Improving Fine-Tuning with Latent Cluster Correction

Cédric Ho Thanh

Keywords: Latent space, fine-tuning, deep neural network, clustering

TL;DR: Fine-tuning method that refines clusters of latent representations to improve classification performance

Abstract:

The formation of salient clusters in the latent spaces of a neural network
(NN) during training strongly impacts its final accuracy on classification
tasks. This paper proposes a novel fine-tuning method that boosts
performance by improving the quality of these latent clusters, using the
Louvain community detection algorithm and a specifically designed loss
function. We present preliminary results that demonstrate that this process
yields an appreciable accuracy gain for classical NN architectures
fine-tuned on the CIFAR100 dataset.

Submission Number: 68