Key points are not available for this paper at this time.
Abstract We present the adaptive locally competitive algorithm (ALCA) as an efficient, biologically inspired, sparse front-end for neuromorphic speech processing. Neuromorphic and sparse coding paradigms are increasingly explored to bridge the efficiency gap between the human brain and conventional hardware, particularly in audio processing. Although the original LCA is a promising sparse coding model for many neuromorphic applications, its potential in speech classification has not been fully investigated. The proposed adaptive version extends the original algorithm by dynamically adapting the sensitivity of its dictionary atoms, modeled as a filter bank, through learnable modulation parameters, thereby enhancing lateral inhibition and improving reconstruction quality, sparsity, and convergence time. Our experiments show that both algorithms outperform the LAUSCHER cochlea model in speech classification accuracy. Although LCA improves accuracy, it incurs higher power consumption, whereas ALCA achieves better accuracy while reducing dynamic power usage to between 4 mW and 13 mW on neuromorphic hardware, three orders of magnitude lower than setups using Graphics Processing Units. Moreover, despite being sparse, ALCA matches the classification accuracy of dense Mel spectrogram representation while substantially reducing computational costs. These findings position ALCA as a compelling solution for efficient speech processing systems, promising substantial advancements in balancing accuracy and power efficiency while remaining data-driven and not dependent on the intended task.
Bahadi et al. (Tue,) studied this question.