A team at the Indian Institute of Technology Bombay (IIT Bombay) is tackling the challenge of designing and implementing spiking neural networks to operate in small, low-power applications. Their use of quantum tunneling could spark a new generation of consumer electronics. This enables neural networks in edge devices. As the IIT Bombay team explains, humans and machines are not so different after all.
Ajay Singh, Senior Project Research Scientist, IIT Bombay:
“We develop hardware for neural networks, in a range of consumer applications. Imagine a small chip that works how your brain does – differentiating between, and reacting to, various stimuli. That could be speech commands, where you instruct a device to turn on a light, for example, or using gestures like waving a hand to trigger actions."
Singh says: "That is equivalent to existing virtual assistants, but those devices have to be plugged in and connected to the internet. We are working on ultra-low-power neural network design, which can be used in devices that need to run on batteries and last for days. It’s this focus on edge computing that makes us unique."
“Our goal is to make the chip as low-power as possible, so it can run on battery power for a very long time. These chips will be low cost too, which makes them perfect for use in edge applications.”
"For example, you could have a battery-powered watch that you talk to, to give commands. It does the work without needing to be plugged in all the time. Or it could be a camera you put somewhere remote to detect and record the movement of vehicles or animals.
Our goal is to make the chip as low-power as possible, so it can run on batteries for a very long time. These chips will be low-cost too, which makes them perfect for use in edge applications.”
Professor Udayan Ganguly, Associate Professor, Electrical Engineering, IIT Bombay:
“This has been a six-year journey for us, starting with getting the physics and the algorithm correct, and then building a circuit and system that can bridge the two. There are a couple of PhD theses’ worth of work here."
Professor Ganguly explains: "One significant challenge lay in understanding the algorithms that make neural networks function. What sort of structures do they have? How do they compute? But the main question was always how to implement these networks in silicon. You have to lay out hundreds of these interconnected neurons together, so they form a symphony on the chip, doing exactly what the algorithm is supposed to do."
“We asked what new physics we can use to achieve the necessary functionality at a neuron or synapse level. Quantum tunnelling enabled us to model the neuron behavior in hardware, for a very low-power alternative.”
"Typically, engineers are concerned with circuit design. They’re given the components, such as transistors, which act in a standard manner, and they’ll use a specific engineering trick to implement the algorithm.
We are adopting a different approach. We took those standard transistors, and asked what new physics we can use to achieve the necessary functionality at a neuron or synapse level.
We have been using a technique called quantum tunneling, which enabled us to model the neuron behavior in hardware, for a very low-power alternative.”
Ajay Singh:
“These techniques are very similar to what happens in the brain. A biological neuron gathers and integrates inputs, and when the signal level crosses a threshold, the neuron gives an event trigger and ‘fires’.
For example, when we hear a sound, the cochlea in the ear converts the audio signal into different streams of signal; specialized neurons then convert those signals into spiking signals that go to the brain, which figures out what has been said.
A similar process happens in circuits, when the current comes in and is stored on a capacitor. The main challenge is that, if you want to store a charge on a capacitor, you need a big capacitor, which increases your area and your power requirements.
Quantum tunnelling is a form of leakage current. If the energy barrier between two electrodes is narrow enough, electrons can ‘escape’ these barriers. We can harness this current. It’s very small, so you only need a very small capacitor to store the charge. In fact, we don't even need a separate capacitor at all. The MOSFET, a type of transistor we use, has a body which acts as a capacitor and stores the charge. So, the current and the capacitor can all be integrated via this single transistor."
"Our initial goal was to build a network of 125 neurons... Arm Academic Access helped us scale up our ambition. We got access to Arm IP about 18 months ago, and for the next tape-out, we could now create a design with 195 neurons.”
"We have also used quantum tunnelling to develop an audio signal processing filter. For speech classification, you need to convert the audio signal into a spiking pattern which can be understood by the neural network. We used quantum tunnelling to develop those filters, and built circuits around that. This reduced the area and power requirement on the filter side too. Now, the whole of that audio classification unit can be low-power and low area," says Singh.
"At the moment, we’re working in a completely analogue and asynchronous domain. In the near future, we’re going to try to develop a spiking neural network system using Arm Cortex processors. That’s something which we are very much looking forward to, as it will enable us to reach a wider consumer electronics market,” explains Abhishek Kadam.
“We had a hard two-month deadline for our first tape-out, and we couldn’t delay it. So, we had to work out a size of neural network that would work in the limited timeframe. In a neural network, you have multiple signals coming in and going out. In a brain, there are billions of neurons. Our initial goal was to build a network of 125 neurons, but we’d still need to connect to lots of peripherals, with a level of automation that can handle the inputs and outputs. We taped out a 36-neuron architecture for the first tape out.
Arm Academic Access helped us scale up our ambition. We got access to Arm IP about 18 months ago, and for the next tape-out, we could now create a design with 195 neurons.
Global Foundries directed us toward Arm, which managed the standard cell libraries of the technology we needed, and this kick-started our involvement with Academic Access. We’re creating things from scratch, so the automation Arm provides us is very helpful. If we have any doubts gaining access to a particular protocol, we can find it on the Product Download Hub (PDH) in our list of IP. And if we want documentation on a certain Arm IP, it’s available on the Arm Developer Hub.”
"We are striving to make a big impact on the world. For example, we are hoping to be able to harvest energy and use that to run AI systems, rather than relying on external power sources. Conventional AI systems are becoming ubiquitous in consumer electronics. We want to shift the world to our low-power solution. We want to save energy.”
Abhishek Kadam, Research & Teaching Assistant, IIT Bombay:
“Arm also accelerated our tape-out cycle, by helping us in the If we did the same task manually, it would take ages.
At the moment, we’re working in a completely analogue and asynchronous domain. In the near future, we are going to try to develop a spiking neural network system using Arm Cortex processors. That’s something which we are very much looking forward to, as it will enable us to reach a wider consumer electronics market. The standard cell and I/O cells libraries provided by Arm are crucial in achieving this feat.”
“We’re still at an early stage – the current projection says we’ll need two to three tape-outs, and roughly three years, before this can be plugged in and used.
But we are striving to make a big impact on the world. For example, we’re hoping to be able to harvest energy and use that to run AI systems, rather than relying on external power sources. Conventional AI systems are becoming ubiquitous in consumer electronics. We want to shift the world to our low-power solution. We want to save energy.”
Arm makes a wide range of IP and tools available at no charge for academic research use. Explore Research Enablement
Professor Udayan Ganguly
The IIT Bombay team interviewed here are:Professor Udayan Ganguly, Associate Professor, Electrical Engineering, IIT BombayAjay Singh, Senior Project Research Scientist, IIT BombayAbhishek Kadam, Research and Teaching Assistant, IIT Bombay