The concept of ‘abduction’ came up in twitter discussion and people asked what it means. This may be a good occasion for a short thread on a recent paper about the challenges in characterising ‘abduction proper’. 1/n docs.lib.purdue.edu/cgi/viewconten…
“The capacity to formulate an explanation for a given observation is called abductive inference (Peirce, 1974).”
This capacity is vital for us to make sense of our every world, but also for scientists to construct scientific explanations for observed phenomena. 2/n
“Often, this capacity is characterized as an inference to the best explanation (IBE)—selecting the “best” explanation from a set of candidate hypotheses ... “
In science this may correspond e.g. to computing which of a few candidate models is most probable given the data. 3/n
“However, accounts of IBE assume that the set of candidate hypotheses is given, & therefore they do not explain the origin of the set of candidate hypotheses, also known as abduction proper”
I.o.w., abduction proper is the creative part. How are novel explanations generated? 4/n
While IBE has formal characterisations (eg define “best” in terms of probability or coherence), we currently lack satisfactory characterisations of abduction proper.
The challenge is how to unify 7 necessary properties in a single characterisation. 5/n
The 7 properties are: (1) isotropy, (2) open-endedness, (3) novelty, (4) groundedness, (5) sensibility, (6) psychological realism, and
... wait for it ... (7) computational tractability. 6/n
"We propose unification [of these properties] can be achieved by viewing the origin of hypotheses as a process of deep analogical inference (...) deep analogical inference allows many consecutive & branching analogical inferences that lead to sets of candidate hypotheses" 7/n
*Intermezzo* -- In Supplementary materials we illustrate the idea of deep analogical inference and its use in (analogical) abduction proper in a case study using the Tacit Communication game (see here a demo of the game ). 8/n
"The computational-level theory of analogical abduction proper unifies 6 out of the 7 necessary properties of abduction proper under one theory."
Which of the properties does it not yet unify with the others do you think? Wanna guess? 9/n
As you may have guessed, the characterization unifies the first 6 properties, but is so far unclear how it can be computationally tractable.
I.o.w. it remains a puzzle how abduction proper can be computed without astronomical resource demands that "blow our minds/brains"
This is in a sense neither surprising (IBE is known to generally be intractable as well, as are many other capacities); nor fatal, since we have tools for dealing w/ intractability.
*Another intermezzo* for an Intro to these tools, check out our new book
The "short" thread turned out a bit longer in the end, but I hope it was interesting & useful.
Two co-authors on the paper () who are on twitter --> @MarkBlokpoel (1st author) & @Pim_Haselager. And co-author on the linked book --> @JohanKwisthout 14/14docs.lib.purdue.edu/jps/vol11/iss1…
• • •
Missing some Tweet in this thread? You can try to
force a refresh
Some ppl may be unclear on relevant formal details of AI-by-Learning, so let me clarify a few things in this thread.
1/n
The AI-by-Learning problem that we formalise grants our hypothetical AI engineer all kinds of idealisations and simplifications, leaving them with a *simpler* problem than IRL. Our intractability result is thus a *lower* bound estimate on the real-world complexity.
⚡️Very excited to share our new preprint "Reclaiming AI as a theoretical tool for cognitive science", by @o_guest @fedeadolfi #ronalddehaan #antoninakolokolova & #patriciarich and myself Highlights/summary in thread 🧵👇 1/npsyarxiv.com/4cbuv
"The term ‘Artificial Intelligence’ (AI) means many things to many people (see Table 1) (...). One meaning of ‘AI’ that seems often forgotten these days is one that played a crucial role in the birth of cognitive science as an interdiscipline in the 1970s and ’80s." 2/n
"This view of ‘AI’ as a research field overlapping with psychology sees computational systems as theoretical tools (...) Accordingly, AI is one of the cognitive sciences (Figure 1), and for decades there was a close dialogue between the fields of AI and cognitive psychology." 3/n
“The pressure to “teach with” generative tools has continued to mount, driven partly by technology companies that have long perceived education as a lucrative market.” 1/n 🧵
“In producing these much-hyped commercial tools, these companies neither focused on education nor consulted with educators or their students. Not only designed without consideration of educational goals, practices, or principles, these models emerge from a technocratic landscape”
“… that often denigrates higher education, imagines teaching to be a largely automatable task, conceives human learning as the acquisition of monetizable skills, and regards both students and teachers as founts of free training data.” 3/n
2. The Goal: To formally characterize cognitive functioning (the ‘what’ of cognition; see Marr’s levels).
3. The Method: hypotheses for cognitive functions can be derived through so-called ‘rational analysis’: Given an agent’s goals and environment of adaptation, what is the optimal function (i.e., derive ‘what’ from ‘why’.)