks (CNNs) have notably reached human level expertise on some constrained image classification tasks. Nonetheless, training CNNs from scratch for new task or simply new data turns out to be complex and time-consuming. Recently, transfer learning has emerged as an effective methodology for adapting pre-trained CNNs to new data and classes, by only retraining the last classification layer. This paper focuses on improving this process, in order to better transfer knowledge between CNN architectures for faster trainings in the case of fine tuning for image classification. This is achieved by combining and transfering supplementary weights, based on similarity considerations between source and target classes. The study includes a comparison between semantic and content-based similarities, and highlights increased initial performances and training speed, along with superior long term performances when limited training samples are available.