Should  AI  imitate cognitive mechanisms?

Avion.jpg Oiseau.jpg
  • Planes do not fly like birds
  • Computers can retrieve trillions of π’s decimals
  • Computers can learn Chinese just by reading Chinese texts
TrillionPi.png Chinese.png

 AI  relies on data

Miner.png Transformers1.png Transformers2.png

And it works!

Muffins.png

Krizhevsky, A., et al. (2012). Imagenet classification with deep convolutional neural networks. NIPS 2012, 1097-1105.


Alphago.png

    

    

    

    

    


Silver, D. et al., (2017). Mastering the game of go without human knowledge. Nature, 550 (7676), 354-359.

And it works!

Advocado0.png Advocado.png

Now what’s the problem?

MangeLieu1.png MangeLieu6.png

Now what’s the problem?

MangeLieu2.png

Now what’s the problem?

MangeLieu3.png

Now what’s the problem?

MangeLieu4.png

Now what’s the problem?

MangeLieu5.png

Now what’s the problem?

    

  • "The senators were helped by the managers."    →    The senators helped the managers.

  • "The managers heard the secretary resigned."    →    The managers heard the secretary.
  • "If the artist slept, the actor ran."    →    The artist slept
  • "The doctor near the actor danced."    →    The actor danced.

    

McCoy, R. T., Pavlick, E. & Linzen, T. (2019).
Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. ArXiv:1902.01007.

Now what’s the problem?


r9Z9V9K2RNuITm5s6lAdLw_thumb_c987.jpg

Marcus, G. & Davis, E. (2020). GPT-3, Bloviator: OpenAI’s language generator has no idea what it’s talking about.
MIT Technology review

[GPT-3] can produce words in perfect English, but it has only the dimmest sense of what those words mean, and no sense whatsoever about how those words relate to the world.

MM2021.jpg

Mitchell, M. (2021). Why AI is harder than we think.
ArXiv:2104.12871.

In 2016 Business Insider assured us that "10 million self-driving cars will be on the road by 2020". Tesla Motors CEO Elon Musk promised in 2019 that "A year from now, we’ll have over a million cars with full self-driving, software..." [...] Despite attempts to redefine "full self-driving" into existence,
none of these predictions has come true.

Now what’s the problem?

GreenBook.png

    

AI  systems are evaluated
on generic contexts.

    

Intelligent systems should be able to adapt to specific contexts.

Now what’s the problem?

GreenBook.png
Bring me the little little bacterium
little book
little galaxy
green #00FF00
book that is to the right of to the right of the lamp
to the right of the bus
to the right of the Eiffel Tower
the lamp on The car is on the road.
The city is on the map.
The fly is on the ceiling.
my desk. She is sitting on my desk.
Go to the reception desk.
A desk at the Foreign Office.

Why is statistical AI limited?

DeepLearning.png

  • (Deep) neural networks are continuous machines.

  • They learn how to interpolate between examples.
GREEN PLANT
Plant2.png Plant2_mask.png 0.817044
Plant1.png Plant1_mask.png 0.630777
Plant7.png Plant7_mask.png 0.223796
Plant5.png Plant5_mask.png 0.110613
Plant6.png Plant6_mask.png 0.087145
Plant3.png Plant3_mask.png 0.076637
Plant4.png Plant4_mask.png 0.027321

Why is statistical AI limited?

  • A trivial example

This system just counts pixels
that are particularly green.

The only thing to learn is the "green" threshold.    →

    
Quite magically,
• small modifications would not alter the decision.
• Resembling plants will be in the same category.

This "magic" results from continuity.

Why is statistical AI limited?

DeepLearning.png

  • (Deep) neural networks are continuous machines.
  • They learn how to interpolate between examples.

  • DNN learn to pay attention to relevant features, but they remain interpolation machines.

Why is statistical AI limited?


Planete1.png

Why is statistical AI limited?


Planete2.png
    Mere interpolation
    cannot abstract          an ellipse.

$$ \frac{x^2}{a^2} + \frac{y^2}{b^2} = 1 $$

Why is statistical AI limited?

Advocado.png

    
Neural nets are essentially
"guessing machines"! MangeLieu3.png

Can we do better?


Intelligent behavior must be computed in part
on the fly!

    

Method: Find out some mechanisms that allow us humans
to be intelligent.

Some cognitive "mechanisms"

GoatDog.png 236

Syntax

    

It barked while
the dog was in the doghouse.

    

        The fact that it barked revealed
        that the dog was in the doghouse.

    

Chomsky, N. (1975). Reflections on language. New York: Pantheon Books.

Crain, S. (1991). Language acquisition in the absence of experience. Behavioral and Brain Sciences, 14 (4), 597-650.

Some cognitive "mechanisms"

Syntax1.svg Syntax2.svg

Syntax

It barked while
the dog was in the doghouse.

    

    The fact that it barked revealed
    that the dog was in the doghouse.

Some cognitive "mechanisms"

Aspect

Mary is sneezing like that every morning now.
Sneezing1.png

Elle n’est pas toujours malade ↔ elle n’est toujours pas malade
Malade1.png Malade2.png
    

Pulman, S. G. (1997). Aspectual shift as type coercion.
Transactions of the Philological Society, 95 (2), 279-317.

Munch, D. & Dessalles, J.-L. (2012). Inferring aspectuality on French sentences:
a minimalist approach. CogSci-2012, 2055-2060.

Some cognitive "mechanisms"

Aspect

Mary is sneezing like that every morning now.
Sneezing1.png Aspect1.svg

Some cognitive "mechanisms"

Aspect

Mary is sneezing like that every morning now.
Sneezing2.png

Aspect2.svg

Some cognitive "mechanisms"

Aspect

Mary is sneezing like that every morning now. 121 Aspect3.svg

Some cognitive "mechanisms"

Contrast

DeepLearning.png

Licorne0.png

Some cognitive "mechanisms"

Contrast

Licorne1.png

Licorne2.png

Gärdenfors, P. (2014). The geometry of meaning - Semantics based on conceptual spaces. Cambridge, MA: MIT Press.

Dessalles, J.-L. (2015). From conceptual spaces to predicates. In F. Zenker & P. Gärdenfors (Eds.), Applications of conceptual spaces: The case for geometric knowledge representation, 17-31. Dordrecht: Springer.

Some cognitive "mechanisms"

Simplicity

  • IQ test

$$ n^{*n} $$
122333444455555

The most intelligent continuation minimizes overall complexity.

complexity = size (in bits) of the shortest description.

    

Solomonoff, R. J. (1964). A Formal Theory of Inductive Inference. Information and Control, 7 (1), 1-22.

Some cognitive "mechanisms"

Simplicity

  • Analogy

rosa    →    rosam
vita    →    . . .

Some cognitive "mechanisms"

Simplicity

  • Analogy

rosa    →    rosam
vita    →    vitam

    

Murena, P.-A., Al-Ghossein, M., Dessalles, J.-L. & Cornuéjols, A. (2020). Solving analogies on words based on minimal complexity transformation. IJCAI, 1848-1854.

Dessalles, J.-L. (2008). La pertinence et ses origines cognitives - Nouvelles théories. Paris: Hermes Science.

Some cognitive "mechanisms"

Simplicity

  • Analogy

setzen    →    setzte
lachen    →    lachte

www.simplicitytheory.science

Murena, P.-A., Al-Ghossein, M., Dessalles, J.-L. & Cornuéjols, A. (2020). Solving analogies on words based on minimal complexity transformation. IJCAI, 1848-1854.

Dessalles, J.-L. (2008). La pertinence et ses origines cognitives - Nouvelles théories. Paris: Hermes Science.

Some cognitive "mechanisms"

"CAN"

    

CAN: Conflict — Abduction — Negation

CAN.png

Dessalles, J.-L. (2016). A Cognitive Approach to Relevant Argument Generation. In M. Baldoni, C. Baroglio, F. Bex, T. D. Bui, F. Grasso & et al. (Eds.), Principles and Practice of Multi-Agent Systems, LNAI 9935, 3-15. Springer.

Should  AI  imitate cognitive mechanisms?

    
Future  AI cannot rely exclusively on "canned" intelligence.

    

DeepLearning.png

ordi_p_2_moins_U.png

Should  AI  imitate cognitive mechanisms?

    
Future AI presupposes to perform
on-line computations.
to implement fundamental cognitive mechanisms
and execute them on the fly.

  • syntactic processes
  • aspect
  • contrast
  • simplicity
  • CAN
  • determination - relations - semantic linking . . .