On Intelligence - Jeff Hawkins

Neo Cortex forms a uniform structure: 6 layers of cortical columns composed mostly of pyramidal neurons. Information flows bidirectionally: down the hierarchy to relate to known patterns and up the hierarchy to predict unknown / complete partial patterns.

Brain does not compute mathematical equations to navigate the world but creates and recalls memories. Catching a flying ball involves (muscle sequence) memories of previous catches that can be efficiently parameterized to adapt to the current flying ball situation.

Knowledge Representation (neo cortex)

• sequential (easy to recite A-Z, hard to do the reverse)
• invariant (to geometric translations in space and time)
• auto-associative (informations connect by observing in shared context)
• hierarchical (complex/abstract patterns are composed of simple patterns)
• SDR: sparse distributed representations
• topographic: neighboring neurons, activate to similar patterns

Learning / Forming Memories

• fire together, wire together (Hebbian learning): similar patterns that repeatedly fire together, will bind together (even across different input domains), translational invariance is not only spatial but also temporal (to bind two stimuli they must not co-occur exactly at same time)

Grandmother cell hypothesis is wrong: Each neuron doesn't detect just one object. This wouldn't allow lifelong learning or resilience to cell death - instead concepts are SDRs.

#book #AI