I am looking for enhancing the accuracy of a deep learning model using some techniques to compress neural networks. Could someone suggest to me some of the tools/techniques, please?
Hello Driss Kaouthar, I would recommend this paper on EfficientNets: Preprint EfficientNet: Rethinking Model Scaling for Convolutional Neu...
Normally, reducing the depth and width of a NN would lead to worse accuracy. The authors of this paper therefore have found an optimal combination. This gives the best accuracy with a small size of the NN. Cheers, Raoul
you can design an over-parameterized model, train it to convergence, it will make it easier to find a compressed network that when trained from scratch, maintains the same performance. .. in general, the compression degrades the accuracy, but by retraining the model you can re-achieve the over-parameterized model's performance.