Archive
Weekly QuEST Discussion Topics, 2 July
QuEST 2 July 2021
We want to continue the discussion of the Chollet work on Measures of Intelligence. First our disclaimer:
• Although our goal is consciousness – it is our position that creating intelligent systems will require understanding the impact of the conscious representation in enabling flexible cognition / flexible relationships between agents / flexible acquisition of new skills (tasks) and that results in more intelligent solutions – the Chollet framework provides interesting thoughts on how to quantify flexibility in the acquisition of new abilities and he ties that to his definition of intelligence. QuEST similarly ties flexibility to our view of the what consciousness affords.
This week we want to hit the parts of the paper by Francios Chollet from Google that touch on generalization and also the mathematics, on the math side we want to examine the terms in the Chollet formalism to understand implications to our views of consciousness – as we posted last week there are several online videos on the Chollet paper – see below – specifically for this week I recommend the video sequence by Yannick Kilcher and specifically his part III – during QuEST we will field questions associated with the details:
we want to have our discussion address our ideas of task flexibility possibly tied to considerations of the Chollet view of generalization (possibly using distributional considerations) and his approach to formalizing complexity,
we also want to discuss the implications of the work on aaco – how do we make our systems better positioned to do more tasks and adapt to issues with our simulators and also the implications to ACE (AI/ML DevSecOps environment)
related information from last week:
On the Measure of Intelligence
Franc¸ois Chollet _
Google, Inc.
November 5, 2019
arXiv:1911.01547v2 [cs.AI] 25 Nov 2019
To make deliberate progress towards more intelligent and more human-like artificial systems, we need to be following an appropriate feedback signal: we need to be able to define and evaluate intelligence in a way that enables comparisons between two systems, as well as comparisons with humans. Over the past hundred years, there has been an abundance of attempts to define and measure intelligence, across both the fields of psychology and AI. We summarize and critically assess these definitions and evaluation approaches, while making apparent the two historical conceptions of intelligence that have implicitly guided them. We note that in practice, the contemporary AI community still gravitates towards benchmarking intelligence by comparing the skill exhibited by AIs and humans at specific tasks, such as board games and video games. We argue that solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience: unlimited priors or unlimited training data allow experimenters to “buy” arbitrary levels of skills for a system, in a way that masks the system’s own generalization power. We then articulate a new formal definition of intelligence based on Algorithmic Information Theory, describing intelligence as skill-acquisition efficiency and highlighting the concepts of scope, generalization difficulty, priors, and experience, as critical pieces to be accounted for in characterizing intelligent systems. Using this definition, we propose a set of guidelines for what a general AI benchmark should look like.
Finally, we present a new benchmark closely following these guidelines, the Abstraction and Reasoning Corpus (ARC), built upon an explicit set of priors designed to be as close as possible to innate human priors. We argue that ARC can be used to measure a human-like form of general fluid intelligence and that it enables fair general intelligence comparisons between AI systems and humans.
There is a Lex Fridman podcast #120 interview with Francios Chollet: some notes from podcast: – one of the things discussed is a historical perspective on intelligence and around the 38 min mark the attributes some statements to Prof Minsky – we will ask our colleague Doug R to set this record straight on the Minsky views.
There is also a video podcast on the article from Yannick Kilcher- some notes from video:
(in fact a 3 part series of podcast on the article) – love some of the conclusions in this presentation – intelligence is not being good at a thing – it is good at getting good at a range of things (tasks) – or the idea of intelligence is not achieved by ‘buying a skill’ either by coding in the intelligence or by data (GPT 3 might fall here) – “solving tasks via experience and priors (what coded in at inception) has nothing to do with intelligence”
There is also a Machine learning street talk podcast episode #51 on the article.
Weekly QuEST Discussion Topics, 25 June
QuEST 25 June 2021
Welcome back – appreciate you coming back to the QuEST meetings. Although our goal is consciousness – it is our position that it creating intelligent systems will require understanding the impact of the conscious representation in enabling flexible cognition and that results in more intelligent solutions. This week we want to hit the paper by Francios Chollet from Google:
On the Measure of Intelligence
Franc¸ois Chollet _
Google, Inc.
November 5, 2019
arXiv:1911.01547v2 [cs.AI] 25 Nov 2019
To make deliberate progress towards more intelligent and more human-like artificial systems, we need to be following an appropriate feedback signal: we need to be able to define and evaluate intelligence in a way that enables comparisons between two systems, as well as comparisons with humans. Over the past hundred years, there has been an abundance of attempts to define and measure intelligence, across both the fields of psychology and AI. We summarize and critically assess these definitions and evaluation approaches, while making apparent the two historical conceptions of intelligence that have implicitly guided them. We note that in practice, the contemporary AI community still gravitates towards benchmarking intelligence by comparing the skill exhibited by AIs and humans at specific tasks, such as board games and video games. We argue that solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience: unlimited priors or unlimited training data allow experimenters to “buy” arbitrary levels of skills for a system, in a way that masks the system’s own generalization power. We then articulate a new formal definition of intelligence based on Algorithmic Information Theory, describing intelligence as skill-acquisition efficiency and highlighting the concepts of scope, generalization difficulty, priors, and experience, as critical pieces to be accounted for in characterizing intelligent systems. Using this definition, we propose a set of guidelines for what a general AI benchmark should look like.
Finally, we present a new benchmark closely following these guidelines, the Abstraction and Reasoning Corpus (ARC), built upon an explicit set of priors designed to be as close as possible to innate human priors. We argue that ARC can be used to measure a human-like form of general fluid intelligence and that it enables fair general intelligence comparisons between AI systems and humans.
There is a Lex Fridman podcast #120 interview with Francios Chollet: some notes from podcast: – one of the things discussed is a historical perspective on intelligence and around the 38 min mark the attributes some statements to Prof Minsky – we will ask our colleague Doug R to set this record straight on the Minsky views.
There is also a video podcast on the article from Yannick Kilcher- some notes from video:
(in fact a 3 part series of podcast on the article) – love some of the conclusions in this presentation – intelligence is not being good at a thing – it is good at getting good at a range of things (tasks) – or the idea of intelligence is not achieved by ‘buying a skill’ either by coding in the intelligence or by data (GPT 3 might fall here) – “solving tasks via experience and priors (what coded in at inception) has nothing to do with intelligence”
There is also a Machine learning street talk podcast episode #51 on the article. Some notes from MLST: ‘need to optimize for generalizability itself’ – generalization = the ability to mine previous experience to make sense of future novel situations, intelligence = generalization power = sensitivity to abstract analogies- generalization is the most important feature of intelligence (ability to handle uncertainty and novelty most important concept of intelligence – task specific performance tells you nothing about intelligence), future of AI will be discrete and continuous (large degree of program synthesis), deep learning good when data is interpolative and has a learnable and smooth manifold that you can traverse via your dense sampling of its surface (DL is interpolative) , missing in today’s AI is adaptation to unknown unknowns, intelligence requires that your adapt to novelty without the help of the engineer who wrote the system, suggest measure of intelligence = the ability to turn experience into future skill, deep learning is a great fit for continuous problems deep learning is a terrible for discrete reasoning problems (algorithmic reasoning problems), f(x)=x, brittleness works in both directions for DL predicting digits of Pi (can write a program to do it but can’t do with a NN – or trying to train a net to determine if a number is a prime which is not interpolative so not good for DL models, perception / intuition issues good for DL – can do it but NOT a good fit) can in fact train a NN to multiply 3 digit numbers and the system can learn that – but there will be glitches you can’t predict – it will multiply many 3 digit numbers correctly but some it will miss – think of a Fourier approximation to a function – it has Gibbs phenomena – doesn’t fit well to rectangular pulses … an approximation to the algorithm but only does local generalization – can’t learn 5 digit numbers … brittle) and in the rules in a program, believer in type 1 and type 2 thinking, all abstraction is born from analogy – the part that is shared is what is said to be the abstraction – the part you can use in the other application – type 1 thinking is abstraction associated to measuring similarity – distance in geometric space – type 1 abstraction is a template – things within a distance are of that type = perception and intuition – everything is a vector, the other way to do abstraction is discrete – topology based grounded about exact comparison by sub graphs (objects are graphs and structure of graphs and using exact comparison), kaleidoscope hypothesis – tiny piece of information can be experienced widely across the space – intelligence being able to face unknown future given past – kaleidoscope suggest the future is similar to the past in some way, intelligence is all about getting to these abstractions (what we’ve call the vocabulary of thought – the qualia) – if share some commonality (share a subgraph) they come from the same source – intelligence is all about reverse engineering the universe to get back to the same source of abstraction – state of the art for hybrid (type1 and type 2 unified systems) is active field of research – currently discrete search using DL to apply to data sets where large and interpolative, and provides guidance to the discrete search process, most cognition is a combination of the two types of thinking, can embed a discrete program into a smooth manifold can do the opposite also by using a discrete program to do perception – not a great idea to do but can do it – alpha fold looking for discrete graph similarities – dreamcoder excites him as a hybrid solution, if a program can solve the ARC challenge = intelligence – No – arc is flawed, a work in progress – Kaggle challenge in ARC last year and learned a lot – found flaws of ARC, Math in paper requires whitebox to see if it is intelligence issue for some – it is all wrapped around idea of conversion ratio of knowledge to use elsewhere (for other tasks) – framework to help us AI think more in different ways – ARC challenge is flawed and being iterated fact the results are only 20% accurate is strange – is it because it is hard or are people not working on it
Weekly QuEST Discussion Topics, 11 and 18 June
QuEST 11 and 18 June
There will not be any QuEST meetings on these two dates – but we will still post relevant links for reading. The first posting is related to a sequence of emails that have been circulating in the group associated with a recent podcast ‘Machine Learning Street talk Episode #54’.
The interview is with Gary Marcus / Luis Lamb. They speak to several issues we’ve discussed – what might the third wave look like – early in the discussion (it is over 2 hours long) – they speak to recent articles by them:
The next decade of AI – Gary Marcus – https://arxiv.org/abs/2002.06177
The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence Gary Marcus Robust AI 17 February 2020
Neurosymbolic ai the 3rd wave – https://arxiv.org/pdf/2012.05876.pdf
Neurosymbolic AI: The 3rd Wave Artur d’Avila Garcez1 and Lu´ıs C. Lamb2 1 City, University of London, UK a.garcez@city.ac.uk 2 Federal University of Rio Grande do Sul, Brazil luislamb@acm.org December, 2020
We suggest both articles are good to read – also about 20 min into the podcast the interviewers provide their definitions for some terms we’ve discussed in detail. I provide below for your consideration. I think they missed some of the conclusions we’ve reached but I found the discussion interesting.
6 most important words intension, extension, reasoning, knowledge, semantics and understanding
Intension – internal structure of an object – ‘the tutor of alexander the great was Aristotle, some other statement whose answer is Aristotle (best student of Plato) – intension is about the understanding the deeper reality where extension is extending without real understanding of the reality
Reasoning – is the act of deriving new knowledge from prior knowledge, given new information – using axioms and rules – ‘my trainer has been milking me’ – reasoning you would say they are taking my money
Knowledge – is a justified true belief – true means it is a fact, justified means it has been established and proved – gold standard
Semantics – is about the interpretation of or mapping to an inner structure to assign meaning (even for pictures not just in linguistics)
Meaning of an utterance – constructed from building materials – infinite – Chomsky universal grammar – everything in the structure – syntax – same as the intension
Understanding is successfully ascertaining, meaning by reconstruction the intension – that inner structure from the syntax / structure we had – if we can describe and reapply the knowledge we gleaned –
Weekly QuEST Discussion Topics, 4 June
QuEST June 4 2021
This will be our last meeting for a couple of weeks – we will meet this Friday 4 June then the next meeting will be on Friday 25 June 2021. This week we want to pick up where our colleagues Kevin and Ancient Mike left off last week. They exposed us to experiments like those developed by George Sperling where a subject can be exposed to an array of characters and then asked them to report what they saw. Sperling found it interesting that a person could only report about 4 items from the array but then it was noted that by directing their attention to a particular part of the display they (for example by a high / medium or low tone) could report those 3-4 items from anywhere in the array. This suggest that the accessing of the information and in the words of Ned Block the conceptualizing of the information forces the sparse representation and thus restricts what can be recalled. Ned Block suggest you do consciously experience the whole array but the sparse cognition representation reduces what you can report as the act of conceptualizing the figures is the low bandwidth / sparse part. There are those like in our previous discussions, Dahanne, who don’t agree with this view. We want to hit the Cleeremans material on computational correlates of consciousness and during that discuss the Ned Block distinct way to speak about consciousness – view of Access – Phenomenal – Monitoring – Self consciousness
Quest Group's Blog 