Controlled Sparsity via Constrained Optimization or: How I Learned to Stop Tuning Penalties and Love Constraints
This work addresses the challenge of direct sparsity control for researchers and practitioners in deep learning, offering a more efficient alternative to penalty-based methods, though it builds incrementally on existing gate mechanisms.
The paper tackles the problem of controlling sparsity levels in neural networks without expensive trial-and-error tuning, by formulating a constrained optimization approach that reliably achieves pre-determined sparsity targets without compromising predictive performance, as validated on datasets like CIFAR-10, CIFAR-100, TinyImageNet, and ImageNet using models such as WideResNet and ResNet.
The performance of trained neural networks is robust to harsh levels of pruning. Coupled with the ever-growing size of deep learning models, this observation has motivated extensive research on learning sparse models. In this work, we focus on the task of controlling the level of sparsity when performing sparse learning. Existing methods based on sparsity-inducing penalties involve expensive trial-and-error tuning of the penalty factor, thus lacking direct control of the resulting model sparsity. In response, we adopt a constrained formulation: using the gate mechanism proposed by Louizos et al. (2018), we formulate a constrained optimization problem where sparsification is guided by the training objective and the desired sparsity target in an end-to-end fashion. Experiments on CIFAR-{10, 100}, TinyImageNet, and ImageNet using WideResNet and ResNet{18, 50} models validate the effectiveness of our proposal and demonstrate that we can reliably achieve pre-determined sparsity targets without compromising on predictive performance.