IIT Guwahati Researchers Present Brain-Inspired AI Model at ICML 2026

Students from the Mehta Family School of Data Science & Artificial Intelligence (MFSDS&AI) IIT Guwahati, presented a poster on an efficient spiking neural network for long-range time series at the prestigious International Conference on Machine Learning (ICML) 2026 in Seoul, South Korea — a top tier, CORE A*-ranked international conference in Artificial Intelligence.

Students from the SustainAI Lab at the Mehta Family School of Data Science & Artificial Intelligence (MFSDS&AI), IIT Guwahati, presented their research titled "A Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model for Time Series," at the International Conference on Machine Learning (ICML) 2026, held at the COEX Convention & Exhibition Center, Seoul, South Korea.

The work introduced a spiking state space model designed to efficiently learn long-range dependencies in sequential data. By combining second-order state space modelling with Harmonic Resonate-and-Fire neurons and neuronal heterogeneity, the proposed approach brings together the long-range modelling capabilities of state space models with the sparse, event-driven computation of spiking neural networks.

Modern AI systems often struggle to efficiently process very long streams of data, particularly on resource-constrained devices such as wearables and sensors with limited battery power. Although models such as transformers are widely used for sequence modelling, their quadratic computational complexity with respect to sequence length makes them increasingly expensive as sequences grow, making them inefficient for long-range modelling.

In this work, the authors introduced the Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model (SH²RFSSM), a second-order spiking state space model designed to efficiently capture long-range dependencies in sequential data. The model combines the sparse, event-driven computation of spiking neural networks with the long-sequence modelling capabilities of state space models. The authors also investigated neuronal heterogeneity, allowing selected neuronal parameters to vary across neurons rather than using identical settings throughout the network.

The proposed approach was evaluated across 17 datasets covering long-range sequence classification, regression, human activity recognition, and long-term forecasting. The model demonstrated competitive performance with state-of-the-art state space models while offering substantially lower estimated energy consumption. These results demonstrate the potential of spiking state space models to combine long-range sequence modelling with sparse, energy-efficient computation, particularly for resource-constrained applications.

The work, co-authored by Kartikay Agrawal, Vaishnavi Nagabhushana, Abhijeet Vikram, Vedant Sharma, and Ayon Borthakur, was presented at ICML 2026 by Kartikay Agrawal, Vaishnavi Nagabhushana, and Ayon Borthakur during the conference poster session (Hall A, Poster 608) on 7 July 2026.

Further reading: https://icml.cc/virtual/2026/poster/62795

Visualization: https://kartikay24.github.io/HRF-SSM/index.html