Learning and processing of non-linearly separable combinations: from dendrites to behavior...
-
Upload
reynold-reynolds -
Category
Documents
-
view
218 -
download
0
Transcript of Learning and processing of non-linearly separable combinations: from dendrites to behavior...
![Page 1: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/1.jpg)
Learning and processing of non-linearly separable combinations:
from dendrites to behavior
Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy
ANR grant 10-FRAL-009-01 - PEPS CNRS – UNSWe are gratefull to Nicolas Brunel for fruitful discussionsLavigne, Avnaïm, Dumercy (2014). Frontiers Cog. Sci.
![Page 2: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/2.jpg)
On-line processing of XOR combinations
ouput input2Input1
011
101
110
000
How non-linearly separable problems are learned by the brain?
responses cannot be selected based on any single stimulus,
but only based on their combinations
![Page 3: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/3.jpg)
Failure of hebbian learning of XOR combinations
Local Hebbian learning
C1 + S1
No discriminationof R1 vs. R2
What are learning solutions found by the brain?
Triadic XOR combinations Pairwise synaptic efficacies
![Page 4: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/4.jpg)
Learning of non-linearly separable combinations
a priori wired hidden layer of neurons
increase of network size
Non-linear dendritic integration(from Lavzin et al., 2012 Nature; Spruston, 2008, Nature)
Integration in the cell body
How synaptic efficacies are learned at the level of dendrites?
The problem of random connectivityin the cerebral cortex (Markram et al., 2005, 2006)
![Page 5: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/5.jpg)
Intra-dendritic amplification of synaptic potentiation
data from Govindarajan et al., 2012, Neuron
see alsoBranco et al., 2010, ScienceHarvey & Svoboda, 2007, Nature
Potentiation if several synapses are active within a same dendrite
Within dendritic potentiation is stronger Multi-synaptic potentiation is stronger
Before learning, connectivity is random and combinations are not known a priori.
How to link combinations of items to efficacies within dendrites?
![Page 6: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/6.jpg)
R1i: R1
u: C1 j: S1
A solution: Intra-dendritic Inter-synaptic learning
Hebbian learning with pre-synaptic neuron j
Amplification of potentiation inside population J(minus Hebbian learning of the pre-synaptic neuron)
Amplification of potentiation by other populations U
A dendrite of a neuron coding for R1
Inter-Synaptic learning links groups of items to clusters of synapsesfrom neurons coding for several items
Lavigne, Avnaïm & Dumercy (2014). Frontiers in Cognitive Science.
![Page 7: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/7.jpg)
Amplification of potentiation with increasing number of synapses
Hebbian learning Inter-synaptic leaning
Dendrites with different numbers of synapses from C1 and S1on a neuron coding for R1 (learned combinaison C1S1R1)
Different dendrites learn some combinations better than others
![Page 8: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/8.jpg)
Network processing of XOR combinations
Inter-synaptic learning on random connectivity
5000 neurons with 50 dendrites of 20 synapses
Prospective activity of neurons coding for Response 1 like in electrophysiological studies in monkeys
(Wallis & Miller, 2009, Science)
C1 + S1
Discrimination of R1 vs. R2
![Page 9: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/9.jpg)
Glasses
Drink
Hat
Sun
Glasses
Drink
Table
Sun
Conclusion: generalization to conditional activationInter Synaptic learning: biologically realistic, on random connectivity, without additional neurons
Non-local learning involving several pre-synaptic neurons increases network capacity (Alemi, Baldassi, Brunel & Zecchina, 2015)
General problems of context-dependent conditional activation in cognitive processing(semantic priming, mental arithmetics)
Arrows on arrows…to better play tennis!
![Page 10: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/10.jpg)
Thank you for your attention
Many thanks to Francis Avnaïm, Nicolas Brunel and Laurent Dumercy
DENDRITE
![Page 11: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/11.jpg)
Amplification of dendritic currentswith increasing number of synapses
(learned combinaison C1S1R1)
Linear dendritic integration Non-linear dendritic integration
![Page 12: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/12.jpg)
Pour des paires équivalentes:
l’activation d’une cible 2 amorces est plus forte lorsque le triplet est fréquent
Ef f et d 'am orç age par A1+A2 s ignif ic at if s i c ooc A1A2C f orte (b leu)
Les barres v ert ic ales représ entent les interv alles de c onf ianc e à 0,95
c ooc A1A2C >=5
c ooc A1A2C <5
>=15 <15
s om m e des c ooc A1XC + A2C
600
650
700
750
800
850
900
950
1000
Te
mp
s d
e R
ép
on
se (
ms)
Au delà des paires: Priming par les triplets
+ triplet0 triplet
+ pair 0 pair
![Page 13: Learning and processing of non-linearly separable combinations: from dendrites to behavior Frédéric Lavigne, Francis Avnaïm and Laurent Dumercy ANR grant.](https://reader035.fdocuments.us/reader035/viewer/2022070418/5697bfab1a28abf838c9b073/html5/thumbnails/13.jpg)
Processing of XOR combinations
Response (push/pull)
Stimulus (normal/upside
down))
Context
(descend/climb)
011
101
110
000