ENERGY MODELS FOR PRUNING NEURAL NETWORKS

Deep neural networks (DNNs) have achieved significant performance improvement in image classification tasks.A DNN utilizes an enormous number of trainable parameters and non-linear functions to learn a mapping from a set of input images to a set of target classes. However, the increasing number of parameters, called the overparameterization problem, can create major challenges in implementation of these models on resource limited devices and can encourage overfitting of the model. Pruning less important trainable parameters in DNNs is one of the major established methods to defeat the above challenges. Pruning methods are also applicable on small models by design to further reduce their size. This thesis proposes the concept of energy models for pruning convolutional filters and hidden units with corresponding incoming and outgoing connections and bias terms for image classification tasks. Specifically, two energy models, namely EPruning and IPruning, are proposed. The first approach uses the Hamiltonian of a given DNN to form an energy loss minimization problem. The loss function is defined based on the difference between the energy of target class and the closest energy of the competitive classes. The second approach uses entropy of feature maps, Kullback–Leibler (KL) divergence between convolution filters in a layer, and activation level of hidden units to form an Ising energy-based minimization problem for detecting less important and redundant parameters in the network. Despite most pruning methods, which are based on defining a pruning threshold per layer, the proposed methods are data-driven and perform a network-wise pruning of the trainable parameters. We define pruning as searching for a sub-network of a given DNN, which has the lowest energy loss value based on a defined energy loss function. Since this is an NP-hard combinatorial optimization problem, a population-based global optimization framework is proposed to minimize the pruning objective functions. The proposed framework is evaluated on different flavors of DNNs (e.g. ResNets, AlexNet, and SqueezeNet) and datasets (e.g. CIFAR-10, CIFAR-100, and ImageNet) for image classification tasks.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.