The Lottery Ticket Hypothesis (LTH) suggests there exists a sparse LTH mask and weights that achieve the same generalization performance as the dense model while using significantly fewer parameters. However, finding a LTH solution is computationally expensive, and a LTH sparsity mask does not generalize to other random weight initializations. Recent work has suggested that neural networks trained from random initialization find solutions within the same basin modulo permutation, and proposes a method to align trained models within the same loss basin. We hypothesize that misalignment of basins is the reason why LTH masks do not generalize to new random initializations and propose permuting the LTH mask to align with the new optimization basin when performing sparse training from a different random init. We empirically show a significant increase in generalization when sparse training from random initialization with the permuted mask as compared to using the non-permuted LTH mask, on multiple datasets (CIFAR-10/100 & ImageNet) and models (VGG11 & ResNet20/50).

Dynamic Sparse Training (DST) methods achieve state-of-the-art results in sparse neural network training, matching the generalization of dense models while enabling sparse training and inference. Although the resulting models are highly sparse and theoretically less computationally expensive, achieving speedups with unstructured sparsity on real-world hardware is challenging. In this work, we propose a sparse-to-sparse DST method, Structured RigL (SRigL), to learn a variant of fine-grained structured N:M sparsity by imposing a constant fan-in constraint. Using our empirical analysis of existing DST methods at high sparsity, we additionally employ a neuron ablation method which enables SRigL to achieve state-of-the-art sparse-to-sparse structured DST performance on a variety of Neural Network (NN) architectures. Using a 90% sparse linear layer, we demonstrate a real-world acceleration of 3.4×/2.5× on CPU for online inference and 1.7×/13.0× on GPU for inference with a batch size of 256 when compared to equivalent dense/unstructured (CSR) sparse layers, respectively.

In recent years, Dynamic Sparse Training (DST) has emerged as an alternative to post-training pruning for generating efficient models. In principle, DST allows for a much more memory efficient training process, as it maintains sparsity throughout the entire training run. However, current DST implementations fail to capitalize on this. Because sparse matrix multiplication is much less efficient than dense matrix multiplication on GPUs, most implementations simulate sparsity by masking weights. In this paper, we leverage recent advances in semi-structured sparse training to apply DST in the domain of classification with large output spaces, where memory-efficiency is paramount. With a label space of possibly millions of candidates, the classification layer alone will consume several gigabytes of memory. Switching from a dense to a fixed fan-in sparse layer updated with sparse evolutionary training (SET); however, severely hampers training convergence, especially at the largest label spaces. We find that the gradients fed back from the classifier into the text encoder make it much more difficult to learn good input representations, despite using a dense encoder. By employing an intermediate layer or adding an auxiliary training objective, we recover most of the generalisation performance of the dense model. Overall, we demonstrate the applicability of DST in a challenging domain, characterized by a highly skewed label distribution, that lies outside of DST's typical benchmark datasets, and enable end-to-end training with millions of labels on commodity hardware.

Sparse Neural Networks (NNs) can match the generalization of dense NNs using a fraction of the compute/storage for inference, and have the potential to enable efficient training. However, naively training unstructured sparse NNs from random initialization results in significantly worse generalization, with the notable exceptions of Lottery Tickets (LTs) and Dynamic Sparse Training (DST). In this work, we attempt to answer: (1) why training unstructured sparse networks from random initialization performs poorly and; (2) what makes LTs and DST the exceptions? We show that sparse NNs have poor gradient flow at initialization and propose a modified initialization for unstructured connectivity. Furthermore, we find that DST methods significantly improve gradient flow during training over traditional sparse training methods. Finally, we show that LTs do not improve gradient flow, rather their success lies in re-learning the pruning solution they are derived from — however, this comes at the cost of learning novel solutions.

We propose a new method for creating computationally efficient and compact convolutional neural networks (CNNs) using a novel sparse connection structure that resembles a tree root. This allows a significant reduction in computational cost and number of parameters compared to state-of-the-art deep CNNs, without compromising accuracy, by exploiting the sparsity of inter-layer filter dependencies. We validate our approach by using it to train more efficient variants of state-of-the-art CNN architectures, evaluated on the CIFAR10 and ILSVRC datasets. Our results show similar or higher accuracy than the baseline architectures with much less computation, as measured by CPU and GPU timings. For example, for ResNet 50, our model has 40% fewer parameters, 45% fewer floating point operations, and is 31% (12%) faster on a CPU (GPU). For the deeper ResNet 200 our model has 48% fewer parameters and 27% fewer floating point operations, while maintaining state-of-the-art accuracy. For GoogLeNet, our model has 7% fewer parameters and is 21% (16%) faster on a CPU (GPU).

Deep Convolutional Neural Networks (CNNs) have recently evinced immense success for various image recognition tasks [11, 27]. However, a question of paramount importance is somewhat unanswered in deep learning research - is the selected CNN optimal for the dataset in terms of accuracy and model size? In this paper, we intend to answer this question and introduce a novel strategy that alters the architecture of a given CNN for a specified dataset, to potentially enhance the original accuracy while possibly reducing the model size. We use two operations for architecture refinement, viz. stretching and symmetrical splitting. Stretching increases the number of hidden units (nodes) in a given CNN layer, while a symmetrical split of say K between two layers separates the input and output channels into K equal groups, and connects only the corresponding input-output channel groups. Our procedure starts with a pre-trained CNN for a given dataset, and optimally decides the stretch and split factors across the network to refine the architecture. We empirically demonstrate the necessity of the two operations. We evaluate our approach on two natural scenes attributes datasets, SUN Attributes [16] and CAMIT-NSAD [20], with architectures of GoogleNet and VGG-11, that are quite contrasting in their construction. We justify our choice of datasets, and show that they are interestingly distinct from each other, and together pose a challenge to our architectural refinement algorithm. Our results substantiate the usefulness of the proposed method.

Despite having high accuracy, neural nets have been shown to be susceptible to adversarial examples, where a small perturbation to an input can cause it to become mislabeled. We propose metrics for measuring the robustness of a neural net and devise a novel algorithm for approximating these metrics based on an encoding of robustness as a linear program. We show how our metrics can be used to evaluate the robustness of deep neural nets with experiments on the MNIST and CIFAR-10 datasets. Our algorithm generates more informative estimates of robustness metrics compared to estimates based on existing algorithms. Furthermore, we show how existing approaches to improving robustness “overfit” to adversarial examples generated using a specific algorithm. Finally, we show that our techniques can be used to additionally improve neural net robustness both according to the metrics that we propose, but also according to previously proposed metrics.