Guwahati: What if artificial intelligence could work more like the human brain—processing only what matters instead of constantly analysing every piece of information?
Researchers at the Indian Institute of Technology, Guwahati (IIT-Guwahati) are working on a new AI model based on this idea. The model is designed to process long sequences of data while using significantly less energy than many conventional AI systems.
The research was presented at the International Conference on Machine Learning (ICML) 2026 in Seoul, one of the leading international conferences on artificial intelligence.
Developed by researchers from IIT Guwahati’s SustainAI Lab at the Mehta Family School of Data Science and Artificial Intelligence, the model is called Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model (SH²RFSSM).
At its core, the model takes inspiration from how neurons in the brain communicate. Unlike conventional AI systems that continuously process information, spiking neural networks activate when there is something important to respond to. This can reduce unnecessary computation and, in turn, energy use.
The IIT Guwahati team combined this approach with state-space modelling, allowing the AI to identify patterns across long stretches of data without the heavy computational requirements of some traditional sequence models.
The researchers say this could be particularly useful for devices that need to analyse data continuously but have limited battery power or computing capacity.
Potential applications include smartwatches and other wearable health devices, IoT sensors, environmental monitoring, smart manufacturing, autonomous systems and long-term weather or traffic forecasting.
The model also introduces neuronal heterogeneity, meaning its artificial neurons do not all behave in exactly the same way. According to the researchers, this helps the system capture the complex patterns that occur in real-world data over time.
The team tested SH²RFSSM on 17 benchmark datasets, covering tasks such as long-range sequence classification, regression, human activity recognition and long-term forecasting.
The model achieved performance comparable to leading sequence-processing AI models while showing substantially lower estimated energy consumption, pointing to its potential for so-called edge AI—where AI processing happens directly on devices rather than relying heavily on cloud servers.
Ayon Borthakur, Assistant Professor at the Mehta Family School of Data Science and AI, said the growing use of AI for analysing long streams of health, environmental, industrial and forecasting data is creating a need for models that are more computationally efficient.
“Modern AI systems increasingly rely on analysing long streams of sequential data,” Borthakur said, adding that the computational cost can make such systems difficult to use on battery-powered and resource-constrained devices.
Kartikay Agrawal, PhD Research Scholar and co-author of the study, said the combination of spiking neural networks and state-space modelling allows the system to learn long-range patterns while reducing unnecessary computation.
The research was co-authored by Kartikay Agrawal, Vaishnavi Nagabhushana, Abhijeet Vikram, Vedant Sharma and Ayon Borthakur.
Agrawal, Nagabhushana and Borthakur presented the work at the ICML 2026 poster session in Seoul on July 7.
The researchers now plan to further test and improve the model for real-world applications, with a focus on making AI more efficient and adaptable for resource-constrained devices.