Study Notes – Embedded Cognition
CSC447 – Spring 2009
Overview:
Randall D. Beer in Dynamical Systems and Embedded Cognition makes a strong case for focusing on cognition in AI research as well as giving a good historical overview of this approach. He presents a conceptual paradigm he calls the situated, embodied, dynamical (SED) framework and breaks this down as follows:
1. Situated Activity
a. Concrete action
b. Situatedness
c. Interactionism
2. Embodiment
a. Physical embodiment
b. Biological embodiment
c. Conceptual embodiment
3. Dynamics
a. Dynamical systems theory (DST)
b. The dynamical framework
c. The dynamical hypothesis
Let’s briefly look at his main points under each, adding some related information from other sources.
Situated Activity: Precursors::
Philosophical roots are in phenomenology, especially Martin Heidegger, as promoted by Hubert Dreyfus, an early critic of AI and author of What Computers Can’t Do, What Computers Still Can’t Do, and Alchemy and Artificial Intelligence. The latter is an attack on the work of Newell and Simon criticizing the idea “that intelligence consists of the manipulation of physical symbols according to formal rules” [http://en.wikipedia.org/wiki/Hubert_Dreyfus].
A key insight of Heidegger – distinction between zhanden and vorhanden. A hammer is zuhanden if during use it becomes an extension of the arm.
Important precursor – J.J. Gibson’s ecological psychology based on study of vision in WW2 pilots. Emphasis – structure in organism’s environment and relation of organism/environment to perception. Rejection of cognitivism – “a theoretical approach in understanding the mind using quantitative, positivist and scientific methods, that describes mental functions as information processing models.” [http://en.wikipedia.org/wiki/Cognitivism_(psychology)]. {Other critics of cognitivism include Penrose, on basis of Goedel’s Incompleteness Theorem and The Halting Problem, John Searle, on the basis that syntax is not semantics (Chinese Room argument) and that syntax is not physics, and proponents of embodied cognition, including Rodney Brooks}.
Gibson’s most important contribution – the notion of affordances, possibilities for action presented by the environment, e.g., design of hammer comprises affordance for nailing. Note: affordance is relative to capabilities of particular organism.
Third influence – ethnomethodological approaches to social science; anthropologists Suchman (man-machine interaction) and Hutchins (naval vessel navigation team) that activity is situated, culturally constituted and dependent on unfolding situation. Action not simply execution of a plan but contingent upon circumstances as they unfold.
Early demonstrations that rich behavior can come from simple mechanisms interacting with complex environment – Grey Walter’s tortoises & Braitenburg’s vehicles. {Also Herb Simon’s ant, cellular automata and Reynold’s Boids}
Situated Activity: within AI::
Impetus was reaction against traditional planning view of action {{examples – General Problem Solver (GPS) of Newell & Simon which influenced Stanford Research Institute Problem Solver (STRIPS) which was “brains” for Shakey the robot. These techniques encountered computational complexity and knowledge representation issues, including the Frame problem}}. Agre & Chapman, with Pengi, showed capabilities of simple internal machinery. Rosenschein & Kaelbling compiled agent’s specification into simple machinery.
Rodney Brooks rejected vertically layered sense-model-plan-act cycle for autonomous robots and proposed a horizontally layered subsumption architecture as a control system. {{Basic and simple actions were placed at the lower layers and thoroughly tested. Higher level functions were placed at higher layers which were given the power to subsume the lower layers. In addition he rejected the notion of extensive knowledge representation for agents saying that the world is its own best representation and claiming to produce intelligence without representation. This approach did solve the frame problem, but his claims were somewhat overstated}}.
Cliff and Beer developed a computational neuroethology model of hoverfly and cockroach.
Situated Activity = behaviorism?
Controversial question arises – how fundamental is environmental situation to agent’s behavior? Is it behaviorism? Approach does emphasize concrete behavior over abstract reasoning; but does not reject abstract reasoning, though relegates it to supporting role. Much work is based in reactive architectures which are reminiscent of stimulus-response paradigm of behaviorism, which has well-known limitations. But emphasis on situated activity does not require purely reactive architectures and there can be a role for internal state.
Extended Mind:
A controversial idea. Begins with observation that agent can organize environment to increase problem-solving ability, e.g., recipe, map. This scaffolding offloads cognitive processing into environment. And through language we can coordinate activity. Extended mind advocates argue that cognition is a distributed phenomenon that belongs only to a larger system of agents transcending individuals. They cite stigmergy of social insects.
Embodiment:
Philosophical roots also in phenomenology, especially Merleau-Ponty. Example – how we perceive an object is rooted in kinds of interactions with it our body allows, which is an early precursor of Gibson’s affordances. His work played role in Dreyfus’ critique of computational theories of mind.
Physical Embodiment:
Within AI importance first emphasized by Brooks who held that AI must address problems of real robots in real environments which would expose inadequacy of classical AI techniques.
Milder form of argument for physical embodiment – material properties of body & environment play key role in behavior and for building robots do not need to model physics of world. In radical form argument is that only physically instantiated AI systems are truly intelligent.
Brooks’ critiques along with trends in situated cognition led to extensive work in behavior-based robotics, active perception, embodied cognitive science, autonomous agents and some aspects of artificial life and philosophy of mind.
Biological Embodiment:
One step further – biological facts of organism’s existence are also important. Conditions for remaining alive constrain behavior and cognitive capacities. And developmental history affects behavioral and cognitive architecture.
Thelen & Smith argue for sensorimotor origins of cognition in development, studying walking in infants & Piaget’s A-not-B error. There is push for neurobiological realism in embodied agents and neuroscience is examining role of neuromechanical interactions in production of behavior.
Basic claim is that biological features matter to behavior and cognition. More radical claims – living state itself is fundamental to cognition and to be truly autonomous system must be capable of creating own laws.
Conceptual Embodiment:
Even abstract concepts are grounded in bodily experience and metaphor. Harnad defines symbol grounding problem – how words & mental states get their meaning – and proposed to ground them in sensorimotor signals. Lakoff and Johnson argue that structure of reason is grounded in embodiment and abstract metaphors derive from sensorimotor domains. Examples – understanding is “grasping”; failing to understand is having it go over our heads.