Authors: Aashman Mehta, AbdelQader AlKilany, Dan F. M. Goodman, Giacomo Indiveri, Melika Payvand
Organizations: Institute of Neuroinformatics, University of Zurich and ETH Zurich, Zurich, Switzerland · University of California, Los Angeles, CA, USA · Imperial College London, London, UK
Gating mechanisms allow a fixed neural network to adapt its computation to the current input by dynamically modulating its effective parameters. In the brain, neuromodulation uses this principle to move neural circuits between dynamical regimes, and deep learning exploits input-dependent gating, e.g., in gated recurrent networks and self-attention. Input-dependent modulation has also recently been shown to improve spiking neural networks, with particularly strong benefits in small, parameter-limited ones. This makes neuromodulation especially promising for neuromorphic processors at the edge, where area, memory and energy budgets constrain network size, while circuit physics limits parameter precision. Here we demonstrate input-dependent, within-trial neuromodulation of a spiking recurrent neural network on a mixed-signal neuromorphic processor. An off-chip modulator reads the on-chip network's input and activity and adjusts its population-level gains while the chip's analog circuits compute with the physics of the silicon. Neuromodulation raises binary spoken-digit classification accuracy by +17.3±10.4 percentage points, tested across five hardware network instantiations.
Figures & tables
Fig. 3 : Chip-in-the-loop modulation on DYNAP-SE (5 RNNs; excitatory and inhibitory gains of the E neurons on chip and somatic gain for off-chip readout neurons rewritten every 40 ms; 213 test trials); ( ) What is fixed, trained and modulated; ( ) Test accuracy of each RNN without and with modulation, the modulator trained with the activity-gated estimated gradient; ( ) Change in test accuracy over the unmodulated baseline for the modulated network and Controls A–C, mean ± SD over RNNs.
Modification
Tests whether improvement requires
A
Retrain readout only (fresh optimizer, same epochs)
Modulation beyond extra readout training
B
Gains held at the mean of the modulated validation run
Dynamic gains rather than a better operating point
C
Another trial’s gain trajectory played through the reservoir
Dependence of the gains on the current input
TABLE I : Controls used in the hardware experiments.
China Nanhu Academy of Electronics and Information Technology, Jiaxing, China · University of Electronic Science and Technology of China, Chengdu, China
Institute of Neuroinformatics, University of Zürich and ETH Zürich, Zürich, Switzerland · AI Center, ETH, Zürich, Switzerland · Digital Society Initiative, University of Zurich, Switzerland