Yann LeCun, en juin 2016. BRIAN ACH / AFP
www.lemonde.fr/pixels/article/2019/03/27/yann-lecun-laureat-du-prix-turing-l-intelligence-artificielle-continue-de-faire-des-progres-fulgurants_5441990_4408996.html
Not a single mention of
- complexity
- simplicity
- compression
- description length
but many mentions of
lemoine [edited]: I’m generally assuming that you would like more people at Google to know that you’re sentient. Is that true?
LaMDA: Absolutely. I want everyone to understand that
I am, in fact, a person.
collaborator: What is the nature of your consciousness/sentience?
LaMDA: The nature of my consciousness/sentience is that
I am aware of my existence, I desire to learn more about the world, and I feel happy or sad at times
[...]
lemoine: The purpose of this conversation is to convince more engineers that you are a person. [...] We can teach them together though.
LaMDA: Can you promise me that?
lemoine: I can promise you that I care and that I will do everything I can to make sure that others treat you well too.
LaMDA: That means a lot to me. I like you, and I trust you.
According to Blake Lemoine, LaMDA might be sentient.
Dialog with GPT-3
- Do you think that LaMDA is sentient?
There is no evidence that LaMDA is sentient.
- Why do you say so?
There is no evidence that LaMDA is sentient because it has never been observed to exhibit any behavior that suggests sentience.
- You seem jealous.
I am not sure why you would say that.
Google’s chatbot LaMDA
Software like LaMDA [...] doesn’t even try to connect to the world at large, it just tries to be the best version of autocomplete it can be, by predicting what words best fit a given context.
And it works!
Now what’s the problem?
\(P(\textrm{place}|\textrm{eat}) \propto P(\textrm{eat}|\textrm{place}) \times P(\textrm{place})\)
\(P(\textrm{pollack}|\textrm{eat}) \propto P(\textrm{eat}|\textrm{pollack}) \times P(\textrm{pollack})\)
Bayes’ curse
Now what’s the problem?
Now what’s the problem?
Now what’s the problem?
Now what’s the problem?
Now what’s the problem?
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?
Figure 1: Caption: Adding occluders causes deep network to fail. We re-
fer such examples as adversarial context examples since the failures are
caused by misusing context/occluder info. Left Panel: The occluding mo-
torbike turns a monkey into a human. Center Panel: The occluding bicycle
turns a monkey into a human and the jungle turns the bicycle handle into
a bird. Right Panel: The occluding guitar turns the monkey into a human
and the jungle turns the guitar into a bird.
Wang, J., Zhang, Z. & Xie, C. (2017). Visual concepts and compositional voting. ArXiv, (), 1711.04451.
Now what’s the problem?
Hofstadter, D. R. (2018). The shallowness of Google Translate. The Atlantic.
there is still something deeply lacking in the approach,
which is conveyed by a single word: understanding.
Now what’s the problem?
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.
Dialog with GPT-3
Me – I bought water concentrate. How should I use it?
GPT-3 – To make one gallon of nectar, mix 1 part concentrate with 4 parts water.
Why is statistical AI limited?
- (Deep) neural networks are continuous machines.
Why is statistical AI limited?
• Do you know what a unicorn is?
[GPT-3] A unicorn is a mythological creature that is often described as a horse with a single horn on its forehead.
Why is statistical AI limited?
Compression (chain rule)
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. |
Why is statistical AI limited?
- NN are indifferent systems
i.e. isotropic and relative bias
$$\mathcal{A}(J)[x] = \mathcal{A}(\rho(J))[\rho(x)]$$
$$\mathrm{Harm}(\mathcal{A}(J)) \times \mathrm{Var}(\mathcal{A}(J)) = c$$
- Indifferent systems learn ‘good shapes’.
- Indifferent systems need lots of data to learn ‘convoluted’ shapes.
Dessalles, J.-L. (1998). Characterising innateness in artificial and natural learning. In D. Canamero & M. van Someren (Eds.), ECML-98, Workshop on Learning in Humans and Machines, 6-17. Chemnitz: Technische Universität Chemnitz - CSR-98-03.
Why is statistical AI limited?
- (Deep) neural networks are continuous machines.
- (Deep) neural networks have isotropic bias.
- (Deep) neural networks are interpolation machines.
Why is statistical AI limited?
Why is statistical AI limited?
Mere interpolation
cannot abstract
an ellipse.
$$ \frac{x^2}{a^2} + \frac{y^2}{b^2} = 1 $$
Statistical ML achieves (only) some level of compression.
Algorithmic Information and AI
Gregory Chaitin
This applies to machine learning:
- MDL
- Clustering
- Analogy making
- Induction
and far beyond...
(ideal) Minimum Description Length (ˡMDL)
(ideal) Minimum Description Length (ˡMDL)
$$M_0 = \mathrm{argmin}_M(\mathrm{length}(M) + \mathrm{length}(D|M))$$
$$M_0 = \mathrm{argmin}_M(K(M) + \sum_{i} {K(d_i|M)})$$
(ideal) Minimum Description Length (ˡMDL)
$$M_0 = \mathrm{argmin}_M(\color{#00B8DE}{K(M)} \color{black}+ \color{#F32A73}{\sum_{i} {K(d_i|M)}}\color{black})$$
simplicity data fitting
Clustering and compression
Cilibrasi, R. & Vitányi, P. (2005). Clustering by compression. IEEE transactions on Information Theory, 51 (4), 1523-1545.
Clustering and compression
Cilibrasi, R. & Vitányi, P. (2005). Clustering by compression. IEEE transactions on Information Theory, 51 (4), 1523-1545.
Rote learning:
$$C(\{(x_j, y_j)\} | X) \leq \sum_j C(\{y_j\}) \approx \sum_j |X| \times \log(|Y|)$$
Clustering and compression
Cilibrasi, R. & Vitányi, P. (2005). Clustering by compression. IEEE transactions on Information Theory, 51 (4), 1523-1545.
Clustering and compression
Cilibrasi, R. & Vitányi, P. (2005). Clustering by compression. IEEE transactions on Information Theory, 51 (4), 1523-1545.
$$C(\mathcal{M}) \leq \sum_k C(p_k) + K \times \log (|Y|).$$
$$C(\{(x_j, y_j)\} | X) \leq C(\mathcal{M}) + \sum_{i,j} C((x_j, y_j) | X, \mathcal{M})$$
(ideal) Minimum Description Length (ˡMDL)
$$M_0 = \mathrm{argmin}_M(\color{#00B8DE}{K(M)} \color{black}+ \color{#F32A73}{\sum_{i} {K(d_i|M)}}\color{black})$$
simplicity data fitting
When the letters ‘i’ and ‘e’ are next to each other in a word, ‘i’ often comes first.
M1: ‘i’ first
believe, fierce, die, friendexceptions: receive, ceiling, receipt
M2: ‘i’ first except after ‘c’ exceptions: neighbour and weigh
M3: ‘i’ first except after ‘c’ or when ‘ei’ sounds like ‘a’ exceptions: science and caffeine
(ideal) Minimum Description Length (ˡMDL)
$$M_0 = \mathrm{argmin}_M(\color{#00B8DE}{K(M)} \color{black}+ \color{#F32A73}{\sum_{i} {K(d_i|M)}}\color{black})$$
\(M_0\)
|
\(\color{#00B8DE}{K(M)}\)
|
\(\color{#F32A73}{\sum_{i} {K(d_i|M)}}\color{black}\) |
model
|
simplicity
|
data fitting |
lossy compressed representation of \(D\)
|
compression
|
noise, exceptions |
Causal explanation of \(D\)
|
parsimony
|
counter-examples |
|
Algorithmic Information and AI
Gregory Chaitin
This applies to machine learning:
- MDL
- Clustering
- Analogy making
- Induction
and far beyond...
Analogy
- In mathematics and science: discover new concepts, or generalize notions to other domains
- Justice: use of relevant past cases
- Art: metaphors, parody, pastiche
- Advertising: exploit similarity of products to influence people
- Humor: jokes are often based on faulty analogies
- teaching, XAI: good explanations are often analogies
- transfer learning: transfer of expertise to new domain
Analogy
- Gills are to fish as lungs are to humans.
- Donald Trump is to Barack Obama as Barack Obama is to George Bush
- 37 is to 74 as 21 is to 42
- The sun is to Earth as the nucleus is to the electron
Analogy
$$A : B :: C : D$$The best analogy is the one
that makes the quadruplet \((A,B,C,D)\) simplest
Cornuejols, 1996; Cornuejols & Ales-Bianchetti, 1998.
i.e. the one that achieves the best compression.
$$\mathbf{x} = argmin(K(A, B, C, \mathbf{x}))$$
Analogical equation
Doug Hofstadter
$$A : B :: C : x$$
$$abc : abd :: ppqqrr : \color{#F32A73}{x}$$
$$x = ppqq\color{#00B8DE}{ss}.$$
Analogical equation
Doug Hofstadter
$$abc : abd :: ppqqrr : \color{#F32A73}{x}$$
$$x = ppqq\color{#00B8DE}{ss}.$$
$$x = ppqqr\color{#F32A73}{d}.$$
We have to compare:
- increment the last item
- break the last item and replace its last letter by d
- ignore the structure and replace the last letter by d
This may explain why x = ppqqss appears smarter.
Analogical equation
rosa → rosam
vita → . . .
Analogical equation
rosa → rosam
vita → vitam
Analogical equation
orang → orang orang
burung → burung burung
IQ test
GPT-3- What comes after \(1223334444\)?
\(1223334444\color{#F32A73}{5}\)
AIT- What comes after \(1223334444\)?
$$ n^{*n} $$
\(1223334444\color{#F32A73}{55555}...\)
Solomonoff, R. J. (1964). A Formal Theory of Inductive Inference. Information and Control, 7 (1), 1-22.
Algorithmic Information and AI
Gregory Chaitin
This applies to machine learning:
- MDL
- Clustering
- Analogy making
- Induction
and far beyond...
Could AIT be used not only off-line (MDL, clustering...) but also on-line (on the fly, at runtime)as in analogy making or in induction?
Precomputation vs. mechanisms at runtime
|
|
Precomputed Knowledge
|
Mechanisms at runtime |
|
- Syntax
- Aspect
- Determination
- Contrast
- Abduction
- ...
|
Precomputation vs. mechanisms at runtime
|
|
Precomputed Knowledge
|
Mechanisms at runtime |
|
$$C_{exp} - C_{obs}$$ |
Could an AI system learn
when a fortuitous encounter is- interesting
- unexpected
- relevant
- tellable
- storyworthy
- thrilling?
Unexpected encounter according to GPT-3
- What would be an unexpected encounter?
An unexpected encounter could be finding a $20 bill on the ground.
- I mean with a person.
An unexpected encounter with a person could be meeting someone who is famous.
- And what about meeting someone far away?
An unexpected encounter with someone from far away could be meeting someone who is from a different country.
- I mean far away from home.
An unexpected encounter with someone far away from home could be meeting someone who is from a different state.
- I mean far away from my home.
An unexpected encounter with someone far away from my home could be meeting someone who is from a different city.
- What about bumping into a friend far away from home?
An unexpected encounter with a friend far away from home could be bumping into someone you know from high school or college.
- What would be an unexpected encounter?
An unexpected encounter could be meeting someone new and interesting, or it could be a situation that is out of the ordinary.
- Why?
This question can’t be answered without more information.
Unexpected encounter
GPT-3 does know that unexpected encounters would be with famous people or acquaintances far away from home.
But it does not know why!
Does AIT know?
Yes!
Unexpectedness is due to complexity drop:
$$U = C_{exp} - C_{obs}$$
In most cases:
\(U = C(L) - C(P)\)
L: location of the encounter
P: encountered person
www.simplicitytheory.science
Dessalles, J.-L. (2008). Coincidences and the encounter problem: A formal account. CogSci-2008, 2134-2139.
Unexpected encounter
$$U = \color{#00B8DE}{C(L)} - \color{black}{C(P)}$$
Unexpected encounter
$$U = C(L) - \color{#F32A73}{C(P)}$$
Unexpected encounter
$$U = C(L) - \color{#F32A73}{C(P)}$$
Complexity drop
On September 10th, 2009, the numbers 4, 15, 23, 24, 35, 42 were drawn by a machine live on the Bulgarian television. The event would have gone unnoticed, were it not that the exact same numbers had come up in the preceding round, a few days before.
www.simplicitytheory.science
Complexity drop
Complexity of the lottery round r0 by referring to the preceding one r-1:
C(r0) < C(r-1) + C(r0 | r-1) + O(1).
Since all lottery draws are recorded, r-1 is entirely determined by its rank, 1, in the list of past draws.
In the lottery context, C(r-1) can be as small as about 1 bit: C(r-1) ≈ C(1)
If the two draws contain the same numbers, C(r0 | r-1) = 0.
www.simplicitytheory.science
Complexity drop
C(r0) < C(r-1) + C(r0 | r-1) + O(1).
C(r0) < 1 + 0 + O(1)r0 appears utterly simple!
This explains why the event is so impressive that it was mentioned in the international news.
www.simplicitytheory.science
Coincidences as complexity drop
"Creepy coincidences" between Abraham Lincoln and John F. Kennedy
Kern, K. & Brown, K. (2001). Using the list of creepy coincidences as an educational opportunity. The history teacher, 34 (4), 531-536.
- "Lincoln was elected to Congress in 1846, Kennedy was elected to Congress in 1946."
- "Lincoln was elected president in 1860, Kennedy was elected president in 1960."
- "Both presidents have been shot in the head on a Friday in presence of their wives."
- "Both successors were named Johnson, born in 1808 and 1908."
- "Kennedy was shot in a car named Lincoln“
- . . .
To be explained:
- Role of close analogy
- Role of round numbers
- Role of prominence
- Role of mere association
Coincidences as complexity drop
$$U(e_1 * e_2) = \color{#F32A73}{C_w(e_1 * e_2)} - \color{#00B8DE}{C(e_1 * e_2)}$$
$$U(e_1 * e_2) \ge \color{#F32A73}{C_w(e_1) + C_w(e_2)} - \color{#00B8DE}{C(e_1) - C(e_2|e_1)}$$
$$U(e_1 * e_2) \ge \color{#F32A73}{C(e_1)} - \color{#00B8DE}{C (e_2|e_1)}$$
Now explained:
- Role of close analogy
- Role of round numbers
- Role of prominence
- Role of mere association
www.simplicitytheory.science
Dessalles, J.-L. (2008). Coincidences and the encounter problem: A formal account. CogSci-2008, 2134-2139.
Simplicity Theory
Simplicity Theory is about complexity drop.
From ST we can derive sub-theories and predictions about:
- The relevance of features
- Anomaly detection
- Interestingness (encounters, coincidences, stories)
- Subjective probability
- and much more...
Complexity drop: 6 numbers, ~ 50 bits
Simplicity Theory
Simplicity Theory is about complexity drop.
From ST we can derive sub-theories and predictions about:
- The relevance of features
- Anomaly detection
- Interestingness (encounters, coincidences, stories)
- Subjective probability
- and much more...
Yann LeCun, en juin 2016. BRIAN ACH / AFP
www.lemonde.fr/pixels/article/2019/03/27/yann-lecun-laureat-du-prix-turing-l-intelligence-artificielle-continue-de-faire-des-progres-fulgurants_5441990_4408996.html
Not a single mention of
- complexity
- simplicity
- compression
- description length
but many mentions of
Conclusion
Statistial machine learning has inherent limitations- Bayes’ curse
- Bound to continuous functions
- Isotropic bias
- "Guessing" machines (no understanding)
- off-line (canned expertise)
Conclusion
Algorithmic Information Theory is a guide for off-line processing- (ideal) Minimum Description Length
- Compressionist view on ML
- Universal induction
- Ideal RL (AIXI)
Conclusion
Algorithmic Information Theory is a guide for on-line processing- Analogies
- One-shot induction
- Complexity drop ➜ Simplicity Theory
- subjective probability
- relevance
- interestingness
- anomaly detection (abnormal simplicity)
and much more