towards which intelligence? cognition as design key for building artificial intelligent systems
TRANSCRIPT
Towards which Intelligence? Cognition as Design Key for building Artificial Intelligent Systems
Antonio Lieto
University of Turin, Dipartimento di Informatica, ItalyICAR - CNR, Palermo, Italy
Symposium “Bridging the Gap between Cognitive Psychology and Artificial Intelligence”
ESCOP 2015, Paphos, Cyprus, 19th September 2015
Outline
• Which “Intelligence” in Artificial Intelligence (AI)?
• Cognitive AI: methodological and technical considerations.
• A case study (Time permitting !): System dealing with Common Sense Reasoning (Conceptual Categorization, paper presented at IJCAI 2015).
2
Which“Intelligence” in AI? (partial overview)
Early AI
Cognitive Inspiration for the Design of “Intelligent Systems”
M. Minsky
R. Shank
Modern AI
“Intelligence” in terms of optimality of a performance
(narrow tasks)
J. McCarthy
mid‘80s
A. Newell
H. Simon
deep learning
A focus shift in AIP. Langley (2012) “Vision the early days of AI: Understanding and reproducing, in computational systems, the full range of intelligent behavior observed in humans”.
Why this view has been abandoned?
- Emphasis on quantitative results and metrics of performance: (“machine intelligence”: achieving results and optimize them !)
- Commercial success of narrow applications etc.
Nowadays it is regaining attention since “The gap between natural and artificial systems is still enormous” (A. Sloman, AIC 2014)
4
Cognitive AI
Attention to the heuristics-based solutions adopted by humans (e.g. Gigerenzer & Todd, 1999) for combinatorial problems (“bounded rationality heuristics”).
Heuristics realize/implement some cognitive functions and are responsible of the macroscopic external behaviour of an agent.
5
2 Perspectives
The “Cognitive Systems” one (Brachman and Lemnios, 2002), referring to the discipline that: “designs, constructs, and studies computational artifacts that exhibit the full range of human intelligence”.
The“nouvelle AI” (e.g. Parallel Distributed Processing (Rumhelarth and McLelland, 1986): based on bio-plausibility modelling techniques allowing the functional reproduction of heuristics in artificial systems (neglecting the physical and chemical details).
6
Cognitive Systems Paradigm
Langley (2012):
• Focus on high level human cognitive functions
• Assuming structured representations (physical symbol system, Simon and Newell, 1976)
• Architectural Perspective (integration and interaction of all cognitive functions)
• Inspiration from human cognition
“Nouvelle” AI paradigm
• Focus not only on high level cognitive functions (but also on perception, vision etc.)
• Assuming unstructured representation (e.g. such as neural networks etc.) and also integration with symbolic approaches.
• System perspective (not necessary to consider a whole architectural perspective).
• Bio-inspired computing, bottom-up approach (for learning etc.). 8
A Matter of Levels
• Both the “classical” and “nouvelle” approach can realize, in principle, “cognitive artificial systems” or “artificial models of cognition” provided that their models operate at the “right” level of description.
• A debated problem in AI and Cognitive Science regards the legitimate level of descriptions of such models (and therefore their explanatory power).
Functionalist vs Structuralist Models 9
Functionalism • Functionalism (introduced by Putnam) postulates a weak equivalence
between cognitive processes and AI procedures.
• AI procedures have the functional role (“function as”) human cognitive procedures.
• Multiple realizability (functions can be implemented in different ways).
• Equivalence on the functional macroscopic properties of a given intelligent behaviour (based on the same input-output specification).
• This should produce predictive models (given an input and a set of procedures functionally equivalent to what is performed by cognitive processes then one can predict a given output).
10
Problem with Functionalism
• If the equivalence is so weak it is not possible to interpret the results of a system (e.g. interpretation of the system failures…)
• A pure functionalist model (posed without structural constraints) is a black box where a predictive model with the same output of a cognitive process can be obtained with no explanatory power.
• An analogy: A Bird and a Jet can fly but a jet is not a good explanatory model of a bird since its flights mechanisms are different from the mechanism of bird.
• Purely functional models/systems (like IBM Watson) are not “computational models of cognition” 11
Structuralism
• Strong equivalence between cognitive processes and AI procedures (Milkowski, 2013).
• Focus not only on the functional organization of the processes but also on the human-likeliness of a model (bio-psychological plausibility).
• Wiener’s paradox: “The best material model of a cat is another or possibly the same cat”
12
A Design ProblemPylyshyn (1979): “if we do not formulate any restriction about a model we obtain the functionalism of a Turing machine. If we apply all the possible restrictions we reproduce a whole human being”
• Need for looking at a descriptive level on which to enforce the constraints in order to carry out a human like computation.
• A design perspective: between the explanatory level of functionalism (based on the macroscopic stimulus-response relationship) and the mycroscopic one of fully structured models (reductionist materialism) we have, in the middle, a lot of possible structural models. 13
“Cognition” as Design Constraint
• We need a function-structure coupling for the design of cognitive artificial system.
• The interpretations of the experimental results coming from cognitive psychology indicate us the algorithm procedures (the heuristics/design constraints) that we can implement in our system in a functional-structural way.
• I.e. the implementation can be done in different ways (multiple realizability account of the functionalism) but the model needs to be constrained to the target system (needs to be structurally valid). 14
Case Study
A Common Sense Reasoning System
IJCAI’15 Paper
Antonio Lieto, Daniele P. Radicioni and Valentina Rho, “A Common-Sense Conceptual Categorization System Integrating Heterogeneous Proxytypes and the Dual Process of Reasoning”. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), Buenos Aires, July 2015, pp. 875-881. AAAI press.
15
System Assumptions– Representation:– conceptual structures as heterogeneous proxytypes (Lieto 2014);
compliance with the computational frameworks of cognitive architectures (General Architectures for Intelligence).
– Reasoning: – 2 types of common sense inference (based on prototypes and
exemplars). – Dual process reasoning (Common sense + Standard
categorization).
– Integration into the ACT-R cognitive architecture (Anderson et al. 2004).
Prototypes and Prototypical Reasoning• Categories based on prototypes (Rosh,
1975)• New items are compared to the prototype
atypical
typical
P
Exemplars and Exemplar-based Reasoning• Categories as composed by a list of exemplars. New
percepts are compared to known exemplars (not to Prototypes).
Heterogeneous Hypothesis
The different proposals that have been advanced can be grouped in three main classes: a) fuzzy approaches, b) probabilistic and Bayesan approaches, c) approaches based on non-monotonic formalisms.
Different representational structures have different accessing procedures (reasoning) to their content.
Prototypes, Exemplars and other conceptual representations can co-exists and be activated in different contexts (Malt 1989).
System Conceptual Structure
20
for a given concept can be represented by adopting differ-ent computational frameworks: i) from a symbolic perspec-tive, prototypical representations can be encoded in termsof frames [Minsky, 1975] or semantic networks [Quillian,1968]; ii) from a conceptual space perspective, prototypes canbe geometrically represented as centroids of a convex region(more on this aspect later); iii) from a sub-symbolic perspec-tive, the prototypical knowledge concerning a concept can, onthe other hand, be represented as reinforced patterns of con-nections in Artificial Neural Networks (ANNs). Similarly,for the exemplars-based body of knowledge, both symbolicand conceptual space representations can be used, as well asthe sub-symbolic paradigm. In particular, exemplars can berepresented as instances of a concept in symbolic systems,as points in a geometrical conceptual space, or as a partic-ular (local) pattern of activation in a ANN. Finally, also forthe classical body of knowledge it is –at least in principle–,possible to use the same frameworks. However, this seemsto be a case where symbolic and conceptual levels are moreappropriate w.r.t. the sub-symbolic one.
Summing up, all the different types of conceptual repre-sentations can be implemented in cognitive artificial systemsand architectures. In addition, different computational mech-anisms for “proxyfying” conceptual representations can beapplied. In the next Section we illustrate and discuss the rep-resentational levels and the associated computational frame-works we adopted for each type of body of knowledge.
3 The DUAL-PECCS SystemAs mentioned, the DUAL-PECCS relies on the heteroge-neous proxytypes approach and on the dual process theory.It is equipped with a hybrid knowledge base composed ofheterogeneous representations of the same conceptual enti-ties: that is, the hybrid knowledge base includes prototypes,exemplars and classical representations for the same concept.Both prototypes and exemplars are represented at the con-ceptual level (see Section 3.1), while classical information isrepresented through standard symbolic formalisms (i.e., bymeans of a formal ontology).
The retrieval of such representations is driven by differentprocess types. In particular, prototype and exemplar-based re-trieval is based on a fast and approximate kind of categoriza-tion, and benefits from common-sense information associatedto concepts. On the other hand, the retrieval of the classicalrepresentation of concepts is featured by explicit rule follow-ing, and makes no use of common-sense information. Thesetwo differing categorization strategies have been widely stud-ied in psychology of reasoning in the frame of the dual pro-cess theory, that postulates the co-existence of two differ-ent types of cognitive systems [Evans and Frankish, 2009;Kahneman, 2011]. The systems of the first type (type 1) arephylogenetically older, unconscious, automatic, associative,parallel and fast. The systems of the second type (type 2) aremore recent, conscious, sequential and slow, and featured byexplicit rule following. We assume that both systems can becomposed in their turn by many sub-systems and processes.According to the hypotheses in [Frixione and Lieto, 2012;Frixione and Lieto, 2014], the conceptual representation of
is-a: felinecolor: yellowhasPart: furhasPart: tailhasPart: stripes
...
conceptual space representation
concept Tiger
Kingdom: AnimaliaClass: MammaliaOrder: CarnivoraGenus: PantheraSpecies: P. tigris
prototype of Tiger exemplars of Tiger
white-tigeris-a: felinecolor: whitehasPart: furhasPart: tailhasPart: stripes...
...
ontological representation
classical information
Typicality-based knowledge
Classical knowledge
Hybrid Knowledge Base
Figure 1: Heterogeneous representation of the tiger concept
our system includes two main sorts of components, based onthese two sorts of processes. Type 1 processes have been de-signed to deal with prototypes- and exemplar-based retrieval,while Type 2 processes have been designed to deal with de-ductive inference.
The two sorts of system processes interact (Algorithm 1),since Type 1 processes are executed first and their results arethen refined by Type 2 processes. In the implemented sys-tem the typical representational and reasoning functions areassigned to the System 1 (hereafter S1), which executes pro-cesses of Type 1, and are associated to the Conceptual Spacesframework [Gardenfors, 2000; Gardenfors, 2014]. On theother hand, the classical representational and reasoning func-tions are assigned to the System 2 (hereafter S2) to executeprocesses of Type 2, and are associated to a standard ontolog-ical representation.
Figure 1 shows the heterogeneous representation for theconcept tiger, with prototypical and exemplar-based repre-sentations semantically pointing to the same conceptual en-tity. In this example, the exemplar and prototype-based rep-resentations make use of non classical information. Namely,the prototypical representation grasps information such asthat tigers are wild animals, their fur has yellow and blackstripes, etc.; the exemplar-based representations grasp infor-mation on individuals (such as white-tiger, which is a partic-ular tiger with white fur). Both sorts of representations acti-vate Type 1 processes. On the other hand, the classical bodyof knowledge is filled with necessary and sufficient informa-tion to characterize the concept (representing, for example,the taxonomic information that a tiger is a mammal and acarnivore), and activates Type 2 processes.
In the following we introduce the two representational andreasoning frameworks used in our system, by focusing i) onhow typicality information (including both prototypes and ex-emplars) and their corresponding non monotonic reasoningprocedures can be encoded through conceptual spaces; andii) on how classical information can be naturally encoded interms of formal ontologies.
Dual Process Reasoning
11
Reasoning Harmonization based on the dual process (Stanovitch and West, 2000; Kahnemann 2011).
In human cognition, type 1 processes are executed fast and are not based on logical rules. Then they are checked against more logical deliberative processes (type 2 processes).
Type 1 Processes Type 2 Processes
Automatic Controllable
Parallel, Fast Sequential, Slow
Pragmatic/contextualized Logical/Abstract
Overview
NL Description-The big fish eating plankton
Typical Representations
IE step and mapping
List of Concepts : -Whale 0.1 -Shark 0.5 -…
Output S1 (Prototype or Exemplar)
Check on S2Ontological Repr.
-Whale NOT Fish -Whale Shark OK
Output S2 (CYC)
Output S1 + S2
Whale Whale Shark
Cognitive Architectures
23
Allen Newell (1990) Unified Theory of Cognition
A cognitive architecture (Newell, 1990) implements the invariant structure of the cognitive system.
It captures the underlying commonality between different intelligent agents and provides a framework from which intelligent behavior arises.
The architectural approach emphasizes the role of memory in the cognitive process.
Heterogeneous Proxytypes
The different proposals that have been advanced can be grouped in three main classes: a) fuzzy approaches, b) probabilistic and Bayesan approaches, c) approaches based on non-monotonic formalisms.
Concepts as Heterogeneous Proxytypes (Lieto, 2014).
Stored in our long term memory and, activate in our working memory only specific representations (“go proxy”).
Evaluation (Human-Machine Comparison)
• 122 textual stimuli created by a team of neuropsychologists and tested in categorisation task during neuromaging experiments (FMRI).
• 40 human subjects were requested to categorize such stimuli (and to specify the corresponding representation activated).
• The provided answers are a Gold Standard and represents the expected results.
• 84% of system answers overlapping with human responses. The computational experimentation also provided insights for the refinement of the theory (details in the paper!)
Take home message
• After two decades, Cognitive AI field is gaining a renewed attention.
• Cognitive Artificial Models of AI are proxyies of a target system and have an explanatory power only if they are structurally valid models (realizable in different ways).
• Cognitive Artificial Systems built with this design perspective can have an explanatory role for the theory they implement and the “computational experiment” can provide results useful for refining of retaining theoretical aspects.
Join us at AIC 2015 in Turin ! 28-29 Sept. 2015
27
Thanks !
Contacts:
Symposium “Bridging the Gap between Cognitive Psychology and Artificial Intelligence”
ESCOP 2015, Paphos, Cyprus, 19th September 2015