Improving Fine-Tuning with Latent Cluster Correction

The formation of salient clusters in the latent spaces of a neural network(NN) during training strongly impacts its final accuracy on classificationtasks. This paper proposes a novel fine-tuning method that boostsperformance by improving the quality of these latent clusters, using theLouvain community detection algorithm and a specifically designed lossfunction. We present preliminary results that demonstrate that this processyields an appreciable accuracy gain for classical NN architecturesfine-tuned on the CIFAR100 dataset.