In conversation with Emre Ozer from Arm
With the recent publication in Nature Electronics about Arm’s hardwired machine learning processor fabricated on Pragmatic’s unique flexible electronics technology platform, we caught up with Dr Emre Ozer, Principal Research Engineer at Arm, to find out a bit more about the project.
Background
Three years ago Arm and Pragmatic began collaborating on a project whose aim was to develop a system that could recognise the intensity level of odour detected by an array of flexible e-nose sensors. The high-level block diagram of the system was an array of sensors, interfaced to a machine learning (ML) processor. The idea behind the project was quite wide ranging: was it possible to have a low-cost solution that could be attached to everyday objects (e.g. food, personal care items, wound dressings, air quality monitors) that would give a simple good/bad output signal to alert the user/consumer to do something – eat or throw out the food, reapply cream, change a dressing, evacuate the premises.
Challenging a traditional view
The initial work was designing the most compact ML engine possible. The system contains multiple sensors that are sampled at regular intervals. The traditional view is that the data from one of those samples, for one of those sensors, is a single data point. But that isn’t how a machine learning algorithm works. It is perhaps simpler to describe it in visual terms – to train a visual ML engine to recognise an object, you can’t give it a series of pixels and tell it this one is blue and this one is green. You have to give it a whole picture: this is a car, this is a dog, etc. Once it has finished its training, then you give it an unknown photo and ask it to compare it to what it has seen before and respond with what it thinks the object is. It is the same with smell: it is more complex than an individual chemical, so a datapoint in ML terms for odour is a series of sensor values. The team at Arm then took these datapoints and trained different conventional ML algorithms with these datapoints. The performance of these conventional ML algorithms was sufficient but their implementations in metal-oxide TFTs as a processing engine were not efficient. So, we had to come up with a resource-efficient ML algorithm that has a similar performance to conventional ML algorithms but its implementation in hardware is more compact.
A whole new definition to low cost
This brings us to the other reason for the project – the team wanted to design a system that could be embedded into items we use every day. This brings a whole new definition to low cost – if we are talking about a 50 cent milk bottle, then the cost of the electronics to tell you if it is starting to sour has to be in order of cents. This ruled out the use of conventional, silicon-based electronics, where the high costs of fabrication make it impossible to manufacture them at the cost point needed. Arm and Pragmatic have a relationship that goes back to 2014, so we knew that Pragmatic’s unique technology platform was perfectly suited for the project with its fast turnaround times and low cost to manufacture. The article in Nature Electronics gives more in-depth information about the ML processing engine design that we implemented.
Next steps
As we move forward, we will be continuing to explore other ML algorithms to find the best solution in terms of efficiency and speed of resolution. Plus, we are looking into what we need to adapt to be able to extend the use to many more situations.