Read more: 鈥�Instant Expert: Artificial intelligence鈥�
More than half a century after the introduction of AI, three key developments could augur the emergence of machine intelligence. New insights from neuroscience and cognitive science are leading to new hardware and software designs. The development of the internet provides access to a vast store of global data. It may even be possible for AI to evolve on its own
Advertisement
Brainy machines
If we want to build a computer that performs with human-level intellect, why not just copy the brain? Humans, after all, are our best example of intelligence. In the last decade, neuroscience has provided many new insights about how we process and store information.
The human brain is a network of 100 trillion synapses that connect 100 billion neurons, most of which change their state between 10 and 100 times per second. Our brain鈥檚 layout makes us good at tasks like recognising objects in an image.
A supercomputer, on the other hand, has about 100 trillion bytes of memory and its transistors can perform operations 100 million times faster than a brain. This architecture makes a computer better for quickly handling highly defined, precise tasks.
But some tasks would benefit from more brain-like processing, even with the attendant trade-offs. For example, uncertain tasks like recognising faces don鈥檛 necessarily require highly accurate circuitry in which processing follows a precise path.
Some researchers are looking into brain-like hardware architectures to mimic the brain鈥檚 low-power requirements. The brain does all of its computations on roughly 20 watts, the equivalent of a very dim light bulb. A supercomputer capable of a roughly analogous computations requires 200,000 watts. Other groups are interested in learning from the brain鈥檚 ability to process information and store it in the same place (see diagram). For these reasons, projects are underway to build novel computer circuits inspired by the brain: more parallel rather than serial, more analogue rather than digital, slower, and consuming far less power.
Computers call on intuition
Humans persistently fail to live up to the ideal of rationality. We make common errors in our decision-making processes and are easily influenced by irrelevant details. And when we rush to a decision without reasoning through all the evidence, we call this trusting our intuition. We used to think the absence of such human quirks made computers better, but recent research in cognitive science tells us otherwise.
Humans appear to have two complementary decision-making processes, one slow, deliberate and mostly rational, the other fast, impulsive, and able to match the present situation to prior experience, enabling us to reach a quick conclusion. This second mode seems to be key to making human intelligence so effective.
While it is deliberative and sound, the rational part requires more time and energy. Say an oncoming car starts to drift into your lane. You need to act immediately: sound the horn, hit the brakes, or swerve, rather than start a lengthy computation that would determine the optimal but possibly belated act. Such shortcuts are also beneficial when there is no emergency. Expend too much brain power computing the optimal solution to details like whether to wear the dark blue or the midnight blue shirt, and you鈥檒l quickly run out of time and energy for the important decisions.
So should AI incorporate an intuitive component? Indeed, many modern AI systems do have two parts, one that reacts instantly to the situation, and one that does more deliberative reasoning. Some robots have been built with a 鈥渟ubsumption鈥� architecture, in which the lowest layers of the system are purely reactive, and higher levels serve to inhibit the reactions and organise more goal-directed behaviour. This approach has proved to be useful, for example, for getting a legged robot to negotiate rough terrain.
There has been a similar push to motivate AIs to make better decisions by giving them emotions. For example, if an autonomous robot tries the same action a few times and fails repeatedly, a 鈥渇rustration鈥� circuit would be an effective way of prompting it to explore a new path.
鈥淚f an autonomous robot tries the same action a few times and fails repeatedly, a 鈥榝rustration鈥� circuit would be an effective way of prompting it to explore a new path鈥�
Creating machines that simulate emotions is a complicated undertaking. Marvin Minsky, one of the founders of AI, has argued that emotions arise not as a single thing that brains do, but as an interaction involving many parts of the brain, and between the brain and the body. Emotions motivate us to choose certain decisions over others, and thinking of the parts of a computer program as if they were motivated by emotions may help pave the way for more human-like intelligence.
Evolution
Most modern AI systems are too complex to program by hand. One alternative is to allow the systems to evolve themselves. In this approach, an iterative process tries out variations of the program in a virtual environment, and chooses the variations that are most successful and make the best decisions, using trial and error.
First, the designers build a simulation of the program鈥檚 environment 鈥� perhaps a desert environment or a pool. Then they set several different AI designs loose inside the simulation, and measure the quality of their decisions 鈥� in other words, their fitness. Perhaps one variant of the AI program quickly finds rewards but another variant is slow. Unfit designs are discarded, modifications are made to the fit designs, and the process repeats.
Modifications that change some parameter of a single individual design can be likened to random mutations in natural selection. Modifications that combine two different designs to produce a third 鈥渙ffspring鈥� design are not unlike sexual reproduction. For that reason, they are referred to as genetic algorithms.
