A recent article on the BBC (and the highly recommended MIT news) breaks the news on an innovative silicone chip that models neuronal architecture and neuronal communication. The chip’s 400 transistors mimic the head of a neuron: they summate the analog signals received from other chips. When such signals reach an adjustable limit, they cascade into an action potential, just as in neurons. Depending on their arrangement and organization, these action potentials can have an excitatory, or inhibitory effect on their neighbours, analogous to their biological counterparts.
This kind of modelling is exciting and interesting – for it is profoundly different from other contemporary methods of modelling brain activity. While a great overgeneralization, most other programmes model the brain’s circuitry – the neurons, the synaptic connections, the action potentials – in a virtual space. They exist as computer code, or interacting objects created by such code. These coded objects, whatever existence they have, model the function of neurons. These chips, in comparison, are an actual model of a neuron. And this is the important difference between the two paradigms: that between modelling function and form.
This distinction has profound implications for Artificial Intelligence, as both articles rightly mention. In other paradigms, predominantly connectionist or GOFAI (Good Ol’ Fashioned AI), speed has been the limiting factor. The ability to model of human-like intelligence rested on Moore’s law, the ever doubling number of transistors on chips, and the associated increase in calculations per second. So, while we already have the theory, and the architecture with which one could construct and entire connectionist model of the brain – we just don’t have the computers yet to execute this model. (The Blue Brain Project is perhaps the most famous of recent attempts to do just this.)
However, when it comes to modelling human-like intelligence with MIT’s neuron-like chip, the limiting factors change: the modelling is limited by the number of chips (though silicon chips are cheap and easy to make, the article does not make entirely certain that these chips are so easily mass produced), and the extent to which these chips can be physically linked together. Real neurons, as we know, can have hundreds of dendritic connections to other neurons. The article does not mention the extent to which these chips can be connected – and it may turn out to be a limiting factor in this current generation of chips.
These engineering problems do not immediately strike me as daunting. But we must be cautious – there are many other signalling methods in the brain: the diffusion of gases; sensitivity-altering neurotransmitters; and so on. These chips, while an intriguing model of the biology of the brain, are but the first step to building silicone brains like ours. And there are many interesting conceptual and philosophical issues to be sorted out along the way.