NEURAL ENGINEERING Phil/Psych 256 Chris Eliasmith
Metaphors in cognitive science None of these should be surprising: Symbolicism: mind as computer Connectionism: mind as brain-like Dynamicism: mind as Watt governor Note that all three positions hold materialism (i.e. that the mind is the brain). Connectionists also hold that the best cognitive/psychological descriptions share processing structure (and representational structure) with brains. The others deny this.
Different roles of metaphor Explanation: relate some unknown domain (the target) to an already familiar domain (the source). as explanations such metaphors only 'go so far'. They help illuminate some elements of the target but usually differ in significant ways as well (e.g. atom as solar system) Discovery: metaphors have been cited as the source of inspiration for understanding some unknown domain. the role of the metaphor isn't just explanatory, it's also the original source of a new idea. Theoretical: justify importing theories into the target domain from the source domain. the mind as computer metaphor might be the best example: it justified the introduction of Turing machines and logic-like structuers to psychological theorizing. The paper takes issue with the theoretical use (via a metaphor): Just as theoretical metaphors in the case of light limited insight, so theoretical metaphors in the case of cognition will limit insights.
A (really) brief history of cog sci Summary: Behaviourists said 'don't look in the black box' and cyberneticists didn't. The cognitive revolutionaries, inspired by the mind as computer metaphor, said 'you have to look inside the black box' Connectionists said 'the black box isn't a Von Neumann machine, and the representations are distributed' Dynamicists said 'there aren't representations or computations in there -- it's a physical, continuous, temporal system'
A (really) brief history (cont.) Suggestion: the focus on computers and computational theory resulted in an oversight. Namely, modern control theory (which rectified the important problems of classical control theory) was ignored as a way of understanding minds Modern control theory seems amenable with most of the positive suggestions from other paradigms: it describes physical, continuous, temporal systems by looking inside the system it does so in an architecturally and representationally independent way But, since the architecture is there for study and representations are explanatorily useful, more work needs to be done to use MCT to describe cognitive systems
R&D Theory Three principles: 1. Neural representations are defined by nonlinear encoding (seen in neuron tuning curves) and weighted linear decoding 2. Transformations (computations) of these representations are determined by alternately weighted linear decoding 3. Neural representations are (control theoretic) state variables in specifications of the systems dynamics.
R&D Theory (cont.) Principle 1: any complex mathematical object (functions, vectors, what have you) can be represented in neurons in a well-characterized way this should be sufficient for representing anything: sounds, pictures, concepts, uncertainties, motor commands, etc. mathematical representation of things like sounds and pictures (i.e. analog or digital encodings on tapes/cds) is quite common-place.
R&D Theory (cont.) Principle 2: computing any function (transformation) of these representations can be done in pretty much the same way the representation is defined rather than trying to extract the thing represented by the representation, the same information is used to extract some variation of the thing represented
R&D Theory (cont.) Principle 3: the most important because it unifies the previous two the representations and transformations can be used to underwrite a dynamic description of the systems function the description consists of differential equations describing the updating over time of the relevant representations control theory provides a means of analyzing and synthesizing differential equations in a systematic way.
Simple example: Working memory Typical working memory effect (LIP)
Simple example: Integration A=0 B=1
Implementation
Path Integration (another example) Path integration is the ability of rats to return directly to an initial location using only idiothetic cues. It seems to be performed in the subiculum (Sharp 1997) since hippocampal lesions don t stop it (Alyan et al. 1997) Start X
Path Integration Representation and dynamics in action insert movie
R&D vs. Symbolicism Similarities: quantifies over representations and computations looks 'inside the black box' explains cognitive phenomena (structure-sensitive processing with HRRs) supposed to be a unified general theory Differences: integrates time extensively neurally plausible should explain non-cognitive phenomena (e.g. motor behavior, perceptual processes, emotion)
R&D vs. Connectionism Similarities: brain-inspired architecture (robust to damage, etc.) distributed representations Differences: increased neural plausibility doesn't rely on learning (learning can be included) provides a method for systemmatically integrating top-down and bottom-up data regarding a phenomena (e.g. working memory and tuning curves in LIP; Kalman filter for visual processing and V1 responses and connectivity) helps identify the computed function given the results of learning
R&D vs. Dynamicism Similarities: essentially includes time high-level dynamic systems theory (differential equation) descriptions of phenomena Differences: The mapping between high-level descriptions (lumped parameters) and the neural implementation is explicit Limitations of classical control are explicitly avoided Inclusion of representations as a useful explanatory tool for understanding cognitive behaviour
Discussion It's a CRUM-y theory The approach is an extension of the 'standard' CRUM, but not completely different. It suggests that functionalism needs to be/can be reconceived to include time. Limitations: a demonstration across a wide variety of (especially cognitive) systems that the approach is successful there are nothing like the number of R&D models as connectionist or symbolicist ones. But the models continue to come, and continue to prove novel and useful (see http://compneuro.uwaterloo.ca/)