Neocortex

Primary Cortex

Squire et al. 1989

Association Cortex
Hippocampus

CA1

CA3

Doughnuts in the brain: periodic boundary conditions on the brains spatial coordinate system
neocortex

Dentate Gyrus

Brainbow micrograph by Dr. T. Weissman

Neocortex

Bruce L. McNaughton Ph.D. AHFMR Polaris research Chair, Dept. Neuroscience The University of Lethbridge Lethbridge, Alberta, Canada Visiting Professor KUL, NERF bruce.mcnaughton@uleth.ca

CA1

CA3

Dentate Gyrus

Brainbow micrograph by Dr. T. Weissman

Neocortex

CA1

CA3

Dentate Gyrus

Carol Barnes Bill Skaggs Alexei Samsonovich Matt Wilson Jim Knierim Kati Gothard Min Jung David Redish Alex Terrazas Francesco Battaglia Drew Maurer Zaneta Navratilova May-Britt Moser Edvard Moser Jill Leutgeb Stefan Leutgeb Francesca Sargolini Ole Jensen Laura Colgin

Brainbow micrograph by Dr. T. Weissman

iPad 0.5 GB DDR2 RAM Processor speed 1 GHz Power consumption 1.5W Volume ~ .3 cc Weight ~0.7g

Rat brain 150,000,000 neurons 15,000,000,000 synapses ~1.9 GB RAM Processor speed ~1.5-15 G logic operations per second Power consumption ~0.05W Volume 2 cc Weight ~ 2 g

5 x 104 neurons 3 x 108 synapses

per mm3
Vias ~ 100 nm Fully 3-D interconnects

Gate pitch ~ 100 nm Connectivity very low compared to brain, and mostly limited to 2-D

through-wafer direct die-to-die copper area-array interconnections

Hippocampus is essential for new memory formation

The Hippocampus receives highly processed inputs from a variety of cortical and subcortical areas. Hippocampal outputs are widespread throughout the cortex and subcortical regions.
Figure adapted from Pinel, 2000

Hippocampus is part of cortex

The cortex is arranged as a hierarchical set of vertically and horizontally interacting modules. The Hippocampus is the TOP of the hierarchy
hippocampus
Felleman & Van Essen '91

hippocampus

CONNECTIONS OF THE VISUAL SYSTEM

retina

Squire et al., 1989

 In 1957, an epileptic patient (H.M.) underwent a bilateral temporal lobectomy (Scoville & Milner, 1957).

This procedure had an immediate and devastating impact on H.M.s memory capabilities, although his cognitive capabilities were relatively intact (e.g. IQ).

Hippocampus is essential for new memory formation

Hippocampal dysfunction leaves old knowledge relatively intact, but disrupts recent memory and new
learning:

why?

Time
Consolidated memory Not yet consolidated memory

Retrograde amnesia

Anterograde amnesia

Time of hippocampal damage

Present

Recall efficiency as a function of the time (before or after hippocampal damage) when the item was experienced

WHAT IS MEMORY? Associative memory is the ability to retrieve the whole of an experience from one or more of its parts Seeing the glass fall is typically followed immediately by a

WHAT IS MEMORY? Associative memory is the ability to retrieve the whole of an experience from one or more of its parts Seeing the glass fall is typically followed immediately by a

CRASH

WHAT IS MEMORY? Associative memory is the ability to retrieve the whole of an experience from one or more of its parts
Hearing the crash Seeing a glass fall

Associative synaptic links form

During learning, the near simultaneous occurrence of two sensory inputs results in the formation of associative synaptic links among the brain cells that represent the events

WHAT IS MEMORY? Associative memory is the ability to retrieve the whole of an experience from one or more of its parts
Remembering the crash Seeing a glass fall

Activation spreads via associative synaptic links

During recall, the occurrence of part of a stored experience causes retrieval of the rest of the experience via associative synaptic links that were formed during learning. We call this pattern completion

WHAT IS MEMORY? Associative memory is the ability to retrieve the whole of an experience from one or more of its parts
Remembering the crash Seeing a glass fall

Activation spreads via associative synaptic links

During recall, the occurrence of part of a stored experience causes retrieval of the rest of the experience via associative synaptic links that were formed during learning. We call this pattern completion

WHAT IS MEMORY? Associative memory is the ability to retrieve the whole of an experience from one or more of its parts
Remembering the crash Seeing a glass fall

Activation spreads via associative synaptic links

During recall, the occurrence of part of a stored experience causes retrieval of the rest of the experience via associative synaptic links that were formed during learning. We call this pattern completion

WHAT IS MEMORY? Associative memory is the ability to retrieve the whole of an experience from one or more of its parts
Hearing the crash Remembering a glass fall

Recall can be bidirectional

Activation spreads via associative synaptic links

First some basics: Brain cells (neurons) and how they communicate with each other.

NEURONS and SYNAPSES

First some basics: Brain cells (neurons) and how they communicate with each other. dendrite synapse

Wij

Neuron j
cell body (soma)

axon

S
The human cerebral cortex contains more than 10 billion neurons Neuron i

Each cell sends and recieves about 10,000

synapses
Wij

Wij

Usually about 300 400 active synapses are needed to excite a cortical neuron

Classification based on neurotransmitter

Classification based on morphology

Guerra et al 2011

Figure 3. Profiles of gene expression and electrical behaviour.

Classification based on gene expression and biophysics

Toledo-Rodriguez M et al. Cereb. Cortex 2004;14:1310-1327

2004 by Oxford University Press

Classification based on connectivity

Thalamo-cortical and cortico-thalamic circuit

http://www.med.yale.edu/neurobio/mccormick/mccormicknew/Index.html

Cerebellar cortex basic circuit

Synaptic circuits of the striatum

Annu. Rev. Neurosci. 2011. 34:44166

Gerfen & Surmeier Annu. Rev. Neurosci. 2011. 34:44166

Symbolic neurons and neural networks

Cells that have fired together form associative groups (assemblies) linked by mutually strengthened synapses, such that reactivation of some members of the group leads to complete retrieval of the active group

D.O. Hebb (1949) provided the key concept underlying the modern theory of neural associative memory

cells that fire together wire together'

For this to work, there must be at least a weak connection between the relevant brain cells to begin with.

David Marr A Theory for Cerebral Neocortex, 1970 Simple Memory: A Theory for Archicortex, 1971

Networks with modifiable recurrent connections could perform autoassociative memory (Marr,1971). The mutual connections among neurons participating in a given cell assembly enable the retrieval of a complete pattern from any pattern that is a unique fragment of the original (pattern completion) or that resembles the original closely enough (error correction). [from McNaughton & Morris TINS, 1986]

Hopfield Net

``Neural networks and physical systems with emergent collective computational abilities'', Proc. Natl. Acad. Sci. USA 79, 2554 (1982)

neural energy landscape model of pattern recognition

J.J. Hopfield

Content Addressable Memory (aka autoassociative memory) The current standard models assume complete (all-to-all) connectivity

The problem: it won't work. Do the math! Cortical connectivity is much too sparse.
Hypothetical Connection Map of Cortex The typical cortical
1010

The synaptic weight matrix Wij

1 1 1010

neuron (j) recieves about 10,000 synapses from other neurons (i); but there are about 10,000,000,000 neurons.

On average, each cell recieves input from only one in a million other cells.

Indirect association in a modular, hierarchical network

Modular (small-world), hierarchical, reciprocal structure may provide a solution to the sparse connectivity problem

Primary Cortex

Hippocampus

Association Cortex
Hippocampus

Association Cortex Primary Cortex

Squire et al., 1989

Indirect Association:
During the initial encoding of the memory, output of the top module may become associated with the patterns in each lower module. This provides an index code for each memory. Recreating the index pattern evokes the corresponding patterns in the lower levels, thus completing the retrieval of the whole memory

Compound event encoded over different modules

Retrieval cued by external event

Spontaneous retrieval

Why does damage to the hippocampus impair new learning? Because there is no index module to enable indirect association

The cortex is highly modular, meaning that there is are groups of cells that are more densly interconnected with each other than with cells in other modules.
Modules contain a few thousand neurons.

Top-down projections terminate in NMDAR rich superficial layer of neocortex

CA1 CA3

Subiculum

Entorhinal Cortex

Dentate Gyrus

Listening to a single hippocampal neuron.

Neurons in the rat hippocampus are very selective. The traditional question had been: "what do they encode?".

Firing Rate Map

Listening to a single hippocampal neuron.

Neurons in the rat hippocampus are very selective. The traditional question had been: "what do they encode?".

Firing Rate Map

Firing rate maps for 80 simultaneously recorded hippocampal neurons while rat ran in a square arena. (High rate cells are interneurons) (Wilson & McNaughton, 1993)

How the hippocampus creates top-down spatiotemporal links for neocortical memory

Firing Rate Map

Firing rate maps for 80 simultaneously recorded hippocampal neurons while rat ran in a square arena. (High rate cells are interneurons) (Wilson & McNaughton, 1993)

How is the spatial code updated as the animal changes its location? What sets the scale?

Firing Rate Map

Firing rate maps for 80 simultaneously recorded hippocampal neurons while rat ran in a square arena. (High rate cells are interneurons) (Wilson & McNaughton, 1993)

Is the code 'purely' spatial? What kinds of information are actually stored in hippocampus?

Multi neuron recording enables us to read out the contents of brains codes. For example, it allows us to compute in real-time where the rat is and where he is planning to go, just from its brain activity.

Blue: Actual path of rat in a box Red: Path predicted from firing rate maps

"Path integration" in rodents. Without vision or any other external cue, rats can make a tortuous outward journey from 'home' and then return on a direct route. To do this they must keep track of changes in direction and distance travelled. (from Mittlestaedt & Mittelstaedt, 1980)

Head direction cells located at input to hippocampal formation and within it

90 120 10 Hz 150 5 Hz 30 60

180

210

330

240 270

300

Single HD cell tuning function in polar coordinates

Ranck, 1984; Taube et al., 1990;

information about distance traversed. Velocity tuning (McNaughton et al., 1981)

There is no evidence for distance cells in the hippocampusyou cannot read distance directly from place or grid cells. But explicit distance cells are not necessary: the firing rate of most hippocampal pyramidal cells increases linearly with speed (over most of the normal range), but place field size does not seem to change appreciably with speed. The distance moved by the rat can be accurately gauged by an external observer by integrating the firing rate over time. But integration in the brain means simply updating the position on the map.

Roughly, a constant number of spikes is emitted during each lap through the place field regardless of speed. It is unknown where the speed signal comes from!

25 Speed (cm/s)

Ekstrom et al '01

information about distance traversed.

Theta power increases linearly with running speed(there is also a small change in frequency)
Raw EEG 7 Hz 150 Hz

Spikes

(2 sec)

Position

rat position

theta integral
Time

The distance moved by the rat can be accurately gauged (by an external observer) by integrating the theta power over time
Terrazas & McNaughton '03

Under conditions of restraint, hippocampal neurons (and head direction cells) become silent
Foster, Castro, McNaughton 1989. Science 244:1580-1582

The place code is updated by path integration N

W
Average

During random foraging in two dimensions, place cell firing is independent of direction.

Muller et al., J. Neurosci. 1994

The place code is maintained in darkness

Once initial position is established in light, firing locations persist in dark. (Each cluster of colored dots represents spikes from different neurons) ALSO, once initial position is established in dark, firing locations persist in light!

External cues do not specify firing fields, but can (sometimes) realign the path-integrator
(Fuhs, VanRhoades, Casale, McNaughton & Touretzky, J.
Neurophysiol. 2005, 94, 2603-2616; Skaggs & McNaughton 1995)

In the Opposite orientation condition (O), the path integrator overrides the effects of the visual landmarks; but in the Same orientation condition (S), the cues reset the path integrator.

The size of place fields is essentially independent of the rate at which external inputs change
180 cm track

CUE-RICH SECTION

CUE-POOR SECTION
Overlap of Population Vectors along Track Bin Number 1 90 80

Rich

0.9 0.8 0.7 0.6 0.5 0.4

The place code is updated by path integration

Bin Number along Track

70 60 50 40 30 20 10

Poor
10 20 30 40 50 60 70 Bin Number along Track 80 90

0.3 0.2

Decorrelation distance is independent of cue density

Battaglia et al., J. Neurosci., 2004, 24(19):4541-4550

Knierim, Kudrimoti, McNaughton. J. Neurophysiol. 1998 Place fields before and after fast or slow rotations of the apparatus with rat

The place code is updated by path integration In uniform cylinder there is random drift of the directional coordinate. Head direction cells and place cells always co-rotate as an integrated ensemble.

How does the path integrator work?

Hippocampal topology: hippocampus is like a folded sheet of extended cortex. It has two principal axes the longitudinal axis and the transverse axis. Transverse slices look very similar anywhere along the longitudinal axis, but there are important gradients.

dorsal
CA1 CA3 Dentate Gyrus Entorhinal Cortex

Subiculum

middle

ventral

Left Hemisphere

Right Hemisphere

Basic wiring diagram of hippocampal formation

1989 - attractor neural networks. 1991 continuous attractor neural networks (with M. Tsodyks) Daniel Amit
1938-2007

neural energy landscape for continuous functions

Firing rate

Theoretical distribution of firing rates in a population of parietal cortical neurons tuned to different horizontal positions of eye in the orbit
Population tuning function (aka activity packet,bump, or hill )
Left 0 Right

Neurons (conceptually) ordered according to preferred value of parameter

Continuous attractor neural network in 1-D)

Neurons (conceptually) ordered according to preferred value of parameter. Local excitatory connections. Global feedback inhibition (not shown) limits net activity. Bumps become stable states.

Continuous attractor neural networks in one and two dimensions

The effect of local excitatory connections and global inhibitory connections is a single bump of excitation

Quicktime movie

Continuous attractor neural networks in one and two dimensions

To perform angular or 2-D path integration, you have to make the bump move consistently with the rats movement.

Attractor network model for head direction cells. Key elements are a ring attractor, and a set of cells conjunctive for direction and angular velocity with offset return connections. This is the so-called hidden layer of classical neural network theory.

head-direction cells

conjunctive cells
90 120 10 Hz 150 5 Hz 30 60

180

210

330

240 270

300

(McNaughton et al., 1989; Skaggs et al., 1995)

The identical concept in 2-D (Samsonovich & McNaughton, 1997)

Head direction signal

Linear motion signal

Conjunctive cells

Intermediate layer contains cells that are conjunctive for location, head direction (and are linearly sensitive to speed)

?
There is a finite number of cells, so the map cant be infinite. What happens when the rat gets to the edge of its map?

Samsonovich and McNaughton 1996

Samsonovich & McNaughton (1997) implemented the map on a standard torus (periodic boundary condition), which predicts regularly repeating square grid of place fields.
0 20 40 60 80 100 120 140

160 180
200 0

50

100

150

200

Microstructure of a spatial map in the entorhinal cortex

Hafting, Fyhn, Molden, Moser & Moser1 Nature 2005

Place fields of layer II medial entorhinal cortical cells have a very regular triangular (rhomboidal) grid-like structure

b
0

20 40
60 80 100 120

Rhomboidal grid could arise from a simple distortion of a rectangular map with periodic boundaries: a twisted torus

140 160 180 200 0

200 0 50 100 150 200 180

50
160 140 120

100
100 80

150
60 40 20

200
0

R O L L

Sargolini, Fyhn, Hafting, McNaughton, Witter, Moser, and Moser (2006)

head direction cell

Conjunctive grid x head direction cell

In addition to grid cells in MEC, there are head direction cells and conjunctive head direction x grid x speed cells in deeper layers. These are exactly the cells required to implement path integration according to the continuous attractor model. EC is likely to be the path integrator.

What sets the spatial scale ?

Firing Rate Map

Firing rate maps for 80 simultaneously recorded hippocampal neurons while rat ran in a square arena. (High rate cells are interneurons) (Wilson & McNaughton, 1993)

One simple definition of scale is the average size of the place fields.

Firing Rate Map

Firing rate maps for 80 simultaneously recorded hippocampal neurons while rat ran in a square arena. (High rate cells are interneurons) (Wilson & McNaughton, 1993)

Another definition of scale is the rate at which the population activity changes as position changes.

PV

Cell #
1 2 3

. . .
N

Population vector for a one dimensional space

PVPV

Cell #
1 2 3

. . .
N

Nearby vectors are correlated

PV

PV

Cell #
1 2 3

. . .
N

Correlation decreases with distance

PV

PV

Cell #
1 2 3

. . .
N

For bigger place fields, correlation decreases more slowly with distance

Spatial scale can be defined as the distance over which successive population vectors become decorrelated. One dimensional case:
Overlap of Population Vectors along Track Bin Number 1 90 80 0.8

1
0.9

0.8

Bin Number along Track

70

50 40 30 20

0.6 0.5 0.4 0.3

pop. vector corr.

60

0.7

0.6

0.4

0.2

10 0.2 10 20 30 40 50 60 70 Bin Number along Track 80 90

-0.2 -50 0 distance on track (cm) 50

Population vector correlation matrix

Battaglia et al. 2004

Dorsal

Place fields get bigger, and fewer cells are active at a given location as recording location move ventrally dorsal

middle Middle

ventral

Jung, Wiener, McNaughton (1994). J. Neurosci. 14(12): 73477356.

dorsal

middle

ventral

Spatial scale increases systematically along the dorso-ventral axis of the hippocampal formation
Jung, Weiner, McNaughton 1994 Maurer, VanRhoades, Sutherland, Lipa, McNaughton, 2005

Hafting et al., 2005

Grid scale also increases systematically along the septotemporal axis. Grid orientation is constant.

"Fourier Synthesis" of Place Fields Combining grid fields at different spatial scales could lead to nonrecurring place fields such as observed downstream of the entorhinal cortex in hippocampus

(McNaughton, Battaglia, Jensen, Moser and Moser. Nature Rev. Neurosci. 2006); Solstad et al. (2006)

Head direction signal

Linear motion signal

Conjunctive cells

Reducing the gain of the motion signal would make the bump move slower grid scale would look bigger

0 20 40 60 80 100 120 140

0 20 40 60 80 100
0 50 100 150 200

160 180 200

120 140 160 180 200 0

50

100

150

Bump speed sets the spatial scale

20

Terrazas, Krause, Gothard, McNaughton, Barnes (J. Neurosci 2005)

Deletion of self-motion signals makes the hippocampal code change more slowly with position: place fields get bigger

Place field expansion after degradation of movement signal is as predicted by slower bump movement: overlap distribution preserved.

Like this

Not like this

Also, population sparsity is preserved, in-field peak rates decrease

Grid Scale Quantization A prediction of the

Dorsal MEC small scale = high speed
toroidal attractor model (but not the interference model): independent modules from dorsal to ventral MEC, so each module can have a different movement speed gain. One bump cant move simultaneously at different speeds. This prediction was recently confirmed

Ventral MEC large scale = low speed

Speed signal

tetrodes were moved tangentially along layer II of MEC

This showed discrete steps in grid spacing
Stensola, Stensola, Solstad, Frland, Moser, Moser, unpublished (slide courtesy May-Britt Moser)

Thus, grid cells are organized into modules with similar spacing and orientation, consistent with a network mechanism

Is the code 'purely' spatial? What kinds of information are actually stored in hippocampus?

Classical view of activity bump is a smooth distribution with rate falling off as a function of distance

Firing rate

0.8

0.6

0.4

0.2

-0.5

-0.4

-0.3

-0.2

-0.1

0.1

0.2

0.3

0.4

0.5

Cells ordered by location parameter relative to current rat location

Constant Place - Variable Cues

Independent Codes for Spatial and Episodic Memory in Hippocampal Neuronal Ensembles
Leutgeb, Leutgeb, Barnes, Moser, McNaughton, Moser. Science, 2005

Constant Cues - Variable Place

In Samsonovich & McNaughton attractor map model, external (cortical) inputs (V) become Hebbian associated with bump locations, to enable correction of the PI by landmarks. This implies that external inputs affect how much cells fire, but not where they fire!

Constant Place - Variable Cues - RATE REMAPPING Variations in cues affect how strongly cells fire, but not which cells can fire. The potentially active population is determined by the path integrator. Independent Codes for Spatial and Episodic Memory in Hippocampal Neuronal Ensembles
Leutgeb, Leutgeb, Barnes, Moser, McNaughton, Moser. Science, 2005
1

0.8

0.6

0.4

0.2

1,2,3,n

-0.5

-0.4

-0.3

-0.2

-0.1

0.1

0.2

0.3

0.4

0.5

Fluctuations in the activity bump are not all noise but reflect the influence of external cues on the cells encoding a given location

0.8

0.6

0.4

0.2

1,2,3,n Cell # / location

-0.5

-0.4

-0.3

-0.2

-0.1

0.1

0.2

0.3

0.4

0.5

Constant Cues - Variable Place - GLOBAL REMAPPING Variations in cues affect how strongly cells fire, but not which cells can fire. The potentially active population is determined by the path integrator. Independent Codes for Spatial and Episodic Memory in Hippocampal Neuronal Ensembles
Leutgeb, Leutgeb, Barnes, Moser, McNaughton, Moser. Science, 2005
1

0.8 0.6

0.4

0.2

1,2,3,n

-0.5

-0.4

-0.3

-0.2

-0.1

0.1

0.2

0.3

0.4

0.5

Fluctuations in the activity bump are not all noise but reflect the influence of external cues on the cells encoding a given location

0.8 0.6

0.4

0.2

1,2,3,n Cell # / location

-0.5

-0.4

-0.3

-0.2

-0.1

0.1

0.2

0.3

0.4

0.5

Inbound

Outbound

Independent codes for places and events

Constant Place Variable Cue Constant Cue Variable Place

Place 1

Place 2

r>0

r~0

The population vectors of active neurons on two visits to the same place span statistically correlated subspaces, even if the cues are different. The population vectors of active neurons in two different places span statistically independent subspaces, even if the cues are similar.

Cue 1 Cue 2

Place 1 Place 2

spatial context sensitive context insensitive

Hippocampal recipient layers are modality AND context sensitive

How could a toroidal synaptic matrix be wired up by self-organization?

What if inhibition is not global?

Range of excitatory connections

Range of inhibitory connections Mexican hat function or DOG (difference of Gaussians)

What if inhibition is not global?

Range of excitatory connections

Range of inhibitory connections Network forms multiple bumps at closest packing density allowed by range of inhibition

What if inhibition is not global?

Range of excitatory connections

Range of inhibitory connections Network forms multiple bumps at closest packing density allowed by range of inhibition
0 20 40 60 80 100 120

McNaughton, Battaglia, Jensen, Moser & Moser (Nat. Rev. Neurosci. 2006)

The effect of local excitatory connections and local inhibitory connections can be multiple bumps of excitation in one or two dimensions This phenomenon was first The mexican hat function: Short range described theoretically by excitation & long range inhibition the mathematician Alan Turing (but in a different context).

Spontaneous symmetry breaking in a topographically structureless Turing layer with short range excitatory connections and longer range inhibitory connections

McNaughton, Battaglia, Jensen, Moser & Moser (Nat. Rev. Neurosci. 2006)

Can a topographically structureless cortical sheet that gives rise to a grid of activity in brain space be used to organize connections in a network that will give rise to
The mexican hat function: Short range excitation & long range inhibition

cells that fire topographically in a grid like fashion in physical space? In other words, can a toroidal synaptic matrix be self-organized during early development?

McNaughton, Battaglia, Jensen, Moser & Moser (Nat. Rev. Neurosci. 2006)

Turing layer Mexican hat connections Impose random drift of grid phases random top down connections (competitive learning)

developing grid cell layer random connections (Hebbian learning)

McNaughton, Battaglia, Jensen, Moser & Moser (Nat. Rev. Neurosci. 2006)

Output layer self-organizes as a toroidal synaptic matrix