— Gary Marcus (@GaryMarcus) November 7, 2019 If we want robots that can think like us, we’ve got to stop giving them all the answers. Curiosity and exploration are the two key components of the human intellect that deep learning simply doesn’t provide. In a recent article in Quanta Magazine, writer Matthew Hutson describes the work of computer scientist Kenneth Stanley, who is currently working at Uber’s AI lab. Stanley’s pioneering work in the field of “neuroevolution” has paved the way for a new artificial intelligence paradigm that eschews traditional objective-based training models in favor of AI models that have no purpose but to explore and be creative. Hutson writes: Standard deep learning models use a black box – a set of weights and parameters that, ultimately, become too complex for developers to describe individually – to ‘brew’ up machine learning algorithms and tweak them until they spit out the right data. This isn’t intelligence, it’s prestidigitation. Instead of hard-coding the rules of reasoning, or having computers learn to score highly on specific performance metrics, they argue, we must let a population of solutions blossom. Make them prioritize novelty or interestingness instead of the ability to walk or talk. They may discover an indirect path, a set of steppingstones, and wind up walking and talking better than if they’d sought those skills directly. If AI could evolve its own solutions and combine those parameters with deep learning, it’d be closer to imitating human-level problem solving. At least, that’s what Stanley argues. His research involves building evolutionary algorithms that can function in tandem with deep learning systems. In essence, rather than teaching an AI to solve a problem, he develops algorithms that sort of meander about seeing what they’re capable of. These systems aren’t rewarded for solving a problem like normal AI paradigms. They just go until something happens. What’s remarkable is that, without a problem to solve, they still manage to solve many kinds of problems far more efficiently than traditional deep learning models. More from Hutson’s article in Quanta: Deep learning AI doesn’t know what to do when it hits a wall. Once the machine gets stuck, it has to start over again – that’s why it takes millions of training cycles to “teach” an AI how to accomplish a task successfully. With Stanley’s evolutionary algorithm-based hybrid model, the AI isn’t trying to find the exit, it’s basically just doing stuff and then trying to find more stuff to do. The machine’s ‘curiosity’ forces it through the entire maze almost every time because it’s bent on exploring. Evolutionary algorithms aren’t new, but the vein of research surrounding them has been largely swept to the side in favor of more immediately-lucrative development opportunities in standard deep learning technology – the kind that fuels B2B and B2C sales. And they’re also under-explored because they’re expensive. It’s takes a lot less power to train a narrow-minded AI than it does to run evolutionary algorithms. But the payoff could be huge. The big idea here is to backdoor human-level intelligence on accident by letting AI evolve its own algorithms though unfettered exploration. Stanley and others believe it’s possible that AGI could manifest as a byproduct of machine curiosity, just like human consciousness occurred as a result of biological evolution.
