A Small Team of Students Have Beaten Google’s Machine Learning

Fast.ai is a small company dedicated to bringing artificial intelligence (AI) education to anyone who’s interested in learning more about AI, machine learning, and deep learning. It doesn’t matter what operating system you use, the education you’ve received, or however old or new your preferred programming language may be.

Making AI more accessible to those who already show enthusiasm about the field is the company’s mission. Those noble company values are paying off; a small group of students from Fast.ai recently programmed an artificial intelligence (AI) algorithm that beats Google’s algorithm.

AI Is Still Anyone’s Game

Co-founders Jeremy Howard and Rachel Thomas both teach at the University of San Francisco while working on Fast.ai among many other endeavors in healthcare technology, deep learning, and software development.

This triumphant achievement is significant because Google’s budget far exceeds Fast.ai’s. It proves that AI is still anyone’s game, and those who build objectively better algorithms will see the rewards immediately.

As Howard proudly says, “State-of-the-art results are not the exclusive domain of big companies.”

Apples to Oranges?

Students at Fast.ai are part-timers who want to learn machine learning and data science. The company gives its students access to Amazon’s cloud, just like professional machine learning developers would use.

The benchmark that these students beat computes the speed of a common image classification task against the algorithm per dollar of computing power. The group of students blew Google’s custom-built chips made specifically for machine learning out of the water.

The team’s biggest update was giving the training algorithm images that were cropped correctly. Howard points out that “these are the obvious, dumb things that many researchers wouldn’t even think to do.”

As a result, the algorithm runs 40% cheaper than Google’s code. However, without using the same hardware as Google, it’s difficult to directly compare both algorithms. And Matei Zaharia, a Stanford University professor who helped create the benchmark, also expresses some doubt. He says that while Fast.ai’s results are impressive, the algorithm should also be built to handle large amounts of data.

Access for All

Jack Clark is the director of communications and policy at OpenAI, a non-profit dedicated to creating friendly AI that benefits humanity. And even though Fast.ai’s algorithm may not be directly comparable to Google’s, Clark says Fast.ai is an indispensable resource and tool for many other industries and technologies, like language analysis.

Clark adds, “Things like this benefit everyone because they increase the basic familiarity of people with AI technology.” And that’s never a bad thing.

Would you be interested in joining a cohort at the Fast.ai institute? Do you think Google’s algorithm is directly comparable to Fast.ai’s? Let us know in the comments below!