Accelerating the Low-Rank Decomposed Models

TL;DR

The paper tackles the bottlenecks of large AI models by reframing Low Rank Decomposition (LRD) as an accelerator rather than solely a compressor. It introduces four acceleration-oriented strategies—rank-aware selection, layer freezing, layer merging, and branched Tucker—to balance memory savings with actual speedups. Experimental results on ResNet-50/101/152 demonstrate substantial parameter and FLOP reductions with meaningful throughput gains and minimal accuracy loss, validating the practical viability of LRD-based acceleration. The work further suggests that progressive LRD, combined with selective knowledge distillation at high compression, can extend these benefits to transformer-scale models.

Abstract

Tensor decomposition is a mathematically supported technique for data compression. It consists of applying some kind of a Low Rank Decomposition technique on the tensors or matrices in order to reduce the redundancy of the data. However, it is not a popular technique for compressing the AI models duo to the high number of new layers added to the architecture after decomposition. Although the number of parameters could shrink significantly, it could result in the model be more than twice deeper which could add some latency to the training or inference. In this paper, we present a comprehensive study about how to modify low rank decomposition technique in AI models so that we could benefit from both high accuracy and low memory consumption as well as speeding up the training and inference

Figures (7)

• Figure 1: SVD
• Figure 2: Tucker2
• Figure 4:
• Figure 5:
• Figure 7: Mixing consecutive 1x1 Conv layers in ResNet modules after applying Tucker decomposition to the middle 3x3 Conv layer.