On Intelligence by Jeff Hawkins

Created on 2021-01-04T19:31:41-06:00

Previous attempts at modeling intelligence are not well rooted in biology.
Artificial neural networks have only the loosest relation to neural networks.
Alan Turing -- demonstrating all computers are functionally similar regardless of how they are built.
Universal Turing Machine: given the right operators, time, and enough memory any machine can compute anything.
Neurons could be used to create AND, OR, NOT gates.
Behaviorism: altering behavior by changing inputs and outputs.
Multi-step deductive reasoning is seen as a pinnacle of human intelligence.
Deep Blue winning chess by calculating more potential branches of play; having no intuition.
Chinese Room Experiment
Founding of "Grid Systems" who invented one of the first laptop computers.
Traditional AI was about symbol manipulation and had nothing to do with how the brain works.
Artificial neural networks tangenially relate to how brains work but they are incomplete.
Physical brains deal with time spiking (patterns changing through time)
Physical brains also have feedback connections; the organs controlled by the brain return information back, in greater number than data leaves.
Auto-associative memory: cells that learn patterns impressed upon them and are able to take partial matches and suggest the original symbol.
Theories that are intuitive and simple but ultimately wrong.
Input-output fallacy: a particular misunderstanding of the brain relating from observing its inputs and outputs but not its function.
Cognitive wheel: when an engineered solution works nothing like nature but yet still works.
The brain contains heirarchies; areas which adapt sensory data or translate concepts in to physical motion.
Synapses can boost or inhibit the firing of other synapses.
Neocortex is made up of layers.
Topmost layer consists mostly of axons.
Layers other than the top, eight out of ten cells are pyramid cells.
Pyramid cells connect to their immediate neighbors but also have axons to neighboring clusters.
Pyramid cells have thousands of connections to other cells.
Vernon Mountcastle -- "An Organizing Principle for Cerebral Function", makes the case that every region in the neocortex is the same. Clusters that decode one sense look the same as other senses.
Clusters in the neocortex are partly specialized to their role by means of different connection counts/thickness, minor suitability tweaks.
The best ideas in science are always simple, elegant, and unexpected.
Much of extant neuroscience is about imaging which areas are activated corresponding to certain stimuli.
Neuroplasticity: the brain's ability to repurpose itself.
Spatial patterns: a pattern which appears across a spatial field; when seeing a plane, parts of the plane are mapped to pixels in a grid, the pixels on the grid are the spatial pattern.
Temporal pattern: a series of spatial patterns that happen or change across time.
Paul Bach-y-rita -- developing a means to project visual data to the nerves of the tongue.
100 step rule: an idea that brains take only one hundred computation steps to solve a given problem.
Brains "compute" using very slow mechanisms because they do not compute at all; they are recalling previously solved patterns.
Memory recall follows associative paths.
Neurons are slow; they take ~5ms to complete a full computation cycle.
Invariant representation: storing knowledge in a way that is flexible; it can generalize to similar problems.
Many behaviors become automatic and require active effort and focus to change them.
Autoassociative memory is good at recognizing and completing patterns it has previously seen but does not recognize them if they are transformed in some significant way.
Continuously predicting some world state and being surprised when the prediction is violated; attention drawn to parts of a problem which do not fit in to the current predicted state.
Prediction may be the primary role of the neocortex.
Filling in: when errors, incomplete or nonsense input is replaced with something that statistically seems like it should be there.
Intelligence defined as remembering and predicting patterns.
Human neocortical "units" have six layers.
Memory prediction framework of intelligence: intelligence as an emergent property of a system designed to predict future outcomes.
V1<->V2<->V3 layers and such
Receptive field: where a visual neuron will only respond to a subset of the vision space; it is constrained to a particular bounding box.
Visual cells may also be conditioned to only respond to particular patterns; ex. lines at 30 degree rotations.
Higher level visual cells learn features of composites from lower cells; some trigger whenever a face-like object are being seen.
High level cells are activate whenever an object of a given type are visible; ex the cell for "traffic cone" is active if a traffic cone exists anywhere in view.
Stacking receptive fields: lower level sees a small region (say 3x3,) while a higher level sees a 3x3 of its subordinate 3x3s and so forth.
If an unexpected event happens the signal is thrown upward until somebody does know what to do about it.
Column: a vertical unit of cells that work together for a single neural computation.
Inhibitory cells exist so that columns can be disabled if they do not seem necessary. Certain regions only activate a grouping of columns for given stimuli.
>90% of synapses are from cells outside a given column.
Constant name pattern: a pattern which is relatively constant to represent a particular concept.
Formation of name patterns: layer 2 cells fire when they predict an output based on a given input, layer 3b cells fire when a lower pattern is not found, layer 3a cells fire to inhibit layer 3b when the pattern is present.
Identifying objects as patterns being sent up from sensors which try to match different memories, sending signals down until a new context is found where the sensors no longer mispredict what they are seeing.
Some channels from higher levels may exist which go all the way down to lower or base layers.
Distant dendrites as "coincidence detectors;" when multiple of them fire at once they cause a different layer cell to spike.
Hebbian learning: neurons which fire at the same become correlated; "neurons which fire together wire together."
Memories not moving; they appear to move but are actually re-learned elsewhere.
Basal ganglia: motor control
Cerebellum: precise timing
Hippocampus: short term memory
Hippocampus is sometimes theorized as the "top" of the neocortex; where memories are developed at the topmost layers and pushed downwards.
Hippocampus has several specialized regions unlike the neocortex.
Memories which cannot be recognized by anything at a lower level end up pooled in the Hippocampus; they stay there until they are either purged or occur enough times the memory is strong enough to work its way downward.
Attention as a process which changes whether an autonomous process or a neocortical process handles and responds to an event.
Neocortex has two hard rules: it is not fast, and its rules are simple (but repeated at scale.)
Experts recognize patterns which sit on top of patterns.