At a recent client meeting in London, I found myself in a curious debate about artificial intelligence (AI) and the future of machines in decision making. We were debating the pros and cons of neural networks, when the conversation turned suddenly to processing power, and model-less applications of computing.
A recent experiment at the Imperial War Museum in London by Enigma Pattern demonstrated how a simple algorithm could break the famed Enigma code of World War Two. In the experiment, Enigma was broken in less than 14 minutes, largely because of the application of parallel computing: 1,000 servers were used to test 13 million password combinations a second, and a simple language recognition algorithm was used to test if the decoded output was, in fact, German. In total, 15.4 billion combinations were tested, evaluated (for German), at a grand cost of $7 for the rental of the parallel servers. There were no AIs in sight, no sophisticated pattern recognition: this was a phenomenal demonstration of the awesome power of sheer ‘brute force’ search.
The awesome power of ‘brute force’ search
‘Brute force’ search is the act of searching through vast permutations (15 billion in this case) of possibilities to solve a problem or a code, testing and re-testing every possible permutation again and again and again. There is little finesse or elegance in this method, just an enormous tidal force of raw computing power. Consider an analogy of learning to fly an airplane. A learning algorithm would try to understand the difference between the switches and controls, turning them on and off, all the while trying to identify a system by which it could fly the plane. A brute force algorithm would try every single possible permutation of each control with every other control, until (billions of attempts later) the airplane takes off. This is an untenable solution unless you have incredible amounts of computing power at your disposal (or infinite time in which to solve the problem). And, for many current problems, the computing power required doesn’t exist, yet. For example, brute force is currently not powerful enough to win at a game like ‘GO’, as the permutations of moves are beyond those of our processors today. A system was built that could estimate incremental probabilities of each move that would, eventually, lead it to victory, but, importantly, that system had to rely on a model that would estimate probabilistic moves to beat the best human players in the world because it was unable to map out every possible permutation of the game.
Entering the realm of quantum supremacy
However, raw processing speeds are about to get a serious boost with quantum computing looming on the horizon. Both Intel and IBM recently announced that they have functioning 49 and 50-qubit (quantum bit) processors, while Google has also been building its own quantum computer. Quantum computers are still very early in their development, but are widely expected to overtake conventional computers in processing speed, dramatically named ‘quantum supremacy’. The promise of quantum computers is a little like a ‘Machine Hulk’ to put it into superhero terminology. It is raw, unadulterated, processing power, the likes of which we have never seen.
The advent of quantum supremacy has a number of major implications both for AI and the world at large. Just as parallel computing cracked the Enigma code like an egg, so quantum computing will have enough juice to break through just about any present day security system using terrifying brute force, with little reliance on elegance or a model. This will mean a security systems themselves will need to become more complex to keep up.
It also has a bearing on the future of AI because many of today’s accomplishments are, in large part, due to sophisticated models like neural networks that, while flexible, still try to understand their problems using a modelling framework. Both supervised neural networks and unsupervised deep learning networks follow a basic model, a structure for estimation, that have parameters, inputs, and ‘toggles’ that allow the researcher to shape the architecture of the model. With the advent of quantum computing, many doors that, until now, have only been able to be opened by elegant deep learning algorithms, will be smashed in by the ‘Machine Hulk’.
Don’t discount how quickly the tide of computing power is rising.
The world has become increasingly fixated on deep learning as a great methodology, but it is still only one methodology, one class of model. An analogy is perhaps when we teach ourselves to draw a face, first by starting with the outline, then filling in detail. This is a method or a ‘model’ of drawing, and we improve in our skill by going through the components, section by section. The ‘brute force’ equivalent is to draw every face imaginable until one makes sense. While it may not sound like the best way to draw, it does work on many smaller problems. It is easy to imagine, for example, standing at your front door, late at night, trying every key on your keychain until one works.
The model-less application of computing power is a big deal, and will probably be much more critical in the future. No matter how much stock you place in the models behind AI, neural nets and machine learning, the one constant is that computing power is constantly increasing, and problems that seemed complex will increasingly become broken through sheer brute force, without necessarily a reference to a (neural network or otherwise) model. Enigma is an easy example of a puzzle that was simply overtaken by the rising tide of processing power, not by ingenuity in data modelling. With quantum computing on the horizon, the tide is about to start rising much faster.