Using the phenomenon of quantum tunnelling, IIT Bombay researchers have demonstrated, for the first time, a spiking neuron network that is highly compact and shows potential for brain-scale implementation. The research, published in the journal IEEE Transactions on Circuits and Systems, demonstrates the use of a 36-member network of spiking neurons in a speech recognition module.
Energy efficient mode
The inter-disciplinary work, led by Maryam Shojaei Baghini, who brought in expertise in circuit design, and Professor Udayan Ganguly who provided the technology and algorithm knowhow, from the Electrical Engineering department of Indian Institute of Technology (IIT), Bombay, showed that their concept of a neuron could be realised and that it works in an energy efficient mode with low power requirement, suitable for emulating the brain.
“This specific work started in 2020, though the relevant knowledge has built over many years and by many people in the team… Our goal was to show a neuron in a network doing something useful. So, we took a prototypical spoken word recognition problem, that is, mimicking the auditory cortex,” says Mr. Ganguly.
In second generation, artificial neural networks, neurons represent their state in eight-bit precision. This does not mimic biological systems.
Instead in a spiking neural network, the next step, the neuron’s output state is “spike” (equivalent to a “one”) or “no spike” (equivalent to a “zero”). This has a binary representation and is closer to the natural workings of neurons in the brain. If such a neural network should also occupy less space and consume little power, it could meet the decadal challenge of brain scale computing.
“We have conceptualised and experimentally demonstrated a very compact and highly energy efficient neuron — a key building block of a neural network,” says Mr. Ganguly.
Making hardware circuit
First, the algorithm of (audio) processing had to be developed. While mathematically the algorithm may be framed in an ideal situation, in actuality, the developer faces non-linearities and other variables. This part of the development was handled by Vivek Saraswat, a graduate student in the department.
In the next step which involved developing the hardware circuit, researchers used the standard technology with a different applied voltage pattern to operate the technology in the quantum-tunnelling-dominant regime.
“This regime is typically avoided in conventional applications and models may not be very accurate. Hence, we partly developed the circuit in the blind,” says Ajay Singh, a graduate student at the department.
While the group was hoping at least one neuron would work well, the whole network sprang to life. “It was a wonderful surprise. Especially as all this was done during the early COVID-19 days remotely, a very ‘unfamiliar’ mode of working, which has since then become the norm,” says Mr. Ganguly.
Comparing this spiking neural network with existing state-of-the-art technology is encouraging. In a publication that came out in May this year, in IEEE Transaction of Circuits & Systems, the researchers show that they have achieved 5,000 times lower energy-per-spike at a similar area, 50 times less area at a similar energy-per-spike, and ten times lower standby power at a similar area and energy-per-spike compared to the state-of-the-art benchmarks.
“Such overall performance improvement makes our neuron a promising candidate to enable brain-scale computing,” adds Mr. Ganguly.
