1. SISSA limbo Cortex and language? December 7, 2016 25’ 35’ 15’ 15’ to
the
doesn’t
fight
that
those
noble
believes
thinks
stands
prays
gates
with
house
sheep
precious
Yasmine
sword
Cortex and language?
December 7, 2016
15’
35’
25’
15’

2. As to language in humans, I think G-d distinguished
them from beasts only to
show that they are one
and the same Ecclesiastes III, 18-19
monkeys cannot speak
because they lack a lan-
guage acquisition device
with universal grammar
what I mean is that
they cannot speak
because they lack
infinite recursion
but they do
latch from
one state
to the next
Moshe Abeles
yet we cannot speak
why?
King Solomon
the human brain has no extra language organ
Santiago Ramon y Cajal
Valentino Braitenberg
quantitative differences

3. Where does language hide?
Neuropsychology
(the analysis of brain-
lesioned patients)
offers hints:
These for example are
patients who cannot read,
and they have all
suffered lesions in the left
occipito-parietal cortex.

4. Lesions in patients Subject who reads found! area VWFA, where
one decodes the form
of written words

5. Neuropsichology, combined with imaging, says
however that the functions of a given area …….
Neuropsichology, combined with imaging, says
however that the functions of a given area can
can be moved elsewhere: e.g.,
language functions to the right hemisphere
(in the corresponding area)

6. model by Cohen,
Dehaene et al
So, where is the innovation that produced language?

7. A quantitative model to analyse?
The BraitenbergPotts net: a glocal associative memory
N pyramidal cells
√N compartments
√N cells each
A pical synapses
B asal synapses

8. (firing rate adaptation or synaptic adaptation)
Potts units with dilute
connectivity
S+1 Potts states
Sparse Potts patterns
Reduced to a Potts model
(Kropff & Treves, 2005)
Structured long-range
connectivity
“0” state included
Sparse global patterns
updated to remove the
‘memory glass’ problem
(Fulvi Mari & Treves, 1998)
Cortical modules
Local attractor states (=S)
Global activity patterns
A simple semantic network
(O’Kane & Treves, 1992)
..in Braitenberg’s model all cortical modules
share the same organization…
with adaptation
(firing rate adaptation or synaptic adaptation)
Latching dynamics !
(Kanter 1988)
pc  C S 2 !!
pc  S ?!?!

9. May latching dynamics model a train of thoughts ?
(cited by Alistair Knott in his PhD Thesis, 1996)
1) How far are they from deterministic AND from random ?
(see Kropff and Treves, 2006)
2) When can they proceed indefinitely in time (recursively) ?

10. Latching Phases
Eleonora Russo
Supervisor: Alessandro Treves

11. Simple simulations indicate a phase transition as p grows

12. The Energy Landscape

13. The Energy Landscape
Memory without latching

14. The Energy Landscape
Memory without latching

15. The Energy Landscape
Finite sequence of latching transitions

16. The Energy Landscape
Finite sequence of latching transitions

17. The Energy Landscape
Finite sequence of latching transitions

18. Energy and Adaptation

19. Energy and Adaptation
Energy
Adaptation

20. A phase diagram!
We can now understand it...
...and add back correlations

21. How might a capacity for indefinite latching have evolved?
semantics 
semantics  AM AM C S
long-range conn 
(local conn )
Storage capacity (max p to allow cued retrieval)
pc  C S 2
Latching onset (min p to ensure recursive process)
pl  S ?
a spontaneous transition to infinite recursion?

22. To go beyond “free thought” into language, with syntax,
a binding operation is required, with working memory
Differentiating syntax from semantics allows
- perhaps - for the implementation of syntactic rules
“operators”
self-organized
input
driven
pc  C S 2
“fillers”
pl  S ?
“fillers”
(see Battaglia et al, Eliasmith et al, Fusi et al, Huyck, Hashimoto, ..)

23. BLISS (w/ E Kropff, A Grüning, M Katran
Medium
Term:
to assess how any model can learn, one first needs a toy:
a basic language including both syntax and semantics
BLISS (w/ E Kropff, A Grüning, M Katran
with mild guidance by G Longobardi)
Sahar
Pirmoradian
+ semantic correlations
N  boy | girl | cat | dog | tiger | jackal | horse | cow | meat | hay | milk | wood | meadow | stick | fork | bowl | cart | table | house || boys | girls | cats | dogs | tigers | jackals | horses | cows | stables | sticks | forks | bowls | carts | tables | houses
PropN  John | Mary || John and Mary
V  chases | feeds | sees | hears | walks | lives | eats | dies | kills | brings | pulls | is || chase | feed | see | hear | walk | live | eat | die | kill | bring | pull | are |
Compl  that | whether
Prep  in | with | to | of | under
Neg  does not || do not (note singular negation removes sing inflection of verb)
Art  the | a(an)
AdjP  red | blue | green | black | brown | white | yellow | slow | fast | rotten | fresh | cold | warm | hot
AdvP  slowly | rapidly | close | far
Dem  this | that || these | those
Syntactic structure
150 terminal symbols (words)
39 nonterminal symbols
40 production rules

24. Sahar then encodes the BLISS words onto the Potts net
The phase diagram is more complicated, but has similar features...

25. Syntax driving semantics or viceversa?
Sahar, anybody
please, find out
Exploring both, at the moment separately

26. toy language acquisition: Mass / Count Syntax
Input layer of
source nodes
(semantics)
Output layer of
Neurons
(syntax)
Can mass count syntax be inferred from semantic properties?

27. Perceptron Geometric View
The equation below describes a (hyper-)plane in the input space consisting of real valued m-dimensional vectors. The plane splits the input space into two regions, each of them describing one class. x2
decision
region for count
decision
region for mass
w1x1 + w2x2 + w0 >= 0
…suppose semantic
features form clusters
over some mass count
-related dimensions...
decision
boundary
w1x1 + w2x2 + w0 = 0 x1

28. Naïve Expectation – for cross linguistic variability
meubles
furniture

29. Mass / Count Statistical Study
Ritwik Kulkarni since 2009
with Susan Rothstein since 2005
and many others
(published in Biolinguistics 2013)

30. Syntactic Usage Questions (English example)

31. Semantic Questions

32. Mass / Count Usage Table
1434 nouns (survived from a sample of ca 1650)
6 languages (among them 3 with several informants)

33. Mass / Count Statistics is NOT Bimodal

34. Mass / Count Statistics - Marathi
Semantics

35. Single Mass / Count Questions do NOT Match
nor do pairs
+ +
+ -
- +
- -

36. Main Mass / Count Dimension
, water, flour, sugar, ... )

37. Mass / Count (lack of) Agreement along the Main Dimension
High variance
Language
Entropy
*Armenian
1.63
*Italian
1.96
*Marathi
2.15
English
2.66
Hebrew
2.11
Hindi
1.54
*Semantics
(2.01)
1.58 (C) 1.24 (A)
Low MI

38. Mass / Count (lack of) Agreement in the full space
Language
Entropy
*Armenian
2.29
*Italian
3.02
*Marathi
2.71
English
3.92
Hebrew
3.40
Hindi
2.12
*Semantics
(3.72)
2.94 (C)
2.34 (A)

39. Mass / Count Agreement vs the Artificial control

40. Mass / Count in the CHILDES Corpus

41. yet Mass / Count Syntax was ready to help
Mass markers
Count markers
Multi-dimensional scaling in the
space of self-organized markers

42. Mass / Count Tentative Conclusions
Binary rules are non-binary and unruly:
Little agreement across languages
Limited basis on underlying semantics
Grammaticalization matures independently
self-organized
input
driven

43. maybe that is because learning occurs on a rough landscape
semantics
might look
something
like this..

44. Mean-field theory of the Potts glass
Phys. Rev. Lett. 55, 304 – Published 15 July 1985
D. J. Gross, I. Kanter, and H. Sompolinsky
predicts complex glassy behavior in Potts networks..
..what do we see in simulations (with limited asymmetry)?
half
successes
failures
No Symmetry + Disorder + Frustration + High Dimensionality …
*many funs*

45. (+)n+ (+)n+ ++- -(-)- Neocortex Hippocampus Cerebellum Basal ganglia
Lamination, Arealization
Hippocampus
(+)n+
DG input sparsifier
CA1 feed-forward
Who is
cutting-edge, in cerebellar technology?
Cerebellum
++-
Expansion recoding,
Private teachers
Basal ganglia
-(-)-
Massive funnelling
Tonic output firing
Tectum
Olfactory
bulb
Spinal cord
Computational
paradigms
100’s Myrs old
that we fail
to understand