Reproduce “Temporal and Spatial Reservoir Ensembling Techniques for Liquid State Machines” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Anmol Biswas, Sharvari Ashok Medhe, Raghav Singhal, Udayan Ganguly
Researchers from IIT Bombay introduced novel temporal and spatial reservoir ensembling techniques for Liquid State Machines (LSMs), alongside receptive field-based input weights for visual data. These methods achieved up to 98.1% accuracy on N-MNIST and 77.8% on SHD, enabling LSMs to compete with complex backpropagation-trained spiking neural networks while using significantly fewer trainable parameters.
Key insights
- Ensembling diverse reservoir dynamics, both temporally (TEPRE) and spatially (MuLRE), offers a more effective and efficient performance scaling strategy for LSMs than simply increasing reservoir size.
- The optimal temporal partitioning strategy for ensembles can align with the inherent temporal characteristics of the input data, as demonstrated by N-MNIST's three distinct motions.
- Receptive field-based input connections enhance LSMs' ability to process visual information by maintaining spatial context, leading to improved performance on vision benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.