Artificial intelligence will be essential for connected devices because humans alone have no way to handle the tsunami of information that devices and sensors generate. The question isn’t will there be AI, but where will it be?
That question matters because teaching computers how to recognize human speech or different objects currently relies on massive amounts of data and fancy learning models that are derived across huge compute clusters running specialty chips (GPUs or Google’s tensor processing units are popular).
This is why Seattle-based startup Xnor.ai excites me. The company, which is a spin out from the Allen Institute for AI, has built algorithms that use less memory and processing power to run AI models. This means a lower-performance chip could implement skills like natural language processing, translation or computer vision.
Ali Farhadi, CEO and cofounder of Xnor, says that the company is showing potential customers algorithms that run existing models as well as algorithms that can handle training a model on an embedded device. This last bit is a big deal.
AI can be divided into two categories. There’s training a model, which takes massive amounts of data and compute power. This is how companies teach a computer model to recognize a person as opposed to a dog. The other side of AI is inference. This is where the newly trained model goes out into the world, gets new data and then classifies the new data based on the model.
Inference is easier to do on general purpose computers. Companies like Movidius (which was purchased by Intel) and Qualcomm are making investments here. When it comes to training, most people are content to leave that to the giant GPUs in the cloud. But Farhadi, who is a former professor at Washington University, says Xnor will be able to do both inference and training on embedded devices.
What’s surprising to me is that he’s taking a software approach to the problem, instead of building a new chip. When a researcher is training a model in something like object recognition, the computer is handling the data in floating point numbers (they have a decimal). The results are highly accurate numbers suitable for big scientific computing.
Farhadi claims that Xnor takes floating point numbers and converts them to binary numerals that occupy less memory and require less processing power to compute. The expected downside of doing this would be a drop off in accuracy, but Farhadi says that AI models that run through the algorithm are just as accurate.
If this is true, the implications are huge. One of the most exciting pieces of technology I saw at CES was a tiny, sub-2 milliwatt chip from Qualcomm’s R&D group that performed basic object recognition. The low power consumption meant it could become part of a battery powered sensor or a component inside a smart phone.
With it, elements such as people tracking, bio-authentication, gesture-based user interfaces and more, become possible. Qualcomm’s tech requires a specific chip, but Xnor’s only requires software. That makes it far more flexible for developers.