A neurobiological perspective on building intelligent devices

What is ‘intelligence’ and, for that matter, what problems might arise in building an intelligent machine? Are human brains, with their greatly expanded neocortex, the only currently-existing intelligent devices? Apart from some unconvincing computer programs, the only known intelligent devices seem to be animals: particularly birds, and especially mammals. There is, however, at least one other clear example of natural ‘intelligence’: all living organisms, notoriously, appear to be intelligently designed, even though this appearance is achieved by selective amplification of molecular accidents. This form of natural intelligence (i.e. the ‘Darwinian algorithm’ comprised of iterative replication/mutation/transcription/translation/selection steps) is the only other successful exemplar of ‘intelligence’ we have identified to date. It is also a good source of clues to help us navigate the neo-cortical labyrinth. A good place to start our inquiry is to ask what is going on inside the skull (see Figure 1). There are two basic processes: the first is a rapid (millisecond) ‘integration’ step, in which synaptically weighted voltages are collected over the surface of a neuron, combined (possibly in a nonlinear way), and sent via more synapses to other neurons. There is also a slower ‘learning’ process that uses the rapid signals to modify the weights such that performance improves. Learning is done by adjusting the strength of individual synapses according to the voltage across the synapse (Hebb’s Rule). The power of the learned world model will reflect the extent to which the synapses can individually be set (much as the power of a digital computer reflects the number of transistors and memory locations that can be individually— and sufficiently rapidly—controlled. Intelligence boils down to numbers: the combinatorial potential vastness of the world should be matched by a corresponding potential combinatorial vastness of the brain that models it, together with precise rules (such as Hebb’s) for selecting useful combinations. Integration requires voltage spread, but accurate Figure 1. The left hand picture illustrates the interaction between an animal’s brain (B) and its world (W). The brain’s input-output relation reflects its synaptic weights, which depend on the history of ancestors (gray zone, ‘genes’) and, especially in complex animals, on the history of the animal itself (’learning’). The right hand picture shows the two main components of a mammal’s brain; the subcortical structures (which learn pairwise correlations), and the neocortex (which learns higher order correlations). The neocortex relies particularly heavily on learning, and provides corrections to subcortical computations.