A Thousand Brains

Knowledge Graph: A Thousand Brains (Jeff Hawkins, 2021)

Editorial spotlight: 150,000 cortical columns → each builds complete model → all vote simultaneously

Concepts

• Hawkins's Thousand Brains Theory (importance 5): Every cortical column in the neocortex learns complete models of objects and concepts. The brain doesn't have one model of your coffee cup — it has thousands, distributed across ~150,000 columns, all voting to reach consensus.. Source: (from training memory of book).
• Hawkins's cortical column (mini-brain) (importance 5): A vertical slice of neocortex ~1mm across containing ~100,000 neurons. Hawkins argues each column is functionally equivalent — all perform the same learning algorithm on different inputs. The structural unit of intelligence.. Source: (from training memory of book).
• cortical reference frames (grid cells for everything) (importance 5): The brain represents all knowledge using reference frames analogous to GPS grid cells. Every object, concept, even abstract ideas exist as locations and movements through learned reference frames. This is the representational foundation of the Thousand Brains Theory.. Source: (from training memory of book).
• Hawkins's old brain (survival circuits) (importance 4): The evolutionarily ancient subcortical structures (brainstem, limbic system) that generate goals, emotions, drives. The old brain tells the neocortex what to care about; the neocortex figures out how to achieve it.. Source: (from training memory of book).
• Hawkins's new brain (neocortex) (importance 4): The neocortex — the thin sheet that expanded massively in mammals. Hawkins's central claim: this entire structure runs one learning algorithm, replicated across ~150,000 columns. Intelligence emerges from massive parallelism of this algorithm.. Source: (from training memory of book).
• Numenta displacement cells (object locations) (importance 4): Hawkins's extension of grid cells to object modeling. Neurons encode 'where on the object' you're sensing. As you touch different parts of a cup, displacement cells track location relative to cup's reference frame.. Source: (from training memory of book).
• Hawkins's complementary learning (old + new) (importance 4): The old brain learns slowly and emotionally. The new brain learns quickly and models the world. Intelligence requires both: the old brain provides goals, the new brain provides understanding. Remove either and you have dysfunction.. Source: (from training memory of book).
• Hawkins's sensory-motor learning (importance 4): You can't learn object structure from passive sensing alone. Movement is essential: as you move your finger, sensors report changing features, and motor commands provide 'where you moved'. Columns integrate both to infer structure.. Source: (from training memory of book).
• Hawkins's nested reference frames (importance 4): Objects are reference frames nested inside larger reference frames. A cup has a handle (sub-frame). The cup sits on a table (parent-frame). The table is in a room (grandparent-frame). Knowledge is hierarchies of reference frames.. Source: (from training memory of book).
• Hawkins's universal mapping principle (importance 4): Everything the brain knows is represented as maps (reference frames). Visual space is map of retinal locations. Auditory space is map of frequencies. Even abstract concepts are locations in conceptual maps. Maps are the universal format.. Source: (from training memory of book).
• Hawkins's comprehensive world model (importance 4): The sum of all reference-frame models across all columns is your world model. Every object, concept, person, place you know exists as reference-frame structure somewhere in cortex. World modeling is the brain's function.. Source: (from training memory of book).
• Hawkins's knowledge genome (importance 3): Proposal to create shared repository of learned models in reference-frame format. Like Human Genome Project but for knowledge. Machines could download and build on accumulated understanding rather than learning from scratch.. Source: (from training memory of book).
• Hawkins's temporal sequence learning (importance 3): Cortical columns learn time-based patterns — musical melodies, speech phonemes, motor sequences. Columns predict next element; prediction failures update model. Time is fundamental dimension of cortical learning.. Source: (from training memory of book).
• Hawkins's generalization (reference frame transfer) (importance 3): Once you learn reference frame for one object, you can apply to similar objects. All chairs share 'chair-ness' reference frame despite shape differences. Generalization is reference frame abstraction.. Source: (from training memory of book).
• Hawkins's attention (old brain selection) (importance 3): Attention is old brain deciding which sensory inputs matter right now. Thalamus gates information flow to cortex. Neocortex builds models of everything; old brain decides which models to activate based on current goals.. Source: (from training memory of book).
• Hawkins's symbol grounding (reference frames) (importance 3): Classic AI problem: how do symbols connect to real-world meaning? Hawkins's answer: symbols are reference frames. 'Cat' means a particular structure in cat-reference-frame space. Grounding is built-in.. Source: (from training memory of book).
• Hawkins's cross-modal voting (importance 3): Different sensory modalities vote together — touch columns and vision columns vote on 'what object is this'. Cross-modal agreement creates robust recognition. This is why you recognize objects better when you can see AND touch them.. Source: (from training memory of book).
• Hawkins's compositional models (importance 3): Complex objects are compositions of simpler reference frames. A car is wheels + chassis + engine, each with own reference frame, composed into car-frame. Learning reuses components. Explains why humans learn efficiently.. Source: (from training memory of book).
• Hawkins's mental simulation (prediction) (importance 3): Thinking is simulating movement through reference frames without actual movement. You can imagine walking through your house by activating house-reference-frame and generating movement predictions. Thought is internal navigation.. Source: (from training memory of book).
• cortical layer 6 (feedback) (importance 2): Deepest layer of cortex with connections back to thalamus and other regions. Hawkins speculates layer 6 may carry predictions back down the hierarchy, enabling predictive voting.. Source: (from training memory of book).
• catastrophic forgetting (absent in cortex) (importance 2): Problem in neural networks where learning new task erases old knowledge. Cortex doesn't have this — you can learn French without forgetting English. Hawkins argues reference frames + voting solve this naturally.. Source: (from training memory of book).
• cortical long-term memory (synaptic weights) (importance 2): Memories are stored in pattern of synaptic connections within and between columns. Recalling memory means reactivating same neural pattern. Forgetting is synaptic pruning. All memory is cortical structure.. Source: (from training memory of book).
• cortical working memory (sustained activation) (importance 2): Short-term memory is sustained neural activity in cortical columns. You hold phone number in mind by keeping relevant columns active. Limited capacity because only so many columns can stay active simultaneously.. Source: (from training memory of book).
• lateral cortical connections (voting wires) (importance 2): Horizontal connections between columns within same region. These carry votes — Column A tells Column B what it thinks object is. Lateral connections enable within-region consensus.. Source: (from training memory of book).
• cortical feedback (predictions) (importance 2): Connections from higher cortical areas back to lower areas. Carry predictions and context. Feedback is ~10× more numerous than feedforward. Brain sends more predictions down than sensory signals up.. Source: (from training memory of book).
• cortical feedforward (sensory input) (importance 2): Connections from lower to higher cortical areas. Carry sensory features upward through hierarchy. Combined with feedback predictions, these create two-way information flow.. Source: (from training memory of book).
• Hawkins's inverse models (motor) (importance 2): Motor system learns inverse of sensory model — instead of 'what will I sense at this location', it learns 'what movements produce this sensation'. Two sides of reference-frame learning.. Source: (from training memory of book).
• population coding (distributed representation) (importance 2): Information encoded across population of neurons rather than single neuron. No grandmother cell for 'coffee cup' — pattern of active neurons represents coffee cup. Enables graceful degradation and generalization.. Source: (from training memory of book).
• Hebbian learning (synaptic plasticity) (importance 2): Neurons that fire together wire together. Synaptic connections strengthen when pre and post-synaptic neurons activate simultaneously. Foundation of all cortical learning in Thousand Brains framework.. Source: (from training memory of book).
• Hawkins's exploration (curiosity) (importance 2): Old brain generates curiosity drive — motivation to explore and build better world models. Prediction errors signal 'here's something you don't understand yet'. Learning is intrinsically rewarding.. Source: (from training memory of book).
• Hawkins's cultural transmission (importance 2): Language enables sharing reference-frame models between minds. Teaching is transmitting reference-frame structures. Culture is accumulated shared models across generations. Only possible because models are compositional and communicable.. Source: (from training memory of book).
• cortical mini-columns (~80-100 neurons) (importance 1): Finer subdivision within cortical columns. Mini-columns are vertical chains of neurons. Columns contain ~1000 mini-columns. Possibly computational subunit but less clear than full column level.. Source: (from training memory of book).
• cortical dendrites (pattern matching) (importance 1): Branching input structures on neurons. Each dendrite segment can independently recognize patterns. HTM models leverage this: neuron fires when ANY dendritic segment matches its pattern.. Source: (from training memory of book).
• proximal dendrites (feedforward) (importance 1): Dendrites receiving feedforward sensory input. Pattern matching on proximal dendrites determines if neuron will activate in response to current input.. Source: (from training memory of book).
• distal dendrites (context/prediction) (importance 1): Dendrites receiving lateral and feedback connections. Pattern matching on distal dendrites puts neuron in predictive state — ready to fire if expected feedforward input arrives.. Source: (from training memory of book).
• apical dendrites (long-range) (importance 1): Dendrites extending to cortical layer 1, receiving long-range feedback. May carry high-level context and predictions from distant regions. Anatomically distinct from basal dendrites.. Source: (from training memory of book).
• basal dendrites (local context) (importance 1): Dendrites receiving local lateral connections. Carry votes from nearby columns. Pattern recognition on basal dendrites enables local consensus.. Source: (from training memory of book).

Claims

• Hawkins's intelligence = model-building (importance 5): Intelligence is the ability to build models of the world using reference frames, make predictions from those models, and use predictions to plan behavior. Not reasoning, not language — model-building is the essence.. Source: (from training memory of book).
• Hawkins's common cortical algorithm (importance 5): Central thesis: one learning algorithm runs everywhere in neocortex. Vision, hearing, touch, language, reasoning — all use the same computational principles. Differences are input type and connectivity, not algorithm.. Source: (from training memory of book).
• Mountcastle's cortical uniformity (1978) (importance 4): Vernon Mountcastle observed that neocortex looks remarkably similar across all regions and species. Same six-layer structure, same cell types, same connectivity patterns. Hawkins extends this: if structure is uniform, algorithm must be too.. Source: (from training memory of book).
• Hawkins's abstract concepts (still reference frames) (importance 4): Democracy, justice, mathematics — even these use reference frames. A political concept is a location in 'political space'. Mathematical proofs are movements through 'theorem space'. The same grid-cell machinery applies to everything.. Source: (from training memory of book).
• Hawkins's brain-inspired AI path (importance 4): To build AGI, we need: reference frames for all knowledge, thousands of parallel models voting, continuous learning without catastrophic forgetting, sensory-motor integration. Thousand Brains provides the architecture blueprint.. Source: (from training memory of book).
• Hawkins's existential risk (goal misalignment) (importance 4): Advanced AI without old-brain-equivalent (goals, values, survival instincts) is dangerous. Intelligence without intrinsic goals follows whatever objective we specify, which we'll get wrong. Need to embed values at architectural level.. Source: (from training memory of book).
• Hawkins's columnar independence (parallel learning) (importance 4): Each column learns independently from its sensory stream. No central coordinator tells columns what to learn. This massive parallelism is why brains learn so efficiently — 150,000 simultaneous learners.. Source: (from training memory of book).
• Hawkins's AI critique (not brain-like) (importance 3): Modern AI (deep learning, transformers) doesn't work like brains. No reference frames, no voting, no continuous learning with same weights. Impressive but fundamentally different architecture. Won't achieve general intelligence without brain principles.. Source: (from training memory of book).
• Hawkins's neocortex advantage (modeling) (importance 3): Old brain can learn but only through slow reward-based trial-and-error. Neocortex can build detailed world models and simulate outcomes mentally. This is the evolutionary advantage — planning beats trial-and-error.. Source: (from training memory of book).
• Hawkins's language (reference frames) (importance 3): Language understanding uses same reference-frame machinery. Words are locations in semantic space. Grammar is movement rules through language-reference-frames. Speaking is navigating through sentence-space.. Source: (from training memory of book).
• Hawkins's embodiment requirement (importance 3): You can't build reference frames without movement through world. Embodiment isn't optional — it's how reference frames are learned. Disembodied AI that only reads text can't build true world models.. Source: (from training memory of book).
• Hawkins's continuous learning (no forgetting) (importance 3): Brains learn new things continuously without erasing old knowledge. Thousand Brains explains: new learning is new reference-frame models in new columnar subsets. Old models remain intact. Voting integrates old and new.. Source: (from training memory of book).
• Hawkins's value alignment (architectural) (importance 3): Safe AI requires building values into architecture, not training objective. Like old brain provides intrinsic goals for neocortex, we need to embed human values as axioms in AI's goal system. Can't emerge from reward function.. Source: (from training memory of book).
• Hawkins's ML critique (statistical vs structural) (importance 3): Deep learning finds statistical correlations in data. Brains build structural models of how world works. Correlation enables narrow tasks; structure enables understanding and generalization. Current AI is all correlation.. Source: (from training memory of book).
• Hawkins's sample efficiency (humans vs AI) (importance 3): Humans learn from tiny data. Child learns 'cup' from 3-5 examples. Deep learning needs millions. Thousand Brains explains gap: we're building structural models via sensory-motor exploration, not fitting statistical patterns.. Source: (from training memory of book).
• Hawkins's reasoning (navigation in concept-space) (importance 3): Logical reasoning is navigating through reference frames of abstract concepts. Deduction is following valid movement paths. Analogical reasoning is recognizing similar reference-frame structures in different domains.. Source: (from training memory of book).
• Hawkins's future (brain-inspired AGI) (importance 3): True AGI will require: reference frames, thousands of parallel models, continuous learning, sensory-motor grounding, intrinsic goals. We now have architecture blueprint. Implementation is engineering challenge.. Source: (from training memory of book).
• Hawkins's consciousness speculation (importance 2): Consciousness might emerge from voting process — the 'feeling' of being aware is columns reaching consensus. Highly speculative; Hawkins admits we don't know. But voting creates unified perception from distributed models.. Source: (from training memory of book).
• Hawkins's robotics (better than DL) (importance 2): Robots using Thousand Brains principles would learn objects through exploration, build reference-frame models, recognize from any angle/lighting. More robust than deep learning pattern matching. Numenta exploring this application.. Source: (from training memory of book).
• neocortex evolution (mammalian expansion) (importance 2): Neocortex appears with early mammals ~200M years ago, tiny at first. Massive expansion in primates. Humans have biggest ratio of neocortex to body. This expansion enabled complex modeling, planning, culture.. Source: (from training memory of book).
• Hawkins's social modeling (other minds) (importance 2): Understanding other people uses same reference-frame machinery. You build model of someone's mental state as location in 'mental-state space'. Theory of mind is reference-frame modeling applied to minds.. Source: (from training memory of book).
• Hawkins's transfer learning (reference reuse) (importance 2): Brains transfer knowledge naturally — learning piano helps learn guitar because both use similar motor-reference-frames. Deep learning struggles with transfer. Reference-frame compositionality makes transfer automatic.. Source: (from training memory of book).
• Hawkins's creativity (novel frame composition) (importance 2): Creativity is composing existing reference frames in new ways. Inventing wheel is recognizing rolling-reference-frame can be applied to transportation-reference-frame. Novel combinations of known structures.. Source: (from training memory of book).
• Hawkins's evolution (algorithm discovery) (importance 2): Evolution discovered one learning algorithm that works so well, it replicated 150,000 times. This algorithm is evolution's masterpiece — the computational principle that creates intelligence. Understanding it is neuroscience's goal.. Source: (from training memory of book).
• Hawkins's climate models (reference frames) (importance 1): Example application: climate modeling using Thousand Brains principles. Build reference-frame models of atmospheric dynamics, ocean currents, carbon cycles. Thousands of models voting on predictions. More robust than single model.. Source: (from training memory of book).
• Hawkins's medical AI (diagnostic voting) (importance 1): Medical diagnosis could use voting models — thousands of diagnostic models each looking at patient from different angle. Combine radiology, lab results, symptoms via voting. More robust than single deep learning classifier.. Source: (from training memory of book).

Empirical results

• Hawkins's coffee cup (thousands of models) (importance 3): When you see/touch a coffee cup, you don't have one representation. Thousands of cortical columns each build a model from their sensory perspective. Visual columns model what it looks like; touch columns model what it feels like. All vote to reach 'coffee cup'.. Source: (from training memory of book).
• visual cortex → auditory repurposing (importance 3): Experiments where ferret visual cortex is rewired to receive auditory input. The 'visual' cortex learns to process sound. Supports Mountcastle: same algorithm works on any sensory stream.. Source: (from training memory of book).
• Hawkins's object recognition (partial input) (importance 3): Humans recognize objects from tiny fragments — one finger touch, glimpse of edge. Thousand Brains explains how: each column has complete model; even partial input activates model; voting confirms recognition despite sparse data.. Source: (from training memory of book).
• ~150,000 cortical columns (human) (importance 3): Estimated number of cortical columns in human neocortex. Each column ~100,000 neurons. Total ~15 billion cortical neurons. These 150,000 parallel learners create human intelligence.. Source: (from training memory of book).
• cortical self-organization (unsupervised) (importance 2): Infant cortex isn't pre-wired with vision region, language region, etc. Regions self-organize based on input statistics. Whatever sensory stream arrives, the local columns learn its structure. Further evidence for algorithm uniformity.. Source: (from training memory of book).
• adult cortical plasticity (continued learning) (importance 2): Cortex continues learning throughout life. Synapses keep adapting based on experience. Thousand Brains explains how: each column independently updates its model. No centralized weights that must be frozen.. Source: (from training memory of book).
• cortical sheet (2-4mm thin) (importance 1): Neocortex is remarkably thin sheet, 2-4mm thick across entire surface. Folded to fit in skull. All intelligence emerges from this thin sheet running one algorithm. Thickness is surprisingly uniform across regions.. Source: (from training memory of book).

Methods

• Hawkins's cortical voting (consensus) (importance 5): The mechanism by which thousands of columns reach agreement. Each column proposes what it thinks the object/concept is; columns vote through inter-column connections. Consensus emerges from distributed models rather than hierarchical arbitration.. Source: (from training memory of book).
• Hawkins's HTM sequence memory (importance 4): Hierarchical Temporal Memory algorithm developed at Numenta. Models how cortical columns learn sequences and make predictions. Each column predicts what comes next; errors drive learning. Foundation of Thousand Brains implementation.. Source: (from training memory of book).
• Hawkins's location layer (grid cells) (importance 4): Each cortical column has layer with grid-cell-like neurons encoding 'where' in reference frame. As sensors move across object, location layer tracks position. Combined with sensory features, this builds structural model.. Source: (from training memory of book).
• Hawkins's within-region voting (importance 3): Columns within one cortical region (e.g., V1) vote amongst themselves first, reaching local consensus. Then regions vote with each other up the hierarchy. Distributed consensus at multiple scales.. Source: (from training memory of book).
• Hawkins's SDRs (semantic similarity) (importance 3): Sparse Distributed Representations — patterns where ~2% of neurons are active. HTM uses SDRs because overlap between patterns encodes semantic similarity. Two SDRs with 40% overlap represent similar concepts.. Source: (from training memory of book).
• Hawkins's prediction-driven learning (importance 3): Columns constantly predict next input. When prediction fails, learning happens — synapses strengthen to reduce future errors. Prediction is not just output; it's the core learning signal.. Source: (from training memory of book).
• Hawkins's feature layer (sensory input) (importance 3): Cortical layer receiving sensory features — edges, textures, sounds. Feature layer feeds into location layer to answer: 'what feature is at this location?' Joint learning builds object models.. Source: (from training memory of book).
• Hawkins's top-down prediction flow (importance 3): Higher cortical regions send predictions down to lower regions — 'I think I'm seeing a face, so I predict eyes here, nose here'. Bottom-up sensory input confirms or rejects. Perception is prediction meeting sensation.. Source: (from training memory of book).
• HTM neuron model (sequence memory) (importance 2): Simplified neuron model used in HTM: neuron activates when specific patterns occur in specific sequences. Learns temporal structure. Biologically simplified but captures key cortical computation.. Source: (from training memory of book).

Entities

• Moser & Moser grid cells (2005) (importance 4): Nobel Prize-winning discovery of neurons in entorhinal cortex that fire in hexagonal grid patterns as animals navigate. Hawkins extends this: if navigation uses grid cells, why not everything else?. Source: (from training memory of book).
• O'Keefe place cells (1971) (importance 3): Neurons in hippocampus that fire when animal is in specific location. Combined with grid cells, these form the biological GPS system Hawkins generalizes to all cortical learning.. Source: (from training memory of book).
• Hawkins's On Intelligence (2004) (importance 3): Hawkins's first book proposing memory-prediction framework. Argued intelligence is prediction; learning is storing sequences. Thousand Brains builds on and revises these ideas with reference frames and voting.. Source: (from training memory of book).
• Numenta (Hawkins's research company) (importance 3): Company founded by Hawkins in 2005 to understand neocortical learning and build intelligent machines based on brain principles. Research led to HTM algorithm and papers on grid cells in neocortex.. Source: (from training memory of book).
• thalamus (voting relay) (importance 2): Structure that relays information between cortical regions. Hawkins suggests thalamus may play role in cortical voting — carrying votes between regions and helping orchestrate consensus.. Source: (from training memory of book).

Relations

• Hawkins's Thousand Brains Theory requires Hawkins's cortical column (mini-brain)
• Hawkins's Thousand Brains Theory requires cortical reference frames (grid cells for everything)
• Hawkins's Thousand Brains Theory requires Hawkins's cortical voting (consensus)
• Hawkins's cortical column (mini-brain) builds-on Mountcastle's cortical uniformity (1978)
• cortical reference frames (grid cells for everything) builds-on Moser & Moser grid cells (2005)
• cortical reference frames (grid cells for everything) builds-on O'Keefe place cells (1971)
• Hawkins's cortical voting (consensus) requires lateral cortical connections (voting wires)
• Hawkins's cortical voting (consensus) requires thalamus (voting relay)
• Moser & Moser grid cells (2005) builds-on O'Keefe place cells (1971)
• Numenta displacement cells (object locations) builds-on Moser & Moser grid cells (2005)
• Numenta displacement cells (object locations) exemplifies cortical reference frames (grid cells for everything)
• Hawkins's HTM sequence memory exemplifies Hawkins's cortical column (mini-brain)
• Hawkins's HTM sequence memory cites Numenta (Hawkins's research company)
• Hawkins's complementary learning (old + new) requires Hawkins's old brain (survival circuits)
• Hawkins's complementary learning (old + new) requires Hawkins's new brain (neocortex)
• Hawkins's Thousand Brains Theory builds-on Hawkins's On Intelligence (2004)
• Hawkins's common cortical algorithm builds-on Mountcastle's cortical uniformity (1978)
• Hawkins's coffee cup (thousands of models) exemplifies Hawkins's Thousand Brains Theory
• Hawkins's sensory-motor learning enables cortical reference frames (grid cells for everything)
• Hawkins's sensory-motor learning requires Hawkins's location layer (grid cells)
• Hawkins's within-region voting exemplifies Hawkins's cortical voting (consensus)
• Hawkins's abstract concepts (still reference frames) supports cortical reference frames (grid cells for everything)
• Hawkins's SDRs (semantic similarity) requires Hawkins's HTM sequence memory
• Hawkins's prediction-driven learning enables Hawkins's HTM sequence memory
• Hawkins's nested reference frames exemplifies cortical reference frames (grid cells for everything)
• Hawkins's intelligence = model-building requires cortical reference frames (grid cells for everything)
• visual cortex → auditory repurposing evidences Hawkins's common cortical algorithm
• Hawkins's AI critique (not brain-like) supports cortical reference frames (grid cells for everything)
• Hawkins's brain-inspired AI path requires Hawkins's Thousand Brains Theory
• Hawkins's existential risk (goal misalignment) requires Hawkins's old brain (survival circuits)
• Hawkins's knowledge genome requires cortical reference frames (grid cells for everything)
• catastrophic forgetting (absent in cortex) contradicts Hawkins's continuous learning (no forgetting)
• Hawkins's location layer (grid cells) builds-on Moser & Moser grid cells (2005)
• Hawkins's feature layer (sensory input) requires Hawkins's location layer (grid cells)
• Hawkins's object recognition (partial input) evidences Hawkins's cortical voting (consensus)
• Hawkins's generalization (reference frame transfer) requires cortical reference frames (grid cells for everything)
• Hawkins's columnar independence (parallel learning) supports Hawkins's Thousand Brains Theory
• Hawkins's attention (old brain selection) requires Hawkins's old brain (survival circuits)
• cortical self-organization (unsupervised) evidences Hawkins's common cortical algorithm
• Hawkins's neocortex advantage (modeling) supports Hawkins's new brain (neocortex)
• Hawkins's symbol grounding (reference frames) requires cortical reference frames (grid cells for everything)
• Hawkins's consciousness speculation requires Hawkins's cortical voting (consensus)
• Hawkins's universal mapping principle generalizes cortical reference frames (grid cells for everything)
• Hawkins's language (reference frames) requires cortical reference frames (grid cells for everything)
• Hawkins's cross-modal voting requires Hawkins's cortical voting (consensus)
• Hawkins's embodiment requirement requires Hawkins's sensory-motor learning
• Hawkins's top-down prediction flow requires cortical feedback (predictions)
• Hawkins's top-down prediction flow enables Hawkins's prediction-driven learning
• HTM neuron model (sequence memory) requires cortical dendrites (pattern matching)
• proximal dendrites (feedforward) requires cortical feedforward (sensory input)
• distal dendrites (context/prediction) requires lateral cortical connections (voting wires)
• population coding (distributed representation) exemplifies Hawkins's SDRs (semantic similarity)
• Hebbian learning (synaptic plasticity) enables Hawkins's prediction-driven learning
• Hawkins's continuous learning (no forgetting) enables Hawkins's cortical voting (consensus)
• Hawkins's compositional models enables Hawkins's nested reference frames
• Hawkins's comprehensive world model requires cortical reference frames (grid cells for everything)
• Hawkins's mental simulation (prediction) requires cortical reference frames (grid cells for everything)
• Hawkins's value alignment (architectural) builds-on Hawkins's old brain (survival circuits)
• Hawkins's exploration (curiosity) requires Hawkins's old brain (survival circuits)
• Hawkins's social modeling (other minds) requires cortical reference frames (grid cells for everything)
• Hawkins's cultural transmission requires Hawkins's language (reference frames)
• Hawkins's ML critique (statistical vs structural) supports Hawkins's AI critique (not brain-like)
• Hawkins's transfer learning (reference reuse) requires Hawkins's compositional models
• Hawkins's sample efficiency (humans vs AI) requires Hawkins's sensory-motor learning
• Hawkins's reasoning (navigation in concept-space) requires cortical reference frames (grid cells for everything)
• Hawkins's creativity (novel frame composition) requires Hawkins's compositional models
• Hawkins's future (brain-inspired AGI) supports Hawkins's brain-inspired AI path
• Hawkins's evolution (algorithm discovery) supports Hawkins's common cortical algorithm
• Hawkins's new brain (neocortex) requires Hawkins's cortical column (mini-brain)
• Hawkins's inverse models (motor) requires cortical reference frames (grid cells for everything)
• cortical layer 6 (feedback) requires cortical feedback (predictions)
• cortical working memory (sustained activation) requires Hawkins's cortical column (mini-brain)
• cortical long-term memory (synaptic weights) requires Hebbian learning (synaptic plasticity)
• ~150,000 cortical columns (human) evidences Hawkins's cortical column (mini-brain)
• Hawkins's cortical voting (consensus) evidences visual cortex → auditory repurposing
• Hawkins's object recognition (partial input) exemplifies Hawkins's coffee cup (thousands of models)
• Hawkins's robotics (better than DL) exemplifies Hawkins's brain-inspired AI path
• Numenta (Hawkins's research company) cites Numenta displacement cells (object locations)
• adult cortical plasticity (continued learning) evidences Hawkins's continuous learning (no forgetting)
• Hawkins's temporal sequence learning requires Hawkins's HTM sequence memory
• cortical mini-columns (~80-100 neurons) requires Hawkins's cortical column (mini-brain)

© 2026 LOOMUS. Curation, importance scoring, edge typing, and visualization. Source paper © respective authors, quoted under fair use.