These are Jeff Hawkins' slides from the Computational Theories of the Brain Workshop held at the Simons Institute at UC Berkeley on April 17, 2018.
Abstract:
In this talk, I propose that the neocortex learns models of objects using the same methods that the entorhinal cortex uses to map environments. I propose that each cortical column contains cells that are equivalent to grid cells. These cells represent the location of sensor patches relative to objects in the world. As we move our sensors, the location of the sensor is paired with sensory input to learn the structure of objects. I explore the evidence for this hypothesis, propose specific cellular mechanisms that the hypothesis requires, and suggest how the hypothesis could be tested.
References:
“A Theory of How Columns in the Neocortex Enable Learning the Structure of the World” by Jeff Hawkins, Subutai Ahmad, YuWei Cui (2017)
“Place Cells, Grid Cells, and the Brain’s Spatial Representation System” by Edvard Moser, Emilio Kropff, May-Britt Moser (2008)
“Evidence for grid cells in a human memory network” by Christian Doeller, Caswell Barry, Neil Burgess (2010)
Artificial Intelligence In Microbiology by Dr. Prince C P
Does the neocortex use grid cell-like mechanisms to learn the structure of objects? by Jeff Hawkins (04/17/18)
1. Computational Theories of the Brain
Simons Institute
April 17, 2018
Jeff Hawkins
jhawkins@numenta.com
Does the Neocortex Use Grid Cell-Like Mechanisms to Learn
the Structure of the World?
A Framework for Cortical Computation
2. 1) Reverse engineer the neocortex
- an ambitious but realizable goal
- seek biologically accurate theories
- test empirically and via simulation
2) Enable technology based on cortical theory
- active open source community
- foundation for machine intelligence
3.
4. “What is conspicuously lacking is a broad framework of
ideas within which to interpret these different approaches”
Francis Crick, 1979
writing about the state of neuroscience
5.
6. Diversity of Function Commonality of Circuitry
Regions look similar
- layers of cells
- vertical connections (primary)
- horizontal connections
- suggests columnar organization
~100 regions
- vision
- touch
- audition
- languages
- all cognition
Remarkably similar circuitry
Vernon Mountcastle’s Big Idea:
1) All regions do the same thing, the “function” of a region is determined by its inputs.
2) Columns are the functional unit of the neocortex. (~ 150K in human)
3) Understanding the column is the key problem in neuroscience. V. Mountcastle, 1978
Hierarchical
Remarkably diverse functionality
7. L2
L3a
L3b
L4
L6a
L6b
L6 ip
L6 mp
L6 bp
L5 tt
L5 cc
Cortical Columns are Incredibly Complex
L6: Zhang and Deschenes, 1997
50%
L5 CTC: Guillery, 1995
Constantinople and Bruno, 2013
<10%
Cortex
Thalamus
Cortex
Motor
Thalamus
- 100K neurons, 500M synapses (1mm2)
- Ten or more cellular layers
- Dozens of intra- and inter-column connections
- Inhibitory neurons/circuits are equally complex
- Significant region-to-region variability
Observation:
The cortex is constantly predicting its input.
Question:
How does the cortex (column) learn predictive
models of its input?
8. Deciphering the Cortical Column One Layer at a Time
Learn predictive
models of sequences
+Learn predictive
models of sensorimotor
sequences
+ Grid cell-like
location layer
+Learn composite objects
+2nd Location layer
Hawkins and Ahmad,
Frontiers in Neur Circ
2016/03/30
4 other papers
Hawkins, Ahmad and Cui
Frontiers in Neur Circ
2017/10/25
Lewis and Hawkins
Poster: Cosyne 2018
Lewis and Hawkins
Poster: Cosyne 2018
L2/3
L4
L5tt
L6a
L6b
L2/3
L4
L6a
L2/3
L4L4
EC-derived location
9. Proximal synapses: Define classic receptive field of neuron
Distal synapses: Cause dendritic spikes
Put the cell into a depolarized, or “predictive” state
Hypothesis:
Depolarized neurons fire sooner, inhibiting nearby neurons.
A neuron can predict its activity in hundreds of learned contexts.
5K to 30K excitatory synapses
- 10% proximal
- 90% distal
Distal dendrites are pattern detectors
- 8-15 co-active, co-located synapses
generate dendritic spike
- sustained depolarization of soma
HTM Neuron
Prediction Starts in the Neuron
Pyramidal Neuron
(Major, Larkum and Schiller 2013)
10. Properties of Sparse Activations
Example: One layer of cells
5,000 neurons, 2% (100) active
Hawkins, Ahmad, 2016
Ahmad, Hawkins, 2015
1 pattern (100 active cells)
Union
10 patterns (1,000 active cells)
4) Unions of patterns do not cause errors in recognition.
1) Representational capacity is virtually unlimited.
(5,000 choose 100) = 3x10211
2) Randomly chosen representations have minimal overlap.
3) A neuron can robustly recognize an active pattern
by forming just 8 to 20 synapses.
Hypothesis:
Unions are used to represent uncertainty throughout the cortex.
Activity gets sparser as certainty increases.
11. A Input Layer Network Model for Sequence Memory
No prediction Predicted input
(Hawkins & Ahmad, 2016)
(Cui et al, 2016)
- High capacity (learns 100’s K transitions)
- Learns high-order sequences: “ABCD” vs “XBCY”
- Extremely robust (parameters, noise, and faults)
- Learning is unsupervised, continuous, and local
- Satisfies many biological constraints
- Multiple open source and commercial implementations
t=0
t=1
Sparse pattern =
Input in specific context
Next prediction t=2
t=0
t=1
12. Deciphering the Cortical Column One Layer at a Time
Learn predictive
models of sequences
+Learn predictive
models of sensorimotor
sequences
+ Grid cell-like
location layer
+Learn composite objects
+2nd Location layer
Hawkins and Ahmad,
Frontiers in Neur Circ
2016/03/30
4 other papers
Hawkins, Ahmad and Cui
Frontiers in Neur Circ
2017/10/25
Lewis and Hawkins
Poster: Cosyne 2018
Lewis and Hawkins
Poster: Cosyne 2018
L2/3
L4
L5tt
L6a
L6b
L2/3
L4
L6a
L2/3
L4L4
EC-derived location
13. Predicting Sensorimotor Sequences
SensorMotor-related context
How can we modify our input layer to also learn predictive
models of sensorimotor sequences?
Add a motor-related context. The layer can now predict its
input as the sensor moves.
What is the correct motor-related context?
14.
15. Predicting Sensorimotor Sequences
- Input layer represents “features @ locations”.
- Changes with each movement.
Sensed
Feature
Object-centric
Location
- “Object” layer represents object.
- Stable over changing inputs.
This network learns predictive models of objects.
An object is “a set of features @ locations”.
19. Deciphering the Cortical Column One Layer at a Time
Learn predictive
models of sequences
+Learn predictive
models of sensorimotor
sequences
+ Grid cell-like
location layer
+Learn composite objects
+2nd Location layer
Hawkins and Ahmad,
Frontiers in Neur Circ
2016/03/30
4 other papers
Hawkins, Ahmad and Cui
Frontiers in Neur Circ
2017/10/25
Lewis and Hawkins
Poster: Cosyne 2018
Lewis and Hawkins
Poster: Cosyne 2018
L2/3
L4
L5tt
L6a
L6b
L2/3
L4
L6a
L2/3
L4L4
EC-derived location
20. Entorhinal Cortex
Body in environments
A
B C
X
Y Z
R
S T
Room 3
Room 2Room 1
Location
- Encoded by grid cells
- Unique to location in room AND room
- Location is updated by movement
A Room is:
- A set of locations that are connected by
movement (via path integration).
- Some locations have associated features.
Location
- Encoded by grid-like cells in L6a
- Unique to location on object AND object
- Location is updated by movement
Cortical Column
Sensor patch relative to objects
Representing Location with Grid Cells
A
C
B
Stensola, Solstad, Frøland, Moser, Moser: 2012
X
Y
Z
W
An Object is:
- A set of locations that are connected by
movement (via path integration).
- Some locations have associated features.
1) Location representations are dimensionless. Dimensionality is defined by movement.
2) Movements do not have to be physical. They only have to exhibit path integration.
3) Features do not have to be sensory features. They can be outputs of other columns.
Proposal: All knowledge, even abstract concepts, are represented this way in the cortex.
Conceptual Spaces
21. Deciphering the Cortical Column One Layer at a Time
Learn predictive
models of sequences
+Learn predictive
models of sensorimotor
sequences
+ Grid cell-like
location layer
+Learn composite objects
+2nd Location layer
Hawkins and Ahmad,
Frontiers in Neur Circ
2016/03/30
4 other papers
Hawkins, Ahmad and Cui
Frontiers in Neur Circ
2017/10/25
Lewis and Hawkins
Poster: Cosyne 2018
Lewis and Hawkins
Poster: Cosyne 2018
L2/3
L4
L5tt
L6a
L6b
L2/3
L4
L6a
L2/3
L4L4
EC-derived location
22. Rethinking Hierarchy
Every column learns complete models of objects. They operate in parallel.
Inputs project to multiple levels at once. Columns operate at different
scales of input.
Sense
Simple features
Complex features
Objects
Classic
Objects
Objects
Objects
Sensor array
Proposed
Region 3
Region 2
Region 1
23. Rethinking Hierarchy
Every column learns complete models of objects. They operate in parallel.
Inputs project to multiple levels at once. Columns operate at different
scales of input.
Non-hierarchical connections allow columns to vote on shared elements
such as “object” and “composite object”.
Sense
Simple features
Complex features
Objects
Classic
Sensor array
Objects
Objects
Objects
Sensor array
vision touch
Proposed
Region 3
Region 2
Region 1
24. 1) Border ownership cells:
Cells fire only if feature is present at object-centric location on object.
Detected even in primary sensory areas (V1 and V2).
(Zhou et al., 2000; Willford & von der Heydt, 2015)
2) Grid cell signatures in cortex:
Cortical areas in humans show grid cell like signatures (fMRI and single cell recordings)
Seen while subjects navigate conceptual object spaces and virtual environments.
(Doeller et al., 2010; Jacobs et al. 2013; Constantinescu et al., 2016; )
3) Sensorimotor prediction in sensory regions:
Cells predict their activity before a saccade.
Predictions during saccades are important for invariant object recognition.
(Duhamel et al., 1992; Nakamura and Colby, 2002; Li and DiCarlo, 2008)
4) Hippocampal functionality may have been conserved in neocortex:
Six-layer neocortex evolved by stacking 3-layer hippocampus and piriform cortex
(Jarvis et al., 2005; Luzatti, 2015)
Biological Evidence
24