Teaching a machine how to drive
By: Stephanie Wong
If you saw an image of a cat, how did you conclude that it was a cat and not a dog? Can you describe the difference between a cat and a dog? Because you can easily conclude whether a given image is one of a cat or dog but to describe the decision making process is very hard. The size of the thing, the number of legs, the tail and number of ears is the same. The difference is not obvious. That is what engineers at Waymo, Google’s ‘child’ company working on self-driving cars, essentially have to encode in a machine — to decide if an object is a pedestrian or not. But teaching a computer and training it to pick out a pedestrian out of a bunch of data is easier than trying to encode the difference. What the Waymo cars actually “sees” is pedestrians encoded as yellow rectangles, other vehicles as purple boxes and so on. There are still many problems with the software, for example, stickers can be put on stop signs that confuse the system into thinking it’s a speed limit sign.
Waymo uses deep learning to predict and respond to data gathered from its millions of miles driven in both public roads and simulation. Deep learning is a type of machine learning that uses a neural network, containing lots of layers at different levels of abstraction, to analyze data. Engineers build models that sift through huge amounts of data to look for patterns, creating neural networks. They then use Google’s data centers to train its neural nets. It utilizes high-powered cloud computing hardware to go through millions of bytes of data. After being trained by both human labellers and automated processes, this huge dataset then needs to be pruned to be able to deployed in Waymo’s vehicles in the real world.
Programmers at Waymo created an algorithm that, by analyzing street photos, could teach a computer to learn and identify visual patterns that characterize a pedestrian. However, they discovered that most of the errors were made by human labellers. This begs the question of whether or not we can eventually filter out all the original human errors and perhaps then the machine will be better than humans. At the heart of it all though, machine learning is still dependent on a vast source of data. Krizhevesky, one of the original people who made huge discoveries in the field of machine learning, and former employee at Waymo, thinks Tesla has a unique advantage of being able to collect data from a variety of environments because there are Tesla drivers all over the world. He even thinks that from the data side, Tesla is ahead of the game. But even so, driving more miles may not give Tesla a bigger advantage because, up to a certain point, returns start to diminish. What matters is the uniqueness of the data that you collect, it’s the edge cases that matter. Jaywalkers and parallel-parking cars present themselves as rare examples. As such, these scenarios are put through Waymo’s simulator and then transformed into many other interactions that further train the models. And although Waymo may be ahead of the game now, machine learning as a field is growing rapidly and it’s not like Waymo doesn’t control the field so competitors will catch up.
