IIT Guwahati researchers develop energy-efficient AI model for long data sequences

Researchers at the Indian Institute of Technology Guwahati have developed a brain-inspired Artificial Intelligence model that can process long sequences of data with lower energy requirements than many conventional AI systems.

The research was carried out by the SustainAI Lab at IIT Guwahati’s Mehta Family School of Data Science and Artificial Intelligence and was presented at the International Conference on Machine Learning (ICML) 2026 in Seoul, South Korea.

The model, named Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model (SH²RFSSM), combines spiking neural networks with state space modelling to handle long-range patterns while reducing computational demands.

Unlike conventional neural networks, which continuously process information, spiking neural networks activate when significant changes or events occur. This allows them to perform computations more selectively and potentially consume less energy.

The researchers have also incorporated neuronal heterogeneity, allowing artificial neurons within the model to have different characteristics rather than functioning identically. According to the team, this helps the system capture complex patterns in sequential data more effectively.

The model was tested on 17 benchmark datasets covering long-range sequence classification, regression, human activity recognition and long-term forecasting. The researchers said SH²RFSSM achieved performance comparable to leading sequence models while recording substantially lower estimated energy consumption.

The findings could have applications in areas where devices need to analyse continuous streams of information without relying heavily on cloud computing. Potential uses include wearable health-monitoring devices, Internet of Things sensors, smart manufacturing systems, environmental monitoring, autonomous systems and long-term forecasting.

Dr Ayon Borthakur, assistant professor at the Mehta Family School of Data Science and AI, said existing AI architectures can become increasingly computationally expensive as the length of sequential data grows, creating challenges for battery-powered and resource-constrained devices.

Kartikay Agrawal, a PhD research scholar and co-author of the study, said combining spiking neural networks with state space modelling allowed the team to address long-range sequence processing while reducing the computational burden associated with conventional approaches.

The research team plans to further test the model in real-world applications involving continuous data processing and work towards improving its efficiency and adaptability for deployment on resource-constrained devices.

The study was co-authored by Kartikay Agrawal, Vaishnavi Nagabhushana, Abhijeet Vikram, Vedant Sharma and Ayon Borthakur. Agrawal, Nagabhushana and Borthakur presented the work during the ICML 2026 poster session at Seoul’s COEX Convention and Exhibition Centre on July 7.