ThirdAI Corp. founder
says he is frustrated by the inherent inefficiency of artificial intelligence.
Instead of trying to make it more efficient by throwing huge amounts of computing power at it, Dr. Shrivastava wants to rethink the fundamentals of AI itself in an effort to make it cheaper and more widely used.
“I think you can change the whole stack, the whole economics of how AI is being used,” said Dr. Shrivastava, a tenured associate professor of computer science at Rice University in Houston. ThirdAI, the company he founded to bring his ideas to market, last year raised $6 million in a seed round led by Neotribe Ventures, Cervin Ventures and Firebolt Ventures.
Dr. Shrivastava, 37 years old, offers a simple scenario to illustrate the inefficiency, in which billions of elements or parameters in an AI model are routinely updated every time there is an error, even though just a few hundred may actually need the update.
A researcher might show an image of a cat to a neural network that has been trained for a month to recognize one billion characteristics of animals. If the neural network, a kind of AI algorithm, makes a mistake processing the image of the cat, it will go back and update all of the parameters in the model, even though only a fraction of them actually need correcting.
Dr. Shrivastava said that this inefficiency typically is addressed by doubling down on AI with more hardware to power through those extra computations. “We are doing an inefficient process and we are doing it with even more energy. That is something that bothers me,” Dr. Shrivastava said during an interview at the Collision tech conference in Toronto last month.
To make AI more efficient, Dr. Shrivastava has borrowed techniques used in the field of search engines. “If you type something into Google, it doesn’t go and scan all of the web. What is relevant automatically pops up,” he said.
The idea of ThirdAI is to process and train AI models by limiting updates to relevant parameters, which pop up automatically, according to Dr. Shrivastava.
It does that by making use of a technique called hashing, in which data is tagged and stored in memory close to similar kinds of data, according to
professor in the computer science and artificial intelligence laboratory at the Massachusetts Institute of Technology.
The use of hashing to speed up deep learning is a relatively new area of research and it isn’t easy, according to Dr. Indyk. “There is a fair amount of engineering that has to be done well. You have to tune everything very carefully,” he said.
ThirdAI’s technology would allow AI to run on cheap chips known as central processing units, instead of more expensive graphics processing units. AI is often run on GPUs, which can handle many computations at once. CPUs handle fewer computations simultaneously, but have more memory. By reducing the number of computations required to process and train AI models, AI can run on the cheaper CPUs, according to Dr. Shrivastava.
Lowering the cost and energy consumption of AI means it can be applied more widely, he said.
“There are a lot of problems we cannot solve with AI because we are bottlenecked by resources—energy, weather forecasting, processing genomic data sets,” all of which require expensive super computers that draw lots of energy, Dr. Shrivastava said.
the founder and managing director of Neotribe, said he believes that ThirdAI has huge potential, although it faces challenges as an early-stage company that is just beginning to commercialize.
“The beauty of ThirdAI is that it is going to massively increase the footprint of AI. The challenge is that it needs to find the right kind of use cases,” he said.
Dr. Shrivastava said the company is engaging with several companies, and the technology could be applied to a broad range of business applications including search, product discovery, natural language processing, and forecasting.
Write to Steven Rosenbush at [email protected]
Copyright ©2022 Dow Jones & Company, Inc. All Rights Reserved. 87990cbe856818d5eddac44c7b1cdeb8