HOT CHIPS: Beyond Compute – Enabling AI Through System Integration

The keynote for the second day of HOT CHIPS this year was a presentation by Ganesh Venkataramanan of Tesla Motors. He titled it Beyond Compute – Enabling AI Through System Integration. Tesla did not present at HOT CHIPS last year (this year they had two presentations and a keynote). They saved up all their announcements for the company's AI day. Ganesh presented there and I covered the announcement of DOJO, especially its incredible packaging, in my post NOT CHIPS: Tesla's Project Dojo. By the way, Tesla's second AI day is on 30 September. I assume it will be live-streamed like last year, but I can't find details on their website.

Given that two of his colleagues were already presenting details of the current state of Tesla's technology, Ganesh focused on history. Computing has changed over the ages, from human brains, to abaci, to the PC, Cray supercomputers, to today's AI training servers, and even AI training data centers.

He had the usual graphs of how data is growing exponentially, and that the data is increasingly unstructured, 80% was his number; That means that 20% of the data can be processed with traditional programming whereas 80% requires AI techniques. He looked at weather forecasting and how their hardware and software approaches have changed over time, as you can see in the above graph. Until recently it used algorithmic approaches, going from scalar through vector a,d even more scalable. But now it is ML/AI.

He took a moment to define the terms AI, ML, and DL.

But also the differences between traditional computers, that you have to program before they can do anything useful, and learning computers that you have to train before they can do anything useful.

One of the big challenges in AI training is labeling the data, which often has to be done manually. In fact, the reason AlexNet drove so many breakthroughs in vision recognition is because it crowd-sourced the labeling so that suddenly researchers had orders of magnitude more labeled data than before. Tesla is obviously concerned with driving scenes, but the 2D approach is insufficient to automatically label objects such as the above traffic cone.

Even with 4D space and time labeling with multiple cameras, Tesla still required a 1,000-person in-house data labeling team working with fully custom-built data labeling and analytics infrastructure. But it did give a 100X increase in labeling throughout.

But to really get the throughput up requires getting the humans out of the loop. Ganesh had a nice historical quote from Turing from the Second World War.

Using AI techniques to drive the labeling with multiple inputs from multiple cameras means the labeling can be automated.

But what to label? This is where I think Tesla has a huge advantage over all the other competitors in autonomous driving. They have literally millions of cars on the roads. They can ask the fleet to supply interesting video clips. They collect 1,000 interesting clips and automatically label them in a week.

Ganesh then talked about AI research and did a deep dive into DOJO. He had 80 slides so I can't cover everything. Most of the DOJO stuff was in last year's HOT CHIPS presentation that I wrote about. Here's the link again: my post NOT CHIPS: Tesla's Project Dojo.

All of this is part of the march up the staircase from big data to AI, and presumably, one day general AI.

So AI/ML is the next phase in computing, following the path blazed by the abacus, Cray, and laptops. Ganesh ended with another quote from Turing (The Times, 1949):

This is only a foretaste of what is to come, and only the shadow of what it is going to be. We have to have some experience before we really know its capabilities...I do not see why it should not enter any one of the fields normally covered by the human intellect and eventually compete on equal terms.

 

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.

.