Answers are grounded in this paper and cite their sources.
Long Short-Term Memory (LSTM) networks are widely used in sequence-based tasks such as speech recognition and sentiment analysis. However, the presence of gating units significantly increases computational overhead and memory usage, making LSTMs less suitable for resource-constrained edge devices and latency-sensitive applications. To address this, various model compression techniques, such as pruning and structured weight, have been explored to reduce memory access and computational demands. However, deploying these compressed models on conventional LSTM accelerators often suffer from low compute efficiency, limiting their practical applicability. This paper proposes a novel column-wise pruning method that adaptively adjusts the sparsity of gating matrices based on their sensitivity to model accuracy, thereby maximizing overall sparsity. To further reduce computational complexity, the sparse vectors processed using the delta network (DN) efficiently skips zero elements during matrix computations. Moreover, we introduce a highly optimized column-wise hardware architecture tailored for efficient LSTM inference. Our accelerator is implemented on an AMD Zynq 7000 XC7Z100 FPGA and evaluated using the TIMIT dataset. Experimental results demonstrate an 80.5× model compression rate without accuracy loss, and 24.8 Top/s of effective throughput. Compared to state-of-the-art LSTM accelerators, our design achieves 2.6× ∼ 27× speedup and 2.3× ∼ 25× improvement in energy efficiency.
No takes yet. Share an insight, caveat, or question.
Shen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: