Blab AI
IIT Guwahati Unveils Energy-Efficient AI Model for Long-Sequence Data
At the International Conference on Machine Learning (ICML 2026) in Seoul, the Indian Institute of Technology Guwahati (IIT Guwahati) introduced a new artificial‑intelligence model called the Spiking Heterogeneous Harmonic Resonate‑and‑Fire State Space Model (SH²RFSSM). The design promises to handle long streams of data while using far less energy than many conventional neural‑network approaches.
SH²RFSSM blends two well‑established techniques: spiking neural networks (SNNs) and state‑space modelling. SNNs mimic biological neurons that fire only when a membrane potential reaches a threshold, generating spikes only in response to meaningful events. This event‑driven operation cuts down on unnecessary computation. State‑space models, meanwhile, track system behaviour over time through a set of state variables, enabling efficient representation of long‑range dependencies.
The research team, hailing from IIT Guwahati’s SustainAI Lab, consists of Kartikay Agrawal, Vaishnavi Nagabhushana, Abhijeet Vikram, Vedant Sharma, and Ayon Borthakur. According to the team, SH²RFSSM is a second‑order spiking state‑space model that incorporates neuronal heterogeneity—allowing individual neurons to have different characteristics. This diversity helps the network capture complex temporal patterns that arise in real‑world data.
In a series of experiments, the model was evaluated on 17 benchmark datasets covering classification, regression, human‑activity recognition, and long‑term forecasting. Results show that SH²RFSSM achieves performance comparable to state‑of‑the‑art sequence models while using substantially lower estimated energy. The authors emphasize that the energy advantage is not limited to training; it also applies to inference, which is critical for devices that operate on battery power.
"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. "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."
The event‑driven nature of SH²RFSSM means that spikes are generated only when input signals cross a threshold, reducing the number of operations required during inference. This sparsity, combined with the efficient representation of long‑range dependencies, makes the model attractive for edge AI applications where local processing is preferred over constant cloud communication.
Potential use cases identified by the researchers include wearable health‑monitoring devices, Internet‑of‑Things (IoT) sensors, smart‑factory control systems, environmental monitoring stations, autonomous vehicles, and long‑term weather or traffic forecasting. In each scenario, lower computational requirements could extend battery life and reduce the need for high‑bandwidth data transmission.
The team plans to move from benchmark evaluation to real‑world deployment. Vaishnavi Nagabhushana said the next focus would be on improving the model’s efficiency and adaptability so that it can be deployed more effectively on resource‑constrained devices. The research therefore addresses a growing challenge in AI: making sophisticated models capable of continuous, on‑device operation without excessive energy consumption.
The work was presented at ICML 2026, one of the leading international conferences in machine learning. It adds to a growing body of research that seeks to reconcile the performance demands of modern AI with the environmental and practical constraints of edge computing.
In summary, IIT Guwahati’s SH²RFSSM demonstrates that a brain‑inspired, event‑driven architecture can match the accuracy of current sequence models while consuming less energy. The next steps involve testing the model in practical, continuous‑data scenarios and exploring its integration into battery‑powered edge devices.
SH²RFSSM blends two well‑established techniques: spiking neural networks (SNNs) and state‑space modelling. SNNs mimic biological neurons that fire only when a membrane potential reaches a threshold, generating spikes only in response to meaningful events. This event‑driven operation cuts down on unnecessary computation. State‑space models, meanwhile, track system behaviour over time through a set of state variables, enabling efficient representation of long‑range dependencies.
The research team, hailing from IIT Guwahati’s SustainAI Lab, consists of Kartikay Agrawal, Vaishnavi Nagabhushana, Abhijeet Vikram, Vedant Sharma, and Ayon Borthakur. According to the team, SH²RFSSM is a second‑order spiking state‑space model that incorporates neuronal heterogeneity—allowing individual neurons to have different characteristics. This diversity helps the network capture complex temporal patterns that arise in real‑world data.
In a series of experiments, the model was evaluated on 17 benchmark datasets covering classification, regression, human‑activity recognition, and long‑term forecasting. Results show that SH²RFSSM achieves performance comparable to state‑of‑the‑art sequence models while using substantially lower estimated energy. The authors emphasize that the energy advantage is not limited to training; it also applies to inference, which is critical for devices that operate on battery power.
"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. "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."
The event‑driven nature of SH²RFSSM means that spikes are generated only when input signals cross a threshold, reducing the number of operations required during inference. This sparsity, combined with the efficient representation of long‑range dependencies, makes the model attractive for edge AI applications where local processing is preferred over constant cloud communication.
Potential use cases identified by the researchers include wearable health‑monitoring devices, Internet‑of‑Things (IoT) sensors, smart‑factory control systems, environmental monitoring stations, autonomous vehicles, and long‑term weather or traffic forecasting. In each scenario, lower computational requirements could extend battery life and reduce the need for high‑bandwidth data transmission.
The team plans to move from benchmark evaluation to real‑world deployment. Vaishnavi Nagabhushana said the next focus would be on improving the model’s efficiency and adaptability so that it can be deployed more effectively on resource‑constrained devices. The research therefore addresses a growing challenge in AI: making sophisticated models capable of continuous, on‑device operation without excessive energy consumption.
The work was presented at ICML 2026, one of the leading international conferences in machine learning. It adds to a growing body of research that seeks to reconcile the performance demands of modern AI with the environmental and practical constraints of edge computing.
In summary, IIT Guwahati’s SH²RFSSM demonstrates that a brain‑inspired, event‑driven architecture can match the accuracy of current sequence models while consuming less energy. The next steps involve testing the model in practical, continuous‑data scenarios and exploring its integration into battery‑powered edge devices.
