Sparse-IFT: Sparse Iso-FLOP Transformations for Maximizing Training Efficiency

TL;DR

Sparse-IFT introduces iso-FLOP sparse transformations that replace dense layers with sparse equivalents, increasing layer representational capacity without raising training FLOPs. A single hyperparameter, the sparsity level $s$, governs four transform families (Sparse Wide, Sparse Parallel, Sparse Factorized, Sparse Doped), expanding the search space for DST to uncover higher-performing subnetworks. Across vision and NLP, Sparse-IFT yields consistent accuracy gains (e.g., +3.5% on ResNet-18/ImageNet at 90% sparsity) and favorable transfer outcomes, supported by spectral analyses that relate Ramanujan-graph-inspired connectivity to performance. Real-world hardware studies reveal substantial wall-clock acceleration potential when unstructured sparsity is supported by specialized kernels, underscoring the need for co-design between sparse ML techniques and sparse-optimized hardware.

Abstract

Recent research has focused on weight sparsity in deep neural network training to reduce FLOPs, aiming for improved efficiency (test accuracy w.r.t training FLOPs). However, sparse weight training often compromises accuracy, requiring extended training schedules to attain the accuracy of dense models. In contrast, our approach, Sparse Iso-FLOP Transformations (Sparse-IFT), uses sparsity to improve accuracy while maintaining dense model FLOPs. Using a single hyperparameter (i.e., the sparsity level), Sparse-IFTs efficiently replace dense layers, expanding the search space for optimal sparse masks. In addition, dynamic sparse training (DST) with Sparse-IFT models effectively navigate this larger sparse mask-weight space, which is evidenced by a spectral analysis using Ramanujan graph properties. Our study reveals a robust correlation among mask topology, weights, and final performance. Notably, without adjusting any training hyperparameters, replacing dense layers with Sparse-IFT yields significant improvements, such as a +3.5% boost for ResNet-18 on ImageNet and +0.9% for GPT-3 Small on the Open LLM leaderboard. To the best of our knowledge, this is the first work to demonstrate the use of sparsity for improving the accuracy of dense models through a set of simple-to-use sparse transformations. Code is available at: https://github.com/CerebrasResearch/Sparse-IFT.

Figures (7)

• Figure 1: Top-1 Accuracy vs. Training FLOPs for variants of ResNet on ImageNet. Sparse-IFT provides significant accuracy gains across different models and sparsity levels, $s \in \{\text{50\%, 75\%, 90\%}\}$, while using the same training FLOPs as its dense counterpart.
• Figure 2: Different members of the Sparse-IFT family, each parameterized by a single hyperparameter (i.e., sparsity level, $s$). Black and white squares denote non-active and active weights, respectively. Green block indicates a non-linear activation function (e.g., ReLU). Derived with sparsity set at $50\%$ as an example, all transformations are Iso-FLOP to the dense feedforward function $f_{\theta_l}$, making them suitable drop-in replacements for $f_{\theta_l}$. Details about each member are in Section \ref{['subsec:members_of_sift']}.
• Figure 3: Ablation studies with Sparse-IFT on the ResNet-18 model for CIFAR-100 across sparsity $\in \{50\%, 75\%, 90\%\}$. (left) Sparse Wide IFT trained with dynamic unstructured and structured sparsity. (middle) Sparse-IFT family members trained with RigL, where Sparse Wide performs the best. (right) Sparse Wide IFT trained in a sparse and dense manner.
• Figure 4: The relationship between the structure and weights of Sparse-IFT ResNet-18 networks are analyzed through a graph perspective in terms of performance. Top row: we assess the relationship between $\Delta r_{imdb}$ and $\lambda_{imsg}$. Bottom row: investigates the correlation between $\Delta r$ and $\lambda$. The Pareto curvature heatmap visually represents the classification performance, with varying color gradients symbolizing the spectrum from low to high test accuracy on CIFAR-100.
• Figure 5: Benchmarking unstructured sparsity during (left) inference on Neural Magic's DeepSparse runtime and (right) training acceleration on the Cerebras CS-2. In both setups, we measure the relative increase in latency or training speed for Sparse-IFT variants against the dense model.