The other day at the DeepMind blog, someone came up with an idea for improving Waymo’s self-driving cars: Evolution
Waymo’s self-driving vehicles employ neural networks to perform many driving tasks, from detecting objects and predicting how others will behave, to planning a car’s next moves. Training an individual neural net has traditionally required weeks of fine-tuning and experimentation, as well as enormous amounts of computational power. Now, Waymo, in a research collaboration with DeepMind, has taken inspiration from Darwin’s insights into evolution to make this training more effective and efficient.Yu-hsin Chen, “How evolutionary selection can train more capable self-driving cars” at DeepMind blog
In the research collaboration between Yu-hsin Chen and Matthieu Devin of Waymo, and Ali Razavi, Ang Li, Sibon Li, Ola Spyra, Pramod Gupta and Oriol Vinyals of DeepMind, a 24% improvement was said to result from population-based training PBT:
Networks are evaluated periodically and compete with each other for “survival” in an evolutionary fashion. If a member of the population is underperforming, it’s replaced with the “progeny” of a better performing member. The progeny is a copy of the better performing member, with slightly mutated hyperparameters. PBT doesn’t require us to restart training from scratch, because each progeny inherits the full state of its parent network, and hyperparameters are updated actively throughout training, not at the end of training. Compared to random search, PBT spends more of its resources training with good hyperparameter values.Yu-hsin Chen, “How evolutionary selection can train more capable self-driving cars” at DeepMind blog
Industry pros will be able to tell over time if the new method is widely applicable. Thinking of it as “evolution” at work, however, risks misconceptions.
Life forms change over the generations and we call that “evolution.” But selection of the best performing members (Darwinian survival of the fittest) is only a part of the story.
Here are some other things life forms do to change over time:
They use horizontal gene transfer; that is, genes are simply transferred between one type of life form and another, often by bacteria or parasites. For example, a tick might spread a gene from cows to reptiles.
Sometimes events that a life form has survived change the genes and they are passed on in a changed state to offspring. At one time, scientists thought that this wasn’t possible but there is enough evidence today that we call them epigenetic changes. What past generations, whether people or lab rats, did affects the genes their offspring inherit.
Because a life form needs only to live and produce offspring, not to, say, make decisions for a self-driving car navigating heavy traffic, many life forms simply ditch complex machinery in favor of parasitism. For example, mushrooms that live on trees depend on the trees to do their digesting for them.
A variety of life forms use genome doubling, jumping genes, hybridization, and symbiosis to stay alive and they pass on life in an altered form as a result. As one researcher put it, sometimes it’s “almost hilariously complicated.”
… natural selection is daily and hourly scrutinizing, throughout the world, every variation, even the slightest; rejecting that which is bad, preserving and adding up all that is good; silently and insensibly working, wherever and whenever opportunity offers, at the improvement of each organic being in relation to its organic and inorganic conditions of life.
But there they sit, the horseshoe crabs, comb jellies, acorn worms and liverworts of the world, surviving and unimproved.
In short, in evolution, “performance” just means the continued survival of a lineage. Thus it includes hybrids between what you might want for your purposes and what you don’t want. It includes parasitism, no change at all for hundreds of millions of years, and lots of destruction of useful mechanisms.
So if the researchers have specific goals in mind, they are acting as goal-directed intelligent designers. They are using a mechanism (which they call Darwinian natural selection) to produce a specific outcome. It may work; we shall see. But it is not natural evolution in general.
One reason for confusion is that, quite often, school systems have tended to teach only a dumbed-down version of evolution: the natural selection to which the researchers refer.
The much more complex and multi-faceted way that evolution really works is ignored. The bottom line is that “evolution” is not necessarily useful for any purpose other than just staying alive. But then the vast majority of life forms have never asked for more than that anyway. AI really is a different game altogether.
More news from DeepMind:
Confirmed: DeepMind’s deepest mind is on leave. The chess champ computer system just never made money. Co-founder Mustafa Suleyman is a philosopher and social justice activist who hoped to use the technology for fundamental transformations. But his AI ethics board lasted about seven days at Google.
Why is DeepMind in deep water financially? Market analysts are wondering if the money is as smart as the machine. In an all-out botwar with the other tech Bigs, DeepMind could simply be paying top minds not to work for the competition while readying AI tools that pay better than winning at board games. Maybe.
News re Waymo: Will industry pressure loosen self-driving car tests?
Does a Western bias affect self-driving cars? How a driver is expected to act varies by culture. Self-driving cars (autonomous vehicles) will need to adapt to different rules and we will, very likely, need to change those rules to make the vehicles work.