IIT Guwahati researchers have developed SH²RFSSM, a model inspired by how biological neurons communicate.

Model mimics how neurons communicate
Researchers at the Indian Institute of Technology Guwahati have developed a brain-inspired artificial intelligence model that can process long sequences of data while using substantially less estimated energy than many conventional AI approaches.
Called the Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model (SH²RFSSM), the model combines spiking neural networks with state space modelling to analyse patterns that unfold over long periods.
The research was presented at the International Conference on Machine Learning (2026 in Seoul, South Korea, one of the leading international conferences in machine learning.
The work was carried out by researchers from IIT Guwahati's SustainAI Lab at the Mehta Family School of Data Science and Artificial Intelligence.
The research team comprises Kartikay Agrawal, Vaishnavi Nagabhushana, Abhijeet Vikram, Vedant Sharma and Ayon Borthakur.
Key Points
- SH²RFSSM is designed to process long sequences of data with substantially lower estimated energy consumption than conventional approaches.
- Event-driven computing: Unlike conventional neural networks that continuously process information, spiking neural networks activate when meaningful events occur.
- Handles long data streams: The model is designed to identify long-range patterns without the heavy computational burden of some traditional sequence models.
- 17 benchmark datasets: Researchers tested it across classification, regression, human activity recognition and long-term forecasting tasks.
- Competitive performance: The model reportedly performed comparably with state-of-the-art sequence models while using less estimated energy.
- Different artificial neurons: Its neuronal heterogeneity allows individual neurons to have different characteristics, helping capture complex temporal patterns.
- Edge AI potential: The technology could enable AI to run directly on devices rather than depending heavily on cloud computing.
- Possible applications: Wearable health monitors, IoT sensors, smart factories, environmental monitoring, autonomous systems, weather and traffic forecasting.
- Battery advantage: Lower computational requirements could potentially extend battery life in portable and resource-constrained devices.
- International recognition: The research was presented at ICML 2026 in Seoul, one of the leading international machine-learning conferences.
- Bigger significance: The research addresses a growing AI challenge — how to make increasingly sophisticated AI less energy-hungry and practical for continuous, on-device use.
Why processing long data is a challenge
AI systems increasingly have to deal with continuous streams of information -- from health signals recorded by wearable devices to environmental sensors, industrial machines, traffic systems and weather forecasting.
The longer such data sequences become, the greater the computational burden can be for conventional AI architectures.
This creates a particular challenge for devices that operate on limited battery power or have restricted computing capacity.
"Modern AI systems increasingly rely on analysing long streams of sequential data such as health signals from wearable devices, environmental sensor readings, industrial monitoring data, and weather or traffic forecasts," said Dr Ayon Borthakur, assistant professor at IIT Guwahati's Mehta Family School of Data Science and AI.
"However, widely used AI architectures often become computationally expensive as the length of data increases, making them less suitable for battery-powered and resource-constrained devices," he said.

AI model inspired by the human brain
The IIT Guwahati model takes its cue from the way biological neurons communicate.
Unlike conventional neural networks, which generally process information continuously, spiking neural networks are designed to generate signals when particular events or changes occur. This event-driven approach can reduce unnecessary computation.
"Unlike conventional neural networks that continuously process information, spiking neural networks activate only when meaningful events occur, enabling sparse and energy-efficient computation," said Kartikay Agrawal, a PhD research scholar at IIT Guwahati.
The researchers combined this principle with state space modelling, which is designed to capture relationships and dependencies across long sequences of data without the computational burden associated with some traditional sequence-processing approaches.
Different neurons, different characteristics
One of the distinctive elements of SH²RFSSM is neuronal heterogeneity.
Instead of making all artificial neurons behave identically, the model allows individual neurons to have different characteristics.
According to the researchers, this diversity helps the system represent complex temporal patterns occurring in real-world data.
The architecture is a second-order spiking state space model that combines long-range modelling with the sparse, event-driven computation associated with spiking neural networks.
Tested on 17 datasets
The researchers evaluated SH²RFSSM on 17 benchmark datasets, covering tasks including long-range sequence classification, regression, human activity recognition and long-term forecasting.
According to the research team, the model achieved performance comparable to state-of-the-art sequence models while showing substantially lower estimated energy consumption.
That distinction is important because the research points not simply to faster AI, but to AI that could potentially operate more efficiently on devices with limited computing and battery resources.
From smartwatches to industrial sensors
The potential applications are wide-ranging.
The model could eventually be useful in wearable health-monitoring devices, Internet of Things sensors, smart manufacturing systems, environmental monitoring, autonomous systems and long-term weather or traffic forecasting.
For such applications, processing data locally could reduce the need to constantly send information to cloud servers.
Lower computational requirements could also help extend battery life in portable and edge devices.
This makes the work particularly relevant to the growing field of edge AI, in which AI computation is performed closer to where data is generated rather than relying entirely on remote data centres.
Independent reports have similarly highlighted the model's potential for resource-constrained and battery-powered devices.
Next step: Real-world deployment
IIT Guwahati researchers say their next focus is to explore the model's performance in practical applications involving continuous and long-range data processing.
Vaishnavi Nagabhushana, another researcher on the team, said the focus would be on improving the model's efficiency and adaptability so that such AI systems can be deployed more effectively on resource-constrained devices.
The research therefore addresses one of the emerging challenges of AI: As models become more capable and are deployed in more places, making them energy-efficient enough to run continuously on smaller devices is becoming increasingly important.
Feature Presentation: Ashish Narsale/Rediff




