[ CogSci Summaries home | UP | email ]
http://www.cc.gatech.edu/~jimmyd/summaries/

---

Granger, R. (2006). Engines of the brain: The computational instruction set of human cognition. AI Magazine: 27(2), 15-33.

@Article{Granger2006,
  author = 	 {Granger, Richard},
  title = 	 {Engines of the brain: The computational instruction set of human cognition},
  journal = 	 {AI Magazine},
  year = 	 {2006},
  volume = 	 {27},
  number =  	 {2},
  pages =    	 {15-33},
}

Author of the summary: Gwen Gillingham 2007, gwen.gillingham@gmail.com **DISCLAIMER: This was done as a class assignment before I started gradschool.**

Cite this paper for:

• using neuroscience info to inform AI design
• algorithms derived from neurology can have unforeseen characteristics: hierarchical structure, embedded sequences, hash coding [p15]
• anatomical system level layout of neural circuitry constitutes the "instruction set" of the brain, from which major cognitive operations are assembled
• Auditory processing is a "subroutine" of visual processing
• although primary coritcal regions perform specialized processing, subsequent regions treat all inputs the same, regardless of modality of origin [p25]
• human brains are not optimized for any particular processing task, but use coopted circuits to perform novel tasks
• the primitive operations of the brain are clustering, sequencing, hash coding, compression time dilation and reinforcement learning which can be nested

Argument: AI's are underconstrained because they are not informed enough by neurology. The same behaviours can be output by a variety of different algorithms; we need to focus on those functions we can derive from brain anatomy. Functions emerge from bottom-up analysis: hierarchical structure, embedded structures, hash coding. Anatomical layout establishes how individual operators are composed into larger routines = "the 'instruction set' of the brain, constitut[ing] the basic mental procedures from which all major behavioural and cognitive operations are assembled."

Organization of the Human Brain:

Sensory input: posterior cortext receives input from diencephalic thalamic nuclei

Motor output: interactions between anterior cortex and the striatal complex or basal ganglia

Mammalian brains scale well. Those aspects of brains that grow disproportionately large as brain-to-body-size ratio increases are likely to be very important: they are the "key to human intelligence." Sites that are candidates for this:

The connection pathways between anterior and posterior cortex (fasciculi) grow large

Output pathways from striatal complex change relative size: reccurrent pathways through thalamus increase relative to descending motor pathway

Descending output from anterior cortex to brainstem motor systems gets bigger.

Consequences: for instance, if you damage a mouse's cortical motor areas, it doesn't have much effect, because it uses more brain-stem regions for this function; on the other hand, damage a human's cortical motor area and the result is paralysis. Reason: the striatal region has become a "subroutine" that humans' motor cortex can call.

Striatal Complex:

Basal ganglia/striatal complex is the primary brain system in reptiles, second-largest brain structure in humans. Modules within the basal ganglia are connected through excitatory (activating), modulatory (strength-changing) and inhibitory (suppressing) neuro-transmitter pathways.

There are four parts of this system:

Cortex-Matrisomes (Action): Has two sub-pathways; 1 (MSN2) is inhibitory, the other (MSN1) disinhibitory. They can be thought of as "stop" and "go" paths. Yield complex combinations of motor responses (stand, walk, throw) or cortical thought responses.

Cortex-Striosomes (Evaluation): "triggers what can be thought of as an "evaluation" signal corresponding to the "expected" reward of a given action." They gradually adapt their connection until they predict the correct reward

Substantia Nigra Pars Compacta (SNc) Feedback - Matrisomes and Striosomes (Learning): SNc Receives input from environment conveying

"good" or "bad" result state measurement info. Compares this info to striosomes' "expectation" of the outcome. SNc increases or decreases strength of connections from cortext to striosomes. Connection strength varies directly with probability to select an action, size of future "expectation" of value of given action. System adjusts selection of actions based on experienced outcomes.

Tonically Active Cholinergic Neurons (TANs) - Matrisomes (Exploration): TANs receive inhibitory inputs from striosomes and provide modulatory input to matrisomes. For highly negative expected rewards, the connections' modulatory strength is small, which leads to high variability in selected action. For actions with strongly positive expected reward, TANs have a high modulatory effect, leading to low variability in action selection.

In sum, the striatal complex is "an adaptive controller, beginning with preset responses to inputs, tracking the outcomesw of those responses, and altering behaviour to continually improve those outcomes, as in reinforcement learning algorithms." [p 19]

Thalamocortical System:

2 parallel circuit types:

Core: peripheral inputs activate thalamic core cells, which participate in topographic activation of cortical layers [p 19]; Other layers then activated vertically: middle then superficial then deep layers; deep layers then return to the originating core thalamic nucleus (directly and through an inhibitory intermediary, the nucleus reticularis)

Matrix - little or no peripheral sensory input; driven only by corticothalamic feedback. Matrix thalamus provides further inputs to the cortex. Matrix input uses both feedforward and feedback pathways between cortex and thalamus, unlike Core input. Matrix is diffuse, Core retains brain topography. Within both these systems, excitatory potentials occur in brief bursts (15-20 msec); whereas inhibitory potentials last much longer (80-150 msec) Facts about two systems in thalamocortical system lead to different algorithms:

1. Core Loop: Superficial cells that initially respond to a particular input pattern become increasingly responsive to that input and similar inputs = clustering or categorization. Superficial layer responses activate deep layers, retaining topography; is then sent to the nucleus reticularis, which then inhibits the core thalamic nucleus. "Thus it is hypothesized that the predominant component of the next input to cortex is only the uninhibited remainder of the input, whereupon the same operations as before are performed. Thus the second cortical response will consist of a quite distinct set of neurons from the initial response, since many of the input components giving rise to that first response are now inhibited." This gets us category -> subcategory -> subsubcategory activation of the cortex.

Simplified Thalamocortical Core Algorithm:

for input X
	for C &element; win(X,W)
		W <= W1 + k(X-C)
	end_for
X <= X- mean(win(X,W))
end_for
where
	X = input activity pattern (vector); W = layer I synaptic weight matrix;
	C = responding superficial layer cells (col vector); k = learning rate parameter
	win(X,W) = column vector in W most responsive to X before lateral inhibition [p 20]

2. Thalamocortical "Matrix" Circuits:

Diffuse projections from layer V to matrix nuclei, and from matrix nuclei back to cortex sparsify and orthogonalize their inputs. Structural relationships that may obtain among inputs are not retained in resulting projections. The effect of this layer is to "chain" elements in the input sequence through "links" created due to coincident layer V activity. "The implicit data structures created are trees in which initial sequence elements branch to their multiple possible continuations." [p 23] This is conceptualized as "scatter storage" or a hash function.

Simplified Thalamocortical Matrix Algorithm:

for input sequence X(L)
	for C &element; TopographicSuperficialResponse(X(L))
		for V(s) &element; NNtResponse (X(L-1))
			Potentiate (V(s))
			NNt(L) = NontopographicDeepResponse(V)
		end_for
	end_for
end_for
	where L = length of input sequence;
	C = columnar modules activated at step X(L);
	V(s) = synaptic vector of responding layer V cell,
	NNt(L) = response of nonspecific thalamic nucleus to feedback from layer V. [p 22]

Combined Telencephalic Algorithm Operation and the Emergence of Complex Specializations:

The thalamo-cortico-stiatal system gives rise to reinforcement learning of similarity based clusters, and brief sequences (and by extension, brief sequences of similarity based clusters.) The output of any region becomes input to another region, and each region in the thalamo-cortico-striatal pathway performs the same sort of analysis on inputs, "generating learned nested sequences of clusters of sequences of clusters." [p 23]

Auditory processing: statistical filters note a sequence of features and choose a best partial match from prior input to determine what a word is. Visual Image Processing: Beyond specialized cortical areas, auditory and visual processing are similar neurologically, at the thalamocortical circuitry level. "It is here hypothesized that although primary cortical regions perform specialized processing, subsequent cortical regions treat all inputs teh same, regardless of modality of origin." Curve-and-line information from primary visual cortex is later analysed similarly to auditory information, statistically segmented into objects etc using partial best match. Auditory parsing (because it is only left-right, time delimited) can be thought of as a subroutine of visual processing (which is more-dimensional). Though all "regions" are identical in structure, they receive different inputs, and after exposure to many inputs, regional specializations of function arise due to lateral competition among areas [pp 23-24].

Hierarchical Grammatical Structure: "It is notable that the emergent data structure of the thalamo-cortico-striatal model, nested sequences of clusters, is a superset of the structures that constitute formal grammars, that is, ordered sequences of 'proto-grammatical' elements, such that each element represents either a category (in this case a cluster), or expands to another such element (nesting), just as rewrite rules establish new relations among grammatical elements." [p 26] Data structures can grow new rules with more loops added to the thalamo-cortico-striatal system. Higher brain-to-body-size-ratio = more loops, more complex grammars. "Growth of grammars need not be linear; grammars have the property of exhibiting apparently new behaviours due to the addition of just a few rules." [p 27] Processing of linguistic inputs is the same as processing of other sensory inputs, but has become more complex with our brain size's growth.

---

Back to the Cognitive Science Summaries homepage
Cognitive Science Summaries Webmaster:

JimDavies (jim@jimdavies.org)