In the fast-paced world that we live in, it often seems as though the new gadget we just bought is obsolete before we have a chance to even take it out of the box. That can be frustrating, but at least you can upgrade it to get back on the cutting-edge of technology, right? Right??? Chances are you have just climbed back in your chair after a fit of rolling around on the floor laughing, because you know that upgradability largely went out the window long ago. If you want the latest technology these days, you have to toss the old device out and buy a new one.
This state of affairs is surely responsible for untold amounts of glee in the boardrooms of Big Tech companies, but that same excitement is not felt by users of these non-upgradable devices. A group of MIT engineers that knows this pain all too well has recently published the details of a technique to construct artificial intelligence chips that can be stacked and reconfigured in a LEGO brick-like way. The design is composed of layers of sensing and processing elements, with one key difference from previous approaches — light. Rather than using fragile conductive wires that restrict reconfigurability of layers, they communicate with one another optically.
The tiny chips can be built from any arbitrary combination of sensing and computing layers. The processing layers are based on arrays of memory resistors (known as artificial synapses) that the team had previously developed. These arrays form physical neural networks that can be trained to process and classify signals directly on the chip — no connection to the cloud needed. Communication between layers is handled by an LED and photodetector-based optical communications system.
To test the concept, the researchers built a four square millimeter stackable chip that was configured to perform basic image classifications. An image sensor embedded into a layer was used to capture a picture of a letter to be classified. This was followed by three artificial synapse array layers, with each trained to recognize a single letter — “M”, “I”, or “T." The synapses produce a current in response to any image shown to them. The higher the current level, the more certainty that the character it was trained to recognize is present in the image.
It was found that the classifier chip did a great job when the character images were clear, but when they were a bit blurry, it did not perform very well. As such, the project was deemed a total failure and never spoken of again. Wait a minute, this chip is reconfigurable, right? Of course it is, so actually the team built a new “denoising” layer and added it to the layer stack of the chip. With this change, the chip was working great and correctly classifying even blurry images.
The researchers are presently at work adding more sensing and processing capabilities to the chip. They see a future in which the chip will allow consumers to upgrade and endlessly customize their electronic devices. A few of the first application areas they are targeting are cell phone cameras, to help them recognize more complex images, and wearable electronic skins for monitoring health.