Chapter 6: How the Cortex Works

• Traditional approaches to understanding the brain have taken a bottom-up approach.
• Hawkins argues that a top-down approach is needed.
• The memory-prediction model is a top-down approach.

Invariant Representations

• Visual Regions: V1 <-> V2 <-> V4 <-> IT
• V1: spatially specific, fast changing, feature level
• IT: spatially invariant, slow changing, object level
• What happens flows up, predictions flow down.
• Other regions such as auditory, touch and motor are similar.
• The top 2 levels allow the senses and motor control to be integrated.
• In general, the 6 layer hierarchy can be seen as a tree. Each region of the cortex can form invariant representations from its inputs.

A Model of the World

• The cortex learns the hierarchical structure of the world.
• A sequence is a set of patterns that accompany each other but not necessarily in a fixed order.
• The cortex learns sequences of sequences.
• Names (e.g. eye or face) can be associated with each sequence.
• Bottom-up classifications and top-down sequences constantly interact. They can also change over time - this is learning.
• Brain cells are loosely organized in columns. Within a column, 90% of the synapses come from outside the column. This provides the column with context.
• Hawkins claims that columns are the basic unit of prediction.