Learning at the Speed of Light

The past decade has been a transformative time in the world of machine learning. A field that was once heavier on hype than on practical applications grew up and began to deliver major breakthroughs that revolutionized industrial processes and consumer products. But for the field to continue to deliver major gains in these areas and beyond, further advances in the field of tinyML will be needed. Traditional methods of deploying machine learning algorithms – small computing devices that rely on powerful computing resources in the cloud to perform inferences – are limited in their application due to issues with privacy, latency and cost. TinyML offers the promise of eliminating these problems and opening up new classes of problems to be solved by artificially intelligent algorithms.

Also Read :  Internet Stocks In China Struggle Amid Challenges

Of course, a state-of-the-art machine learning model, with billions of parameters, is not exactly easy when memory is measured in kilobytes. But with some creative thinking and a hybrid approach that harnesses the power of the cloud and mixes it with the benefits of tinyML, it might just be possible. A team of researchers at MIT has shown how this is possible with them method called netcast that relies on resource-heavy cloud computers to quickly retrieve model weights from memory, then transmit them almost instantly to the TinyML hardware over a fiber optic network. Once these weights are transmitted, an optical device called a broadband “Mach-Zehnder” modulator combines them with sensor data to perform local, lightning-fast calculations.

The team’s solution uses a cloud computer with a large amount of memory to hold the weights of a full neural network in RAM. These weights are streamed to the connected device as needed through an optical pipe with enough bandwidth to transfer an entire movie in a millisecond. This is one of the biggest limiting factors that prevents tinyML devices from running large models, but it is not the only factor. Processing power is also at a premium on these devices, so the researchers also proposed a solution to this problem in the form of a shoebox-sized receiver that performs super-fast analog calculations by encoding the input data to the transmitted weights.

Also Read :  Bad news: Headlines are indeed getting more negative and angrier | Media

This scheme makes it possible to perform trillions of multiplications per second on a device with the resources of a desktop computer from the early 1990s. In the process, on-device machine learning that guarantees privacy, minimizes latency, and is made highly energy-efficient is possible. Netcast has been tested on image classification and digit recognition tasks with over 50 miles separating the tinyML device and cloud resources. After only a small amount of calibration work, average accuracy rates above 98% were observed. The results of this quality are good enough to be used in commercial products.

Also Read :  Apple's iPhone Will Be Irrelevant One Day. Here's How It Will Happen.

Before that happens, the team works to further improve their methods to achieve even better performance. They also want to shrink the shoebox-sized receiver down to the size of a single chip so it can be incorporated into other devices such as smartphones. With further refinement of Netcast, big things may be on the horizon for tinyML.

Source

Leave a Reply

Your email address will not be published.

Related Articles

Back to top button