A Digital Liquid State Machine With Biologically Inspired Learning and Its Application to Speech Recognition

Area of Science:

• Neuroscience
• Computer Science
• Electrical Engineering

Background:

• Traditional machine learning methods often require significant computational resources and complex hardware.
• Existing liquid-state machines (LSMs) have limitations in terms of learning efficiency and hardware implementation.
• Bioinspired computing offers a promising avenue for developing more efficient and low-power intelligent systems.

Purpose of the Study:

• To present a novel bioinspired digital liquid-state machine (LSM) for low-power very-large-scale-integration (VLSI) applications.
• To introduce the first bioinspired spike-based online learning algorithm for LSMs, enabling on-the-fly information extraction.
• To optimize the LSM for speech recognition tasks, focusing on hardware implementation efficiency and performance.

Main Methods:

• Development of a bioinspired spike-based online learning algorithm for digital LSMs.
• Implementation of local synaptic weight updates based on neuronal firing activities, avoiding global communication.
• Benchmarking the digital LSM using subsets of the TI46 speech corpus for isolated word recognition.
• Investigation of synaptic model impacts on reservoir fading memory and network performance.
• Analysis of tradeoffs between synaptic weight resolution, reservoir size, and recognition accuracy.

Main Results:

• The proposed online learning rule enables efficient information extraction without intermediate data storage.
• The local learning rule facilitates parallel VLSI implementation, outperforming backpropagation-based learning in efficiency.
• The digital LSM demonstrates competitive performance in isolated word recognition, rivaling Sphinx-4.
• The study identifies optimal synaptic models and hardware implementation techniques for reduced complexity and maintained performance.
• The bioinspired LSM outperforms other reported LSM and neural network-based recognizers on the TI46 speech corpus.

Conclusions:

• The bioinspired digital LSM with spike-based online learning offers a low-power, efficient solution for machine learning applications.
• The proposed learning algorithm and hardware optimization techniques pave the way for practical VLSI implementations.
• This work sets a new benchmark for speech recognition using LSMs, demonstrating superior performance and efficiency.