The buzz around artificial intelligence and machine learning invites a great deal of interest, but unfortunately it is also plagued by a great deal of misinformation. It’s quite astonishing to see how so many profess to be experts in machine learning without possessing even the basic understanding of its principles and, more importantly, its limitations. To date, machine learning has made great strides in automating statistical referencing. Simply stated, the machine is trained using historical samples of inputs and outputs and is able to infer outcome from new, unseen data.
Take autonomous vehicles, for example. Most driverless cars utilize an algorithm called “reinforcement learning,” a machine learning technology predicated on behavioral psychology in which “software agents” are trained to determine the optimal decision for a given driving state. Specifically, humans will train an autonomous vehicle by taking a software agent for a ride and allowing the “agent” (or the machine) to observe and record each and every state in which the human makes a decision while driving. The machine also scores how favorable or unfavorable the outcomes of these decisions were. Over time, as the algorithm rewards good behavior while gathering many millions of references, we end up with a pretty good software agent that knows what the optimal decision should be in all common (and most uncommon) driving situations. As remarkable and commercially viable autonomous vehicles are, the technology is relatively simple due to the finite number of states and outcomes (decisions) the driving agent(s) faces at any point in time. In many cases, making a driving decision is nothing more than a brute force search (a simple process of searching through all possible scenarios).
The limitations of machine learning technology are still real as machines still can’t think like humans.
With all the advanced technology and terms such as “genetic algorithms” or “neural networks,” machines still can’t truly develop new intelligence, not to mention apply intuition, instincts, cognitive thinking, or strategic reasoning. Moreover, machines can’t develop superior intelligence to that of humans, as they are bound by the biases and rules of the people who created them.
Until recently many scientists believed that we are, at a minimum, decades away from such technological breakthroughs, but something remarkable took place last year when Google AlphaGo consistently beat the world’s best human Go player.
What Is Go?
Go is an ancient Chinese board game that was invented more than 2,500 years ago. The rules of the game are very simple as the game is played between two players taking turns, with one placing white and the other black stones on the board. When one player surrounds another player’s stone(s), the territory surrounded is then surrendered to him/her by which the player is able to gain a larger footprint on the board. The winner of the game is the one who covers the largest territory on the board. Despite its relatively simple rules, Go is very complex, even more so than chess. Compared to chess, Go has both a larger board (17 * 17) with more scope for play and longer games and, on average, many more alternatives to consider per move.
Image 1: A board state in Go – credit Wikipedia
It is estimated that there are well over 50 million Go players worldwide, with the majority of them living in East Asia. As of December 2015, the International Go Federation has a total of 75 member countries and four association membership organizations in multiple countries.
A Huge Breakthrough: Google’s AI Beats A Grand Master Top Player At The Game Of Go
Traditionally, machines were able to come out on top in games set to measure human intellect. For example, the highly publicized battle of human vs. machine in 2011, in which IBM’s super computer Watson beat Ken Jennings (a sequential champion) in Jeopardy. Even earlier in 1997, Deep Blue topped world chess champion Gary Kasparov. We now know that machine intelligence had little to with the above accomplishments since in both Jeopardy and chess, Watson utilized a simple brute force searching algorithm designed to scan a large and comprehensive database. In chess, for example, Watson analyzed the outcome of every possible move and sequentially all consequential moves, looking further ahead than any human possibly could.
With Go however, this type of search is not possible simply due to the sheer number of available moves. In Go, there are more possible moves than the number of atoms in the universe. The grand master champions of Go all possess a great deal of strategic intellect that until recently could not be matched with machine learning. AI scientists from the likes of Facebook, Google, IBM, and Microsoft searched for new approaches in which they could apply machine learning technology that will be able to consistently win in Go. The research naturally gravitated towards deep learning, an advanced machine learning technology predicated on hierarchical layers of networks suited for pattern analysis of unstructured (or unlabeled) data.
Deep learning is a subset of unsupervised learning, whereby the machine classifies patterns of large amount of data by processing the data from one layer of classification to another without human guidance. Deep learning in combination with neural networks was the technology most suitable for the task. Neural networks (or “neural nets,” for short) is a machine learning discipline that is mainly used for pattern matching in the context of facial, voice, or handwritten recognition. Rather than starting from an empty board and analyzing each and every move, the AI scientists were looking from the end winning state backwards. Their primary goal was to construct a very large database of patterns on the Go board that would most likely lead to a win. By incorporating deep learning in combination with neural nets the machine was able to abstract the makeup of winning patterns based on their characteristics vs. just the raw visual pattern itself. In a way, given enough sample patterns, the system is able to find other patterns that hold the same characteristics and abstract a winning pattern – in other words, – “learn”.
The biggest breakthrough however came when the scientists of Google’s AlphaGo realized that rather than having humans feed the deep learner with winning patterns, why not let computers play against each other and get smarter as they find new winning patterns on their own? In short, take the humans constraints or biases completely out of the equation.
On May 23, 2017, TechCrunch published: “Google’s AlphaGo AI Go Player Has Defeated Ke Jie, Go world champion! The win was by a narrow margin, but AlphaGo has been programmed to ensure victory, not to run up the score or devastate its opponent.” Jie said that AlphaGo’s performance left him “shocked” and “deeply impressed” in post-match statements, noting that the moves the computer played “would never happen in a human-to-human match.”
An achievement that only a few short years ago was thought of as impossible or at a minimum, decades away, is about to become main stream and influence our ways of lives in a big way. The conundrum of creating machines that can learn from themselves and beat human on the intangibles of cognitive intelligence such as intuition, or strategic thinking is a thing of the past. We are entering a new era that is both exciting, and maybe a little frightening. Only time will tell but the big data/machine learning revolution is still in its infancy, with plenty of runway left to accelerate.
Erez M. Katz
CEO / Lucena Research Inc.