the hippocampus and spatial representation · the hippocampus and spatial representation...
TRANSCRIPT
24/10/2019
1
The Hippocampus and spatial representation
Andrej Bicanski
NEUR3041: Neural computation
Aims
2 The hippocampus and spatial representation
• Introduce the hippocampus and place cells
• Explore an unsupervised, competitive learning model of place cells (Sharp, 1991)
• Discuss subsequent experiments emphasising importance of boundaries to place cells, and a fixed feed-forward model sufficient to explain place cell firing (Hartley et al., 2000)
• Revise the model to accommodate results suggesting place cell stability and robustness requires synaptic plasticity (Barry and Burgess 2007)
1
2
24/10/2019
2
The Hippocampal Formation
3
Rat Human
The hippocampus and spatial representation
Hippocampus with its subfields (DG, CA1,2,3)Entorhinal cortex(Parahippocampal cortex)
Importance of the Hippocampus: Oliver Sacks and Jimmie G
4 The hippocampus and spatial representation
”isolated in a single moment of being…. he is a man without a past (or future), stuck in a constantly changing, meaningless moment” (Sacks, 1985)
3
4
24/10/2019
3
5 The hippocampus and spatial representation
In-vivo electrophysiology
Recording single neurons in behaving animals
Hafting et al., 2005
O’Keefe & Dostrovsky 1971
6
Place cells allocentric location,Hippocampus
Grid cells path integration,Entorhinal cortex
In humans too!Robertson et al. 1999; Ekstrom et al. 2003; Jacobs et al. 2010; Doeller et al. 2010, Horner et al. 2016; Shine et al. 2019
Head-direction cells allocentric orientationPapez’ circuit
Ranck, 1984; Taube et al 1997
Single neuron responses in the hippocampal formation
Lever et al. 2009Hoydal et al. 2019
Boundary/Object Vector cells allocentric direction and distance of objects/boundaries
5
6
24/10/2019
4
Place fields are receptive fields (for location)
*
*
*
Compare to well known receptive fields in sensory space
8 The hippocampus and spatial representation
Place cells
Place Cell
7
8
24/10/2019
5
9 The hippocampus and spatial representation
Some properties of place cell firing
1. Firing is independent of the rat’s orientation in open but not narrow-armed mazes
2. Firing is robust to the removal of sensory cues
10 The hippocampus and spatial representation
Recap: Unsupervised, competitive learning
x1 x2 x3 x4 x5
o1 o2 o3
• ‘Winner-takes-all’: For each input pattern xn , output ok
(i.e. hk>hi for all ik). Set ok=1, oi=0 for all ik
• Random initial connection weights. All-to-all connectivity
• Hebbian learning: wij→ wij + oi xjn
i.e. wkj→ wkj + xjn, other weights don’t change
where oi =0
• Normalisation: reduce total size of connection weights to
each output (so |wi| =1) by dividing each by |wi|
• Process repeats with presentation of each input pattern
The output whose weights are most similar to xn wins and its weights then
become more similar. Different outputs find their own clusters in input data
Rumelhart and Zipser 1986.
PDP book freely available on J. McClelland’s website
wij
9
10
24/10/2019
6
11 The hippocampus and spatial representation
Sharp’s (1991) place cell firing model
A
B
C
D
E
F
GH
Cortical inputs:
Cue distances, cue directions
Competitive learning:
Competitive learning:
Entorhinal
cortex
Place cells
A B C D E F G H A B C D E F G Hnear
far
right
left
12 The hippocampus and spatial representation
Sharp’s (1991) model continued
Simulated place cell firing is resistant to cue-removal
11
12
24/10/2019
7
13 The hippocampus and spatial representation
Sharp’s (1991) model continued
Simulated place cells are omni-directional only after random exploration and not following directed exploration
14 The hippocampus and spatial representation
Summary
• Sharp’s competitive learning model explains robustness and directionality of place fields
13
14
24/10/2019
8
15 The hippocampus and spatial representation
What are the inputs to place cells?
‘simple’ place field firing is consistent
4.6 4.1
4.9 5.0
61cm 122 cm
0% 20% 40% 60% 80% peak
Y
X
N
S
W E
window
rat forages in box
O’Keefe and Burgess, 1996
16 The hippocampus and spatial representation
What are the inputs to place cells?
O’Keefe and Burgess, 1996
5.1 3.5
6.7 6.5
14.1 2.7
8.7 4.3
Other place fields appear to respond at fixed
distances from walls, with nearer walls more
important than further ones.
15
16
24/10/2019
9
17 The hippocampus and spatial representation
BVCs as hypothesised inputs to place cells
Hartley et al. 2000
• Each BVC tuned to respond when a barrier lies at a specific distance from the rat in a particular allocentric direction
Firing Rate
Receptive Field
Boundary Vector Cell (BVC)
18 The hippocampus and spatial representation
BVCs as hypothesised inputs to place cells
Hartley et al. 2000
• BVC firing determined by product of two gaussians, defining distance and angle tuning to boundary:
• Sharper tuning for shorter distances, reflecting uncertainty in inputs to BVCs
2
22
2
22
2
]2/)(exp[
)(2
)](2/)(exp[),(
ang
angi
irad
iradii
d
ddrrg
−−
−−
17
18
24/10/2019
10
19 The hippocampus and spatial representation
A fixed, feed-forward model of place cells from BVCs
S
Place field
BVC inputs
Place fields are modelled as the thresholded sum of 2 or more BVC firing fields.
Hartley et al. 2000
reminder
20 The hippocampus and spatial representation
Modelling O’Keefe and Burgess 1996 with BVCs
Simulated Place Fields
Place Cell Data
(O’Keefe & Burgess,
1996)
2-4 BVCs orientated at
right angles to one
another and
thresholded are
sufficient to fit most
fields
Best fitting BVCs
19
20
24/10/2019
11
21 The hippocampus and spatial representation
Modelling O’Keefe and Burgess 1996 with BVCs
DataO’Keefe & Burgess (1996)
22 The hippocampus and spatial representation
Modelling O’Keefe and Burgess 1996 with BVCs
ModelHartley, et al.(2000)
21
22
24/10/2019
12
23 The hippocampus and spatial representation
BVCs predict boundary-referenced firing of PCs across environments
Model
BVCs
Data (Lever et al)
1. Record some
place cells
4. Record same cells
in predicted environments
2. Find BVCs that best fit
those place cells
3. Use those model BVCs to predict
firing in other environments
24 The hippocampus and spatial representation
Discovery of BVCs
Lever et al. 2009. Solstad et al. 2008
23
24
24/10/2019
13
25 The hippocampus and spatial representation
Importance of synaptic plasticity to place field stability
• Synaptic plasticity the putative biological instantiation of weight-updating.
• ‘Vanilla’ plasticity requires NMDA receptors.
• Place field stability falls when NMDA receptors are blocked:
Saline CPP (NMDAR antagonist)
Kentros et al. 1998
26 The hippocampus and spatial representation
Importance of synaptic plasticity to place field stability
• Robustness of place fields to cue removal also depends on NMDA receptors.
Nakazawa et al. 2002
Saline CPP (NMDAR antagonist)
25
26
24/10/2019
14
27 The hippocampus and spatial representation
Slow, experience-dependent changes to place fields
Lever et al. 2002Likely reflect synaptic plasticity. Can’t be explained by a ‘fixed’ model.
28 The hippocampus and spatial representation
Summary
• Sharp’s competitive learning model explains robustness and directionality of place fields
• Subsequent results showing place cell firing is determined by geometry of environmental borders can be explained by a simple feed-forward model, given that boundary vector cells are the inputs to place cells
• But, such models cannot account for long-term dynamics of place fields in similar environments, nor the importance of synaptic plasticity in place field stability
27
28
24/10/2019
15
29 The hippocampus and spatial representation
Extended BVC model: added learning
• Update weights using BCM ruleBVCs
Place cells
Post synaptic activity yi
sliding
threshold θ
yi = f(∑jwij xj)Δwij = xj φ(yi, θ)
θ ~ <yi2>
requires pre & post-synaptic activity (xj and yi)
reduction if post-synaptic activity low
increase if activity high
φ(yi, θ)
Barry & Burgess (2007)
xj
yi
wij
30 The hippocampus and spatial representation
4.7 3.3 3.8 4.0 4.3
4.7 3.3 3.8 4.0 4.3
Iterations 20 40 60 80 100
Calculate place cell
firingBVC firing
Update weights
with BCM
Extended BVC model: added learning
29
30
24/10/2019
16
31 The hippocampus and spatial representation
Extended BVC model: added learning
5.0 3.8 2.7 2.9 3.1
3.0 3.4 3.5 3.8 3.5
2.1 2.6 2.6 2.5 2.5
0 10 20 30 40
Iterations
5.7 3.7 3.9
4.9 2.8 3.6
6.5 2.8 2.8
0 5 10
Iterations
32 The hippocampus and spatial representation
Extended BVC model: added learning
day2
day9
21.2 9.3 20.0 21.1 13.7 20.7
21.1 17.2 14.2 18.4 19.3 16.6
…
Barry (2006)
31
32
24/10/2019
17
33 The hippocampus and spatial representation
Summary
• Sharp’s competitive learning model explains robustness and directionality of place fields
• Subsequent results showing place cell firing is determined by geometry of environmental borders can be explained by a simple feed-forward model, given that boundary vector cells are the inputs to place cells
• But, such models cannot account for long-term dynamics of place fields in similar environments, nor the importance of synaptic plasticity in place field stability
• Combining the feed-forward BVC model and BCM learning accounts for long-term remapping
33