Indian scientists invent artificial neuron that will help create “human-like” machines, study says

0


[ad_1]

India

Get short url

https://cdn1.img.sputniknews.com/img/07e4/0a/15/1080837243_0:200:1920:1280_1200x675_80_0_0_80e96c8deccdd6656ffc1b354600244b.jpg

Sputnik International

https://cdn2.img.sputniknews.com/i/logo.png

https://sputniknews.com/india/202107171083398623-indian-scientists-invent-artificial-neuron-that-will-help-create-human-like-machines-study-says/

According to scientists, Neuromorphic Spiking Neural Networks (SNNs) are promising computational paradigms that depend on the mechanisms of the brain, firing neurons, and synaptic plasticity in mammals. The study found that in the DEXAT neuron model, when neurons fire, the threshold increases and returns exponentially to a baseline.

A team of scientists at the country’s leading institute, the Indian Institute of Technology, Delhi (IIT-Delhi), has invented a new model of neurons that will help build accurate, fast, and energy-efficient neuromorphic artificial intelligence (AI) systems in applications such as speech recognition . The team published its results as a study published in Nature Communication last week.

According to the research team, the neuron model they developed will improve the performance of neuromorphic Recurrent Spiking Neural Networks (RSNNs) by providing faster convergence, higher accuracy, and flexible long-term short-term memory.

In the human brain, neurons and synapses are the most important building blocks of intelligence. The model known as the Double Exponential Adaptive Threshold (DEXAT) will work in a similar way and make the AI ​​system more efficient.

“In recent years we have successfully demonstrated the use of storage technology beyond simple storage. We have used semiconductor memory efficiently for applications such as in-memory computing, neuromorphic computing, edge AI, sensing, and hardware security. This work specifically uses analog properties of nanoscale oxide-based memory modules to build adaptive spiking neurons, ”said Professor Manan Suri, lead author of the study.

According to the team, this invention will help achieve high performance with fewer neurons, and its benefits have been shown in multiple data sets. The researchers also showed that a classification accuracy of 91 percent of the Google Spoken Commands (GSC) record was achieved. In addition, it was found that the neuromorphic network of nanodevices achieves an accuracy of 94 percent even with very high device variability.

[ad_2]

Share.

Leave A Reply