Questions & Answers
As people read On Intelligence, they often have similar questions. Below is a list of questions in which multiple people seem to be interested. I will attempt to answer them as time permits. If there is something you want to know you can look here and in the discussion forum. You are welcome to suggest additional questions by submitting them through the contact form. I will try to answer them, but please don't expect that I will be able to answer each one or respond in a timely fashion. Also, be warned that the information below is speculative and without references. It represents my current opinion.
Here is a list of questions on which I am working. Right now I have only answered one.
- Does prediction only occur in the neocortex?
- What exactly is happening when the cortex makes a prediction?
- Can you explain further how sequences are stored?
- What exactly is the definition of intelligence?
Q: Does prediction only occur in the neocortex?
A: No. Prediction is almost certainly happening in older parts of the brain as well, but the mechanisms for doing so are likely to be somewhat different than those used in the neocortex, and certainly the predictions will not be as sophisticated as those made by the neocortex.
Recall that reptiles and other animals without a neocortex can have sophisticated senses and flexible behavior. All behavior requires some level of prediction. So we can be sure some level of neural prediction does not require a cortex. Sub-cortical areas of mammals will also make predictions. For example, when a newborn horse prances about shortly after birth we can be certain that its nervous system already has a partial model of the world and is making predictions about what it will feel next as each hoof comes down. However, no other part of the brain is built like the neocortex. Most parts of the old brain are highly evolved heterogeneous structures. The cerebellum is an exception; it has a highly regular structure. But no other part of the brain, including the cerebellum, has the large hierarchical organization of the neocortex. Therefore, it appears that the neocortex is the only brain region that has ability to form the kind of nested, hierarchical memories and corresponding predictions described as part of the memory-prediction framework.
One commonality among all parts of the brain is the neuron. In my view every neuron with modifiable synapses is at its core a prediction machine. For example, a typical pyramidal cell has a number of synapses on the cell body or close to the cell body. These are often referred to as proximal synapses. The proximal synapses are believed to have a strong effect and can force the cell to fire. They are sometimes referred to as "driver" inputs. When a cell is driven to fire by its proximal synapses, you can imagine it then looks on its many thin dendrites for synapses that were active at that moment and strengthens them. The synapses on distant or thin dendrites are sometimes referred to as distal synapses. In the future, when those same distal synapses are active again, the neuron recognizes that pattern and fires in anticipation of the driver input. This is classical Hebbian learning, which I argue is just a way for a cell to predict its driver input. The model I describe in the book has a few additional twists. One, it assumes that only a few active distal synapses on thin dendrites are sufficient to have a large effect at the cell body. They can make the cell fire even though they are far away. Two, I suspect synapses are formed through this hebbian-like process, not just strengthened. That is, if a dendrite and axon that pass near each other and the axon is active at the same time as the cell with the dendrite, they can form a new synapse. The formation of new synapses in this way is not a new idea, but it is speculative. But in either case a neuron with Hebbian learning can be viewed as a prediction machine. Therefore prediction is in some sense occurring throughout the nervous system, everywhere Hebbian learning is occurring.