Joanna Bryson
Laboratory for Cognitive Neuroscience and Intelligent Systems
Edinburgh University Level 8, Appleton Tower Edinburgh EH8 9JZ, UK

Will Lowe
Centre for Cognitive Science
Edinburgh University
2, Buccleuch Place

A review of Ballard, Dana H., Hayhoe, Mary M., Pook, Polly K. & Rao, Rajesh P. N. Deictic codes for the embodiment of cognition


Ballard et al. show how minimal state can be made flexible enough for complex cognitive tasks by using deictic pointers, but do so within a specific computational framework. We discuss broader implications in cognition and memory, and provide biological evidence for their theory. We also suggest an alternative account of pointer binding which may better explain their limited number.

Ballard, Hayhoe, Pook and Rao point out in their conclusion (p.38) that deictic coding is intimately connected to theories of intelligence that minimize representational content. As researchers in the field of reactive, behavior based artificial intelligence (Bryson) and modelling human semantic memory (Lowe), we consider Ballard et al's theory a valuable way to reconceptualize intelligence.

Traditional AI considers as its core problem representational redescription from a perception model to an action model. The reactive approach claims to eliminate representation altogether, focussing instead on units of directly coupled perceptions and actions (Brooks 1991). In fact this transfers state from world models to control. The complex representation processing of traditional AI systems is, in a reactive system, precompiled as an intricate and highly customized program structure. However, this transformation appears to result in computational savings. Reactive robots run considerably quicker and more robustly than their traditional counterparts, using computational devices of much lower power.

The increase of power and reduction of combinatoric complexity provided by even four or five deictic variables has also been demonstrated (eg. Chapman 1989, Horswill 1995). Deictic variables allow us to combine the speed and reliability of reactive systems with some of the flexibility of symbol-based AI.

Minimal state implementations of intelligent control are also appealing because they imply correspondingly minimal storage requirements for episodic memory. Binding data, including reference to the control context, could be the sort of indexical information that is stored in the hippocampal system (McClelland 1995 pp.451-452). If episodic memory is stored by reference, then remembering is a constructive process of rationalizing sparse information. This could explain many recall affects, such as the suggestability of witnesses over long periods of coaching or the rapid confabulations of the deranged, without postulating complex representational redescription. For the witness, episodic memory might reference a perceptual routine that changes through learning over time. For the deranged, confabulation from inaccurate pointers may be just as fluent as a normal person engaged in explanation or reminiscence.

The aforementioned `perceptual routine' need not require complex representational manipulations. Work by Tanaka (1996), Perrett (1996) and others indicates that many supposedly high-level cognitive tasks such as recognizing shapes, faces (either general or individual), and even the direction of another's attention may be performed by cells sensitive to higher-order visual features. These cells are ordered topographically with smooth transitions in feature specificity in a way similar to orientation-specific cells in the visual cortices. Perrett's theory also allows an account of the mental rotation task that does not postulate complex representational transformations over time; consistent with Ballard et al's theory, it only requires a single neural pointer moving over precompiled structure.

Surprisingly, Ballard et al's discussion of temporal bands (p.4) makes no reference to the work of Poeppel and colleagues in this area (eg. Poeppel, 1994). Extensive reaction time studies point to a processing window of about 30msec. 30msec is the smallest interval that two stimuli can be temporally ordered; events occurring within the interval are treated as simultaneous. Poeppel suggests that treating stimuli within the window as simultaneous allows the brain to normalize for differing sensory transduction times.

This research may also be relevant to the issue of pointer binding. 40Hz brain waves have been implicated in perceptual processing (see Phillips & Singer, forthcoming, for a review). This frequency defines a sequence of system states of approximately 30msec duration. Current neuronal theories of perception (Malsburg, 1995) use synchronous oscillations to bind features together within system states.

If pointer binding is due to synchronous oscillation. we might also have a more biological explanation than those offered on pp.13-14 for the limited number of available pointers. Oscillation rates are highly dependent on the electro-chemical properties of the nervous system. Only a handful of distinct phases within the 40Hz oscillations can co-exist without mutual interference. This could constitute a neural constraint on the number of pointers simultaneously active.

Ballard, Hayhoe, Pook and Rao's theory constitutes an advance toward an alternative understanding of intelligence based on immense tables of perceptual and motor skills tightly coupled with functional routines, where coherence emerges through dynamically bound deictic variables. There has long been a debate as to what extent our intelligence is constrained and affected by our biology. Perhaps these are some new answers.

Brooks R. A. (1991) `Intelligence without Representation' Artificial Intelligence 47 p.139-159

Chapmann D. (1989) `Penguins Can Make Cake' AI Magazine 10 4 p51-60

Horswill I. D. (1995) `Visual routines and visual search' Proceedings of the 14th International Joint Conference on Artificial Intelligence 1995

von der Malsburg C. (1995) `Binding in models of perception and brain function' Current Opinion in Neurobiology 5 p520-526

McClelland J. L, McNaughton B. L, and O'Reilly R. C. (1995) `Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory' Psychological Review 102 3 p.419-457

Perrett I. D. (1996) `View-dependent coding in the ventral stream and its consequences for recognition' in R. Caminiti et al. Vision and Movement: Mechanisms in the Cerebral Cortex HFSP Strasbourg

Phillips W. A. and Singer W. (forthcoming) `In search of common cortical foundations' Behavioral and Brain Sciences

Poeppel E. (1994) `Temporal mechanisms in perception' International Review of Neurobiology 37 p185-202

Tanaka K. (1996) `Inferotemporal cortex and object recognition' in R. Caminiti et al. Vision and Movement: Mechanisms in the Cerebral Cortex HFSP Strasbourg