Reservoir Network with Structural Plasticity for Human Activity Recognition

TL;DR

This work addresses edge-time-series processing under tight resource constraints by implementing a custom ESN-based neuromorphic chip with on-chip learning and structural plasticity. It introduces neurogenesis-enabled reservoir growth, LFSR-based random weights, and SGD-driven online training of the readout to maintain performance while adapting to non-stationary data. Key results show HAR accuracy of $95.95\pm0.78\%$ and prosthetic finger control accuracy of $85.24\pm2.31\%$, at throughput of $6\times 10^4$ samples/s and about $47.7$ mW power on IBM 65 nm, with robustness to noise and efficient data movement. The design demonstrates practical, low-power edge neuromorphic computing suitable for real-time HAR and prosthetic control applications.

Abstract

The unprecedented dissemination of edge devices is accompanied by a growing demand for neuromorphic chips that can process time-series data natively without cloud support. Echo state network (ESN) is a class of recurrent neural networks that can be used to identify unique patterns in time-series data and predict future events. It is known for minimal computing resource requirements and fast training, owing to the use of linear optimization solely at the readout stage. In this work, a custom-design neuromorphic chip based on ESN targeting edge devices is proposed. The proposed system supports various learning mechanisms, including structural plasticity and synaptic plasticity, locally on-chip. This provides the network with an additional degree of freedom to continuously learn, adapt, and alter its structure and sparsity level, ensuring high performance and continuous stability. We demonstrate the performance of the proposed system as well as its robustness to noise against real-world time-series datasets while considering various topologies of data movement. An average accuracy of 95.95% and 85.24% are achieved on human activity recognition and prosthetic finger control, respectively. We also illustrate that the proposed system offers a throughput of 6x10^4 samples/sec with a power consumption of 47.7mW on a 65nm IBM process.

Figures (9)

• Figure 1: High-level diagram of the ESN, which mainly comprises of input, reservoir, and readout layer. The input layer buffers the input, whereas the reservoir layer and readout layer stochastically extract input features and classify it, respectively.
• Figure 2: The system high-level architecture of the proposed ESN is mainly composed of FIR filters (optional), central control unit (CCU), shifter, and a group of neurons clustered into two layers: reservoir and readout. The CCU controls the data flow and core unit synchronization, and the shifter manages the sequential data transfer from the reservoir (blue) to output (red) neurons. The reservoir neurons provide non-linear expansion ensuring linear separability of input features by the output neurons.
• Figure 3: The RTL schematic of the: (Left) Output neuron, in which the weight-sum of the reservoir output is computed and the weight adjustment is taking place; (Right) Leaky-integrated discrete-time continuous reservoir neuron used in the proposed ESN system.
• Figure 4: RTL design of hyperbolic tangent ($tanh$) activation function modeled using piece-wise functions.
• Figure 5: (a) Data movement from the input layer to the readout layer via an H-Tree network, (b) SH-Tree topology which uses a single H-Tree network shared among all the PEs in the network, (c) local rings topology, in which data within the layers are circulated through separated local rings, (d) MH-Tree topology, where multiple H-Tree networks are used to move data within the reservoir and dedicated unidirectional links to transfer reservoir activation directly to the readout layer.