[Paper Review] Increasing Liquid State Machine Performance with Edge-of-Chaos Dynamics Organized by Astrocyte-modulated Plasticity

Vladimir A. Ivanov, Konstantinos P. Michmizos|arXiv (Cornell University)|Oct 26, 2021

Advanced Memory and Neural Computing9 references24 citations

TL;DR

NALSM uses an astrocyte-inspired modulation of STDP to steer liquid state machines toward near-critical dynamics, achieving state-of-the-art LSM accuracy on MNIST and N-MNIST without dataset-specific tuning, and scaling to larger liquids with competitive performance.

ABSTRACT

The liquid state machine (LSM) combines low training complexity and biological plausibility, which has made it an attractive machine learning framework for edge and neuromorphic computing paradigms. Originally proposed as a model of brain computation, the LSM tunes its internal weights without backpropagation of gradients, which results in lower performance compared to multi-layer neural networks. Recent findings in neuroscience suggest that astrocytes, a long-neglected non-neuronal brain cell, modulate synaptic plasticity and brain dynamics, tuning brain networks to the vicinity of the computationally optimal critical phase transition between order and chaos. Inspired by this disruptive understanding of how brain networks self-tune, we propose the neuron-astrocyte liquid state machine (NALSM) that addresses under-performance through self-organized near-critical dynamics. Similar to its biological counterpart, the astrocyte model integrates neuronal activity and provides global feedback to spike-timing-dependent plasticity (STDP), which self-organizes NALSM dynamics around a critical branching factor that is associated with the edge-of-chaos. We demonstrate that NALSM achieves state-of-the-art accuracy versus comparable LSM methods, without the need for data-specific hand-tuning. With a top accuracy of 97.61% on MNIST, 97.51% on N-MNIST, and 85.84% on Fashion-MNIST, NALSM achieved comparable performance to current fully-connected multi-layer spiking neural networks trained via backpropagation. Our findings suggest that the further development of brain-inspired machine learning methods has the potential to reach the performance of deep learning, with the added benefits of supporting robust and energy-efficient neuromorphic computing on the edge.

Motivation & Objective

• Motivate improving LSM accuracy without backpropagation or dataset-specific tuning.
• Leverage astrocyte biology to regulate synaptic plasticity and network dynamics.
• Place LSM dynamics near the edge-of-chaos to maximize computational capacity.
• Demonstrate scalability to larger liquid sizes and robustness to sparse astrocyte connections.

Proposed method

• Extend a baseline LSM with STDP by embedding a LIM astrocyte model.
• Use an astrocyte to modulate the STDP depression rate based on liquid/input activity to near-critical dynamics.
• Define a proxy branching factor BF_proxy(t) from spike counts to approximate the critical branching factor.
• Train the output layer with gradient descent on spike-count features from the liquid.
• Evaluate accuracy on MNIST and N-MNIST and compare to LSM variants and backpropagation-based networks.

Experimental results

Research questions

• RQ1 Can astrocyte-modulated STDP steer LSM dynamics toward near-critical edge-of-chaos behavior without dataset-specific tuning?
• RQ2 Does NALSM achieve higher accuracy than baseline LSM and STDP-enhanced LSM variants on MNIST and N-MNIST?
• RQ3 What is the impact of liquid size and neuron-astrocyte connectivity on performance?
• RQ4 How close to the critical branching factor does NALSM operate, and how does this relate to kernel quality and generalization?

Key findings

• NALSM achieves top accuracies of 97.61% on MNIST and 97.51% on N-MNIST with 8,000 liquids, surpassing baseline LSM and LSM+AP-STDP under same conditions.
• With 1,000 neurons, NALSM reaches 96.15% (MNIST) and 96.13% (N-MNIST), outperforming LSM, LSM+STDP, and LSM+AP-STDP.
• NALSM maintains accuracy advantages with sparse neuron-astrocyte connectivity down to 10% density.
• Accuracy improves with larger liquids, saturating around 8,000 neurons, and remains competitive on Fashion-MNIST (85.84%).
• Kernel quality (linear separation and generalization) correlates with higher accuracy, supporting near-critical dynamics as beneficial for LSMs.