Researchers at MIT have paved the way to low-power neural networks that can run on devices such as smartphones and household appliances. Andrew Hobbs explains why this could be so important for connected applications and businesses.
Many scientific breakthroughs are built on concepts found in nature – so-called bio-inspiration – such as the use of synthetic muscle in soft-robotics.
Neural networks are one example of this. They depart from standard approaches to computing by mimicking the human brain. Usually, a large network of neurons is developed, without task-specific programming. This can learn from labelled training data, and apply those lessons to future data sets, gradually improving in performance.
For example, a neural network may be fed a set of images labelled ‘cats’ and from that be able to identify cats in other images, without being told what the defining traits of a cat might be.
But there’s a problem. The neurons are linked to one another, much like synapses in our own brains. These nodes and connections typically have a weight associated with them that adjusts as the network learns, affecting the strength of the signal output and, by extension, the final sum.
As a result, constantly transmitting a signal and passing data across this huge network of nodes requires large amounts of energy, making neural nets unsuited to battery-powered devices, such as smartphones.
As a result, neural network applications such as speech- and face-recognition programs have long relied on external servers to process the data that has been relayed to them, which is itself an energy-intensive process. Even in humanoid robotics, the only route to satisfactory natural language processing has been via services such as IBM’s Watson in the cloud.
A new neural network
All that is set to change, however. Researchers at Massachusetts Institute of Technology (MITT) have developed a chip that increases the speed of neural network computations by three to seven times, while cutting power consumption by up to 95 percent.
This opens up the potential for smart home and mobile devices to host neural networks natively.
“The general processor model is that there is a memory in some part of the chip, and there is a processor in another part of the chip, and you move the data back and forth between them when you do these computations,” MIT News reports, in an interview with Avishek Biswas, MIT graduate student in electrical engineering and computer science, who led the chip’s development.
Traditionally, neural networks consist of layers of nodes that pass data upwards, one to the next. Each node will multiply the data it receives by the weight of the relevant connection. The outcome of this process is known as a dot product.
“Since these machine-learning algorithms need so many computations, this transferring back and forth of data is the dominant portion of the energy consumption,” said MIT Biswas.
“But the computation these algorithms do can be simplified to one specific operation, the dot product. Our approach was, can we implement this dot-product functionality inside the memory, so that you don’t need to transfer this data back and forth?”
A mind for maths
This process will sometimes occur across millions of nodes. Given that each node weight is stored in memory, this amounts to enormous quantities of data to transfer.
In a human brain, synapses connect whole bundles of neurons, rather than individual nodes. The electrochemical signals that pass across these synapses are modulated to alter the information transmitted.
The MIT chip mimics this process more closely by calculating dot products for 16 nodes at a time. These combined voltages are then converted to a digital signal and stored for further processing, drastically reducing the number of data calls on the memory.
While many networks have numerous possible weights, this new system operates with just two: 1 and -1. This binary system act as a switch within the memory itself, simply closing or opening a circuit. While this seemingly reduces the accuracy of the network, the reality is just a two to three percent loss – perfectly acceptable for many workloads.
Internet of Business says
At a time when edge computing is gaining traction, the ability to bring neural network computation out of the cloud and into everyday devices is an exciting prospect.
We’re still uncovering the vast potential of neural networks, but they’re undoubtedly relevant to mobile devices. We’ve recently seen their ability to predict health risks in fitness trackers, such as Fitbit and Apple Watch.
By allowing this kind of work to take place on mobile devices and wearables – as well as other tasks, such as image classification and language processing – there is huge scope to reduce energy usage.
MIT’s findings also open the door to more complex networks in the future, without having to worry so much about spiralling computational and energy costs.
However, the far-reaching power of abstraction inherent in neural networks comes at the cost of transparency. Their methods may be opaque – so called black box solutions – and we expose ourselves to both the prejudices and the restrictions that may come with limited machine learning models. Not to mention any training data that replicates human bias.
Of course, the same problems, lack of transparency, and bias be found in people too, and we audit companies without having to understand how any individual’s synapses are firing.
But the lesson here is that, when the outcome has significant implications, neural networks should be used alongside more transparent models, where methods can be held to account. Just as critical human decision-making processes must adhere to rules and regulations.