The gloomy old orthodoxy about human brain cells was that, in the adult at least, they never divide or re-grow. Brain injury or deterioration was therefore irreversible. This belief now seems pessimistic. The human hippocampus, at least, can regenerate neurons; and mouse brain cells, encouraged by epidermal growth factor, have been successfully cultured in vitro. Last week Daedalus was exploiting these facts with his ‘brain pan’ — a single layer of brain cells cultured in a petri dish. The idea was for the cells to put out dendrites and axons, which would connect at random with each other. Electrodes lowered onto the cells could fire them, and eavesdrop on the resulting neural communication.

Daedalus now reckons that he has invented the ultimate neural-net computer. The ‘strong’ theory of neural-net and brain action, he says, is either brilliant or absurd, or possibly both. It asserts that a planless array of randomly wired neurons can learn anything. Neural connections which are reinforced become more easily triggered and more active; those which are neglected become less active. So an utterly planless network like his brain pan could be ‘taught’ to act as a computer. Even its wiring might grow to meet demand. Routes carrying heavy traffic could well grow more dendrites to carry it.

Daedalus is setting up a brain pan to test these claims. He is fitting it with a dense array of peripheral electrodes to simulate input (sensory) nerves, and another dense array of output (motor) connections. He will feed typical neural-net tasks into the input array, and assess the response on the output one. Correct responses will be reinforced by repetition; incorrect ones will be extinguished by neglect or counterexample. The first experiments will use mouse brain cells. Human ones might (or might not) be cleverer, but would raise ethical concerns.

The brain pan's education will start with simple signals for elementary evaluation, and will progress towards ever greater subtlety. Ultimately it will be given microphone and TV-camera inputs and hydraulic-actuator outputs, and encouraged to learn its way round the real world — like a mouse or human being. Success will raise deep philosophical problems about minds, meanings and consciousness. Failure will reinforce Daedalus's nagging suspicion that we know nothing whatever about how the brain works.