The Experience Machine

Unboxing the Experience Machine

• The number of neuronal connections carrying signals backward in this way is estimated to exceed the number carry signals forward by a very substantial margin.
• The brain, at 2% of human body weight is estimated to account for around 20% of total bodily energy consumption.
• The bulk of what the brain does is learn and maintain a kind of model of body and world - a model that can then be used, moment by moment, to try to predict the sensory signal. These predictions help structure everything we see, hear, touch, and feel.
• As a brain encounters new sensory information its job is to determine if there is anything in that incoming signal that looks like important "news" - unpredicted sensory information that matters to whatever it is that we are trying to see or do.
• Hermann von Helmholtz - "given everything I know, how must the world be for me to be receiving the pattern of signals currently present?" This is the question that perceptual systems are built to resolve.
• If the signal is poor, the brain chruns out "good hallucinations" by filling in and fleshing out the missing signal according to what it expects to hear.
• The "top-down" flow of information
• "Controlled hallucination" - When inner guessing completely rules the roost, we are just hallucinating, full stop. But when it is appropriately sensitive to sensory stimulations - via prediction error signals - the guessing is controlled, and the world becomes known to the mind.
• Linear predictive coding (Cluade Shannon) - The trick is trading intelligence and foreknowledge against the costs of encoding and transmitting all the information. Transmit only whatever turns out to be different from the predicted patterns. Wherever there is detectable regularity of any kind, prediction (and hence this form of data compression) becomes possible.
• However complex or high-level the predictions, it is prediction errors that must then carry the news, signaling differences from the expected and thereby keeping us in touch with a changing and sometimes surprising world.
• In the predictive processing architecture of the brain, it is thought that different neuronal populations specialize in different things, so that each "higher" level can use its own specialized knowledge and resources to try to predict the states of the level immediately below it. A level that specializes in predicting whole words might use its knowledge to help predict states at a lower level whose specialty is recognizing letters. The word level might be predicted by a higher level that specializes in sentences.
• All that flows forward is news - deviations from what is expected. This is efficient.
• Our own actions and histories sculpt the onboard prediction machinery that in turn sculpts human awareness, right down to the level of what seem to us to be basic sensory experiences.
• Millions of years of evolution have determined the bedrock configuration of the machinery we command at birth: the early wiring of the brain, the structure of our sense organs, and the shape of our bodies. Courtesy of all that, we start our journey already armed with plenty of hard-won knowledge.
• Creatures with lungs are already structurally "expecting to breathe".
• Creatures like us specialize in learning about their worlds on the basis of repeated sensory encounters. And we drive learning by trying to predict our own sensory flows.
• The brain learns by looking for better and better ways to predict that unruly sensory barrage. Very young infants seem to spend most of their time doing just this, trying to find useful patterns in the sensory stream.
• One way to predict quite a lot about the most likely next word in a sentence is to implicitly know a lot about grammar. And a good way to learn grammar is to try, again and again, to predict the next words you are going to hear. As those attempts continue, your brain can slowly, unconsciously, discover the regularities that will enable you to do a better job.