Apple has acquired Xnor.ai, a Seattle-based startup that builds AI models that run on edge devices, for approximately $200 million.
GeekWire first reported the story, based on information from "sources with knowledge of the deal." Because XNor.ai's technology focuses on running AI models on low-resource edge devices without sending data to the cloud, many observers speculate that the acquisition ties into Apple's strategy for data privacy. However, GeekWire and others report that Apple has given its standard response to inquiries:
Apple buys smaller technology companies from time to time and we generally do not discuss our purpose or plans
Xnor.ai was spun out of the Allen Institute for AI's (AI2) incubator program in 2017, with funding from Seattle VC Madrona Venture Group. Company co-founder Ali Farhadi was featured in a 2016 New York Times article discussing his team's research on reducing compute requirements for neural networks. Xnor.ai's latest products include a solar-powered FPGA chip that runs a computer-vision neural network as well as a person-detection algorithm embedded in Wyze security cameras---a partnership that was recently and unexpectedly terminated.
The key idea behind Xnor.ai's technology is quantization of neural network weights and operations---reducing the number of bits needed to represent the numbers in the model. Quantization is a feature of most deep-learning frameworks, including TensorFlow and PyTorch; however, Xnor.ai's team took the extreme approach of using just a single bit to represent each weight and activation value in the convolutional layers, the main component of state-of-the-art computer vision models. Their result was a neural network that achieves an accuracy only 2.9% less than a full-precision model. In addition to a 32x memory savings, representing the data as single bits also allows the computations to be performed with a simple logic operation, XNOR. Besides giving the company a name, this also means that the vision models can run 58x faster on a regular CPU, obviating the need for a GPU.
The memory and compute savings also mean that the models can be used in devices with resource and power constraints, such as security cameras or mobile phones. This on-edge processing is attractive to the privacy-minded, as it eliminates the need for data to be sent to the cloud for AI processing, which is a major concern for consumers and regulators. Apple CEO Tim Cook's 2019 commencement address at Stanford University stressed the importance of privacy, prompting observers to speculate that this is a motivation for the acquisition. On the other hand, Cook mentioned in a 2019 CNBC interview that Apple acquires a company "every two to three weeks on average," having acquired nearly two dozen in the first half of that year. Apple recently overtook Google as the leader in the number of AI acquisitions among the "tech giants," and has made several high-profile hires of ex-Google AI researchers, including Ian Goodfellow, the inventor of generative adversarial networks (GANs).
Xnor.ai is the second startup from AI2 Incubator to be acquired by a large tech company. In 2017, Baidu acquired Kitt.ai, an AI2 startup that developed a chatbot framework. The Incubator website lists three other funded AI startups, along with alumni Kitt.ai and Xnor.ai.