Researchers at the University of Newcastle have implemented a non-neural-network hardware that can significantly cut the power consumption of artificial intelligence.

The team trained a neural network, and their technology – a ‘Tsetlin machine’ – to recognise hand written digits from the standard MNIST data set.

“Using an out-of-the-box neural network we could get less than 10 trillion operation per Joule”, Newcastle senior lecturer Rishad Shafik (right) told Electronics Weekly. “Using our first design of Tsetlin machine, we got about 65 trillions per Joule. This improvement fundamentally comes from simpler logic-based designs.”

When identifying key words spoken by random speakers, a different Tsetlin machine “could identify 15 words for the same energy budget needed by neural network to identify one word”, he added.

A Tsetlin machine (pronounced ‘setlin’) is a type of ‘learning automata’, a kind of machine learning algorithm invented by Russian scientist Mikhail Tsetlin in the 1960s, not long after the invention of neural networks.

The problem with learning automata, explained Newcastle’s Shafik, is that in their basic form “learning automata are almost impossible to implement in hardware because of the vast number of states that must be accommodated”.

Enter a Norweigen AI professor, Ole-Christoffer Granmo of the University of Agder, who in the last few years found a way to cut the complexity of learning automata by combining them with classical game theory and Boolean algebra, said Shafik.

He implemented the simplified learning automata in software, and gave them the name ‘Tsetlin machine’ after the founder of the subject.

Building on Granmo’s work, and working with him, the Newcastle team has found a way to efficiently map the data types and underlying algorithms of Tsetlin machines onto logic gates, and is implementing them on FPGAs and in a custom ASIC – in a form that can be used both for the training/learning AI phase, and in use after training – the latter frequently called ‘inferencing’ by AI folks.

Shafik puts power savings between Tsetlin machines and neural networks down to the way they map onto hardware: Neural networks, even binary neural networks, are arithmetic – they use many multiply-and-add operations, while in Tsetlin machine hardware does not use arithmetic. “It is purely logic and naturally parallel”, he said.

Newcastle’s connection to Tsetlin Machine has another thread: Inventor Mikhail Tsetlin taught learning automata theory to the supervisor of Alex Yakovlev, who heads the Newcastle Microsystems Research Group and leads its AI team with Shafik.

Yakovlev is a pioneer of clock-less digital designs, said Shafik, who has long held the view that AI applications will need simple automaton blocks that are independent and modular.”

According to Yakovlev: “Energy efficiency is the biggest enabler for AI. Also important, is to be able to explain the AI decisions. For both of these, we need to move away from arithmetic, and our Tsetlin machine hardware design provides just that.”

What can such a machine do?

“Any type of machine learning that requires training”, said Shafik – essentially, anything a neural network can do.

How Tsetlin machines work is explained in detail in the paper ‘Learning automata based energy-efficient AI hardware design for IoT applications’, a paper published in Royal Society Philosophical Transactions A.

Shafik put it in a nutshell for Electronics Weekly:

The hardware Tsetlin machine expresses an image pixel or voice data sample as a set of ‘0’ or ‘1’ values. Internally the machine has as many learning automata as there are these values. These automata are independent learning units that can ‘play games’ during training.

In each step, the ‘game’ involves tossing a coin [the machine includes a random element] to suggest if the values should be included or not. After a few steps, the automata learn to make better decisions using history. Their objective is to select the best combination of values that can define the machine learning problem at hand – for example, detecting objects in an image or recognising a syllable or keyword in a voice sample.

When a training routine is complete, the Tsetlin machine returns a logic expression that is significantly simpler than the long arithmetic sequences generated in neural networks. This logic expression can be used to infer decisions from data.

The Newcastle team used in-house expertise in low-power event-driven design to create the hardware. A game changer was to find parallel patterns with fewer logic components in each path. These paths are only enabled when there is a genuine need for cutting down the unnecessary power consumption and improving performance.

At Newcastle, Shafik and Yakovlev work with: Adrian Wheeldon (hardware), Jie Lei (machine learning applications) and Tousif Rahman (data encoding).

The team has a Tsetlin machine chip in a fab now. Called mignon – French for small and beautiful – it will be the first to implement this form of learning automata according to Shafik.

The work had been supported by platform and impact accelerator grants from the Engineering and Physical Sciences Research Council (EPSRC).