A small team of students AI coders from Fast.ai, a small unit that offers free online courses for machine learning, just beat Google researchers by creating a better performing AI algorithm. These coders are part-time students at Fast.ai who are simply enthusiastic about machine learning and seek a career in data science.
The code created by these students was measured using a benchmark from researchers at Stanford called “DAWNBench.” It uses a common image classification task to track the execution speed of a deep-learning algorithm against each dollar of computing power.
These rankings were previously dominated by Google’s researcher for training category on different machines. They used a customized collection of chips specially designed for machine learning, however, the Fast.ai team was successful in achieving faster results on a similar hardware.
The team was able to beat Google because they took care of simple things like ensuring that the images fed to its training algorithm were cropped correctly. Jeremy Howard, one of Fast.ai’s founders said, “These are the obvious, dumb things that many researchers wouldn’t even think to do.”
Fast.ai’s coders trained the algorithm on the “ImageNet database in 18 minutes using 16 Amazon Web Service instances, at a total compute cost of around $40.”
According to Howard, this result is nearly 40% more efficient than Google’s algorithm but he also admits that this comparison isn’t that simple because the hardware used by them is different.
Nevertheless, this achievement is quite significant because it can squash the common notion that advanced AI research is a task limited to only those with huge resources.