Eliasmith Descriptions Understanding Computation in Brain 2010

Dec 01, 2012 11:48 AM

Chris advocates control theoretic description as being the best way to describe

brain computation.

3.2  A good description provides a CLEAR DECOMPOSITION

Because of the complexity of neural systems, decomposition is an essential

explanatory strategy (Bechtel and Richardson, 1993). The more specific and effective the

decomposition for explanatory progress, the better the description

3.3  

6  Control theoretic descriptions are good descriptions

6.1  Control descriptions provide simple mappings from data to control theory states

By far the best mechanistic descriptions we currently have of single neural cells is as

non-linear electrical circuits. The circuits are naturally described by non-linear systems

theory, the main mathematical tool of control theory. As a result, control theoretic states

are already widely accepts as the simplest, most powerful descriptions of single neuron

behaviour.

As we compose single cells into larger networks, it is useful to adopt the language of

representation and computation. Eliasmith & Anderson (2003) propose the Neural

Engineering Framework (NEF), a detailed theory of how neural systems can understood

in these terms. I will not review the three central principles of this approach here, but

will note the following: the third principle provides a direct mapping from the single cell

data typically collected by neurophysiologists to control theory. This mapping consists

of a nonlinear encoding, determined directly from the data, and linear decoding that is

optimal and mapped directly to the neurophysiology. This low-level neurophysiological

mapping allows for prediction of single cell and aggregate data. In short, the mapping is

simple between control theoretic states and many kinds of neural data.

These methods have been successfully used for a wide variety of modelsincluding the

barn owl auditory system (Fischer, 2005), the rodent navigation system (Conklin &

Eliasmith, 2005), escape and swimming control in zebrafish (Kuo & Eliasmith, 2005),

working memory systems (Singh & Eliasmith, 2006), the translational vestibular ocular

reflex in monkeys (Eliasmith, Westover, & Anderson, 2002), and context sensitive

linguistic inference (Eliasmith, 2005). This variety suggests the mapping is a useful one

for positing and testing neural mechanisms.

7.1  Comparisons to computational descriptions

It is important to emphasize that, compared to computational descriptions, control

theoretic ones unify our description of phenomena of interest to cognitive scientists. For

instance, working memory and navigation, are typical ‘cognitive’ phenomena. However,

locomotion and reflexes are more typical ‘sensory-motor’ kinds of phenomena. Control

theory applies well to all of these. And, to allay concerns that such descriptions do not

apply to ‘higher’ cognitive phenomena, the BioSLIE model presented in Eliasmith (2005)

is a crucial existence proof of the utility of control theoretic descriptions of linguistic,

inference tasks. More generally, this model demonstrates syntactic generalization, and

makes predictions regarding learning history and response times in a cognitive task (the

Wason card task). The details of this approach are beyond the scope of the present

discussion (see Stewart & Eliasmith (in press) for a description).

SOAR is bad  (Newell).   it distinguishes btwn program and memory;

lack of systemic relation to neural hardware.

7.3  STATISTICAL DESCRIPTIONS

are bad cuz

they do NOT provide decompositions or mechanisms

they are predictive but not explanatory

8  If we want a good description of brain function, we ought to

adopt a control theoretic approach

So, how we ought to describe computation in the brain is not as computational theory

would demand. Instead, if we let pragmatic considerations drive our descriptions (as we

must, otherwise we cannot choose), control theoretic descriptions are most promising for

advancing our understanding of neural systems.

Recall the clarification of this conclusion: it does not rule out the other descriptions as

useful. Rather, it suggests that other descriptions are essentially heuristics for stating the

description in a control theoretic manner. In many cases, there are equivalent

formulations of a given model in these purportedly different approaches. However,

control theoretic descriptions should be primary, and hence the goal of providing good,

useful descriptions of brain function.

Eliasmith Descriptions Understanding Computation in Brain 2010

Attachments