122 Are the Androids Dreaming Yet? inhibitors and activators like in real synapses. An individual neuron will fire if the sum of all the connections multiplied by the weights reaches a certain pre-determined threshold. A neural network does not run a program in the conventional sense, and must be trained through experience rather like a human brain. The training process allows the weights in the network table to be adjusted to give the correct result. But, unlike the brain, you can read the weights and even save them to a disk. The neural network tables start with random settings. You show the network the letter ‘A’ and adjust the weights in the tables until it gives a positive answer: ‘It’s an A. Repeat the process with the other letters until the network correctly distinguishes them. As you do this a computer algorithm constantly adjusts the weighting tables using a method called “back propagation. At the end of the training process you can show the network some new input and see how it does. For example, a letter ‘A that is in a slightly different font to anything in the training set. Trained neural networks can perform complex tasks such as recognizing faces or making clinical diagnoses, and they can be allowed to modify their weighting tables as they work so they learn from experience in a similar way to a human brain. Strong AI proponents believe making a thinking machine is just a matter of building a really large, fast neural network and working out how to train it efficiently. Quantum Brains Conventional wisdom says each brain cell is a single processing unit making an on-off decision — fire, or don't fire - depending on the state of its neighbors. But, Stuart Hameroff, Professor of Anesthesiology at the University of Arizona, thinks neurons are not the fundamental information-processing unit in the brain. He suggests that this accolade should go to tubulin. Tubulin is a small, versatile protein that self- assembles into filaments rather like the way b