IISc proposes ‘human brain’ inspired computing platform to boost speeds, energy efficiency of AI tools

This could revolutionize and shatter the glass ceiling of overall AI capabilities present today. The new platform, researchers said, offers flexibility and easy deployability of AI-based tasks onto one’s personal electronic gadgets like smart phones, laptops and desktops. This would also significantly find use in computing applications, in Machine Learning and scientific computing, the IISc team on their work which addressed some of the pressing gaps in the present-day neuromorphic ( following human brain-like computing techniques) operations.

Unlike the energy-heavy operability in the current digital platforms used for extensive computing purposes, the use of molecular films permitted researchers to track free ionic movements, which widened the memory pathways.

The latest paper published in Nature also described how kinetic controls were established over the molecular transition, enabling almost all quintessential neuromorphic traits in a single circuit element. The researchers applied timed voltage pulses and managed to trace large number of molecular movements and map each of them to a distinct electrical signal, forming an extensive ‘molecular diary’ of different states.

” Due to this free ionic movement, countless unique memory states and pathways were generated. Such intermediary states had remained inaccessible, so far, as most digital devices are only able to access two either high and low conductance states. This platform allowed us to control molecular kinetics very precisely inside an electronic circuit powered by nanosecond voltage pulses,” said Sreebrata Goswami from CeNSE.

Though tiny, the molecular change was what allowed the team to create a highly precise and efficient neuromorphic accelerator, the first-of-its-kind that can not only store but also process data within the same location, resembling the working of a human brain.

Such accelerators can be seamlessly integrated with silicon circuits to boost their performance and energy efficiency, the researchers said.

In the realm of AI, several milestones in neuromorphic computing, like edge training, deployment of advanced models such as generative adversarial networks, long or short-term memory or transformers could be unlocked. Several of these endeavours are currently being pursued at IISc.

” This is a completely indigenous effort. Now with the rare discovery, many of the challenges of the neuromorphic computing stand addressed,” said Sreebrata Goswami, visiting professor at the CeNSE.