a-thosuand-brain.md

a-thosuand-brain.md

• we have thousands of models of the world in our brain. what we perceive is a concensus among them.
• without pain, the new brain would completely rule over the old brain.

Summary:

• each neocortical column represents reference frames wrt to various objects.

Part 1: understanding of the brian

• we don't understand how intelligence emergences from cells in the brain.
• author claims to have discovered a framework that gives rise to a new way to think about intelligence.
• intelligence is closely tied to models of the world. claim: "most of the cells in nerocortex are dedicated to creating and manipulating reference frames (learning, planning and thinking)"

Ch1: old + new brain

• old brain: spinal cord + brain stem (digestion, breathing, reflexes, etc.,), new brain: neocortex.
• new cortex can't control directly, it needs to ask the old brain to do things. but the old brain can always take back control. (eg. holding your breath)
• neocortex: no clear division, but can be split into 12 parts (based on anatomy). we don't know what most cells are doing. 6 layers of cells. most connections are vertical (perpendicular to surface). cells look similar irrespective of area (vision, touch, hearing, etc.,).
• every part of the neocortex generates movement (it's not just through the sensory neocortical neurons as conventionally told).

Ch2: VM's big idea

• neocrotex grew by copying the same element over and over. VM claims "every part of the neocortex does the same thing", the only difference is what they are connected to. so finding the basic function of the neocortex is good enough.
• 150k cortical columns stacked beside each other. divided into 100 minicoloumns with 100 cells each.
• people can learn calculus -- nothing to do with evolution. this shows that humans evolved general purpose learning algorithms.
• our quest to understand intelligence can be narrowed down to what a cortical column does and how it does it.

Ch3: model of world in your brain

• thinking of neocortex as input output system is unhelpful. neocortex makes predictions is helpful. each cortical columns makes predictions.
• neocortex makes models of the world and makes predictions based on these models. brain makes unconscious predictions, they come up in conssciousness when the predictions are wrong.
• brain learns a model by observing how sensory info changes as we move. so, Q: how does the neocortex learn a predictive model of the brain through movement? if we answer this question, it will help AI too.
• tenets: (i) thoughts are mental states, (ii) memory is in the neural connections.

Ch4:

• not many prediction neurons in the brian. so seperation between reality neurons and prediciton neurons.
• pyramidal neurons: what are distal synapses responsible for? it's not clear. jeff's theory: distal synapses flip the neuron into a predictive mode (fire earlier than they should). the predictive firing inhibhits other neurons. expected input: only predictive neurons fire. unexpected input: much more firing because there's no inhibhition.
• most predictions happen inside the neuron. AI models built based on these principles were promising.
• how does the prediction depend on the movement? neurons can attach reference frames to objects we're interacting with. neocortex must know the position of every part of our body relative to the object we are interacting with.
• claim: each neocortical column must include reference frames.

Ch5:

• we don't see objects at the sensed locations (eyes / ear) but out there in the world. suggests that brain encodes the locations of objects relative to us.
• maps in the old brain: grid cells are reference frames. place cells are your absolute location.
• maps in new brain: recent theory-- grid cells exist in the neocortex.
• each cortical column models objects with neurons acting like the place cells, grid cells and head direction cells (reference frames for each observed object).

Ch6: not all 150k columns model objects

• imagning objects are also encoded as reference frames. democarcy, money, etc., is also stored as a reference frame. are knowledge is stored wrt to these reference frames.
• "what" region of the neocortex attach frames to objects, "where" ergion of neocortex attach frames to your body.
• abstract concepts are represented with reference frames of 4 or more directions. only with reference frame can we easily parse advanced math concepts
• the brain can reuse reference frames

Ch7: thousand brains theory

• argues conventional theory that brain as a hierarcy of feature detectors is flawed for several reasons (no theory for saccades, no explanation of size of areas, dark neurons in these areas, 3d structures, etc.,).