Deep Learning Is Going to Teach Us All the Lesson of Our Lives: Jobs Are for Machines
Adam Rifkin stashed this in Robot Jobs
We have gone from linear to parabolic.
On December 2nd, 1942, a team of scientists led by Enrico Fermi came back from lunch and watched as humanity created the first self-sustaining nuclear reaction inside a pile of bricks and wood underneath a football field at the University of Chicago. Known to history as Chicago Pile-1, it was celebrated in silence with a single bottle of Chianti, for those who were there understood exactly what it meant for humankind, without any need for words.
Now, something new has occurred that, again, quietly changed the world forever. Like a whispered word in a foreign language, it was quiet in that you may have heard it, but its full meaning may not have been comprehended. However, it’s vital we understand this new language, and what it’s increasingly telling us, for the ramifications are set to alter everything we take for granted about the way our globalized economy functions, and the ways in which we as humans exist within it.
The language is a new class of machine learning known as deep learning, and the “whispered word” was a computer’s use of it to seemingly out of nowhere defeat three-time European Go champion Fan Hui, not once but five times in a row without defeat. Many who read this news, considered that as impressive, but in no way comparable to a match against Lee Se-dol instead, who many consider to be one of the world’s best living Go players, if not the best. Imagining such a grand duel of man versus machine, China’s top Go player predicted that Lee would not lose a single game, and Lee himself confidently expected to possibly lose one at the most.
What actually ended up happening when they faced off? Lee went on to lose all but one of their match’s five games. An AI named AlphaGo is now a better Go player than any human and has been granted the “divine” rank of 9 dan. In other words, its level of play borders on godlike. Go has officially fallen to machine, just as Jeopardy did before it to Watson, and chess before that to Deep Blue.
The combination of deep learning and big data has resulted in astounding accomplishments just in the past year.
Aside from the incredible accomplishment of AlphaGo, Google’s DeepMind AI learned how to read and comprehend what it read through hundreds of thousands of annotated news articles. DeepMind also taught itself to play dozens of Atari 2600 video games better than humans, just by looking at the screen and its score, and playing games repeatedly. An AI named Giraffe taught itself how to play chess in a similar manner using a dataset of 175 million chess positions, attaining International Master level status in just 72 hours by repeatedly playing itself. In 2015, an AI even passed a visual Turing test by learning to learn in a way that enabled it to be shown an unknown character in a fictional alphabet, then instantly reproduce that letter in a way that was entirely indistinguishable from a human given the same task. These are all major milestones in AI.
However, despite all these milestones, when asked to estimate when a computer would defeat a prominent Go player, the answer even just months prior to the announcement by Google of AlphaGo’s victory, was by experts essentially, “Maybe in another ten years.” A decade was considered a fair guess because Go is a game so complex I’ll just let Ken Jennings of Jeopardy fame, another former champion human defeated by AI, describe it:Go is famously a more complex game than chess, with its larger board, longer games, and many more pieces. Google’s DeepMind artificial intelligence team likes to say that there are more possible Go boards than atoms in the known universe, but that vastly understates the computational problem. There are about 10¹⁷⁰ board positions in Go, and only 10⁸⁰ atoms in the universe. That means that if there were as many parallel universes as there are atoms in our universe (!), then the total number of atoms in all those universes combined would be close to the possibilities on a single Go board.
Such confounding complexity makes impossible any brute-force approach to scan every possible move to determine the next best move. But deep neural networks get around that barrier in the same way our own minds do, by learning to estimate what feels like the best move. We do this through observation and practice, and so did AlphaGo, by analyzing millions of professional games and playing itself millions of times. So the answer to when the game of Go would fall to machines wasn’t even close to ten years. The correct answer ended up being, “Any time now.”