SS 246: Perception
Yasser Hashmi
M. Omer Sheikh 2003-02-0129
Exploitation of
Regularities in the Environment by the Brain
A Summary
This paper summarizes a paper by Hoarce Barlow. This paper is a critique on the work of Roger Shepard on the internalization of Kinematic Geometry in favour of exploitation of statistical regularities in the environment for use in perception of events by the brain.
Statistical regularities are important for learning, memory, intelligence, inductive inference and general information processing. Brain exploits regularities to make the organism survive better. It is currently known that neural mechanisms use statistical properties of images. Shepard’s proposal of internalization is not statistical and no benefits or advantages can be derived from it.
Statistical regularities so abundant in the world are important for survival. Shepard believes that the apparent trajectory of motion of an object shown at two different points was guided by the internalized kinematic geometry of moving objects just like the ‘other’ school of thought believes perception to be adapting to statistical regularities. All learning is internalization of regularities in the environment which are reinforced due to repeated occurrence. Learning is influenced by associations between sensory stimuli, between stimuli and reinforcement. Hence we can not be exactly sure of when learning begins and perception stops and vice versa.
Some earlier ideas about use of statistical regularities in the brain are outlined.
Helmoltz (1925) believed that perception is a ‘conclusion’ unconsciously formed. This is the result of the combination of ‘apperception’ or the immediate impact of a stimulus and with the ideas remembered from a past experience.
Mach (1886) and Pearson (1892) believed in the ‘Economy of Thought’ brought about by scientific laws which he argued simplify the complex worldly experiences.
Craick (1943) wrote that the higher cognitive center builds symbolic working models that are expressions of the environmental regularities based on associative structures of objects and events in the environment.
Attneave (1954) incorporated mathematical concepts of information, channel capacity and redundancy introduced by Shannon and Weaver (1949). He directed attention to the fact that natural images have much redundancy and the ‘subjective’ prominence of borders was a psychological mechanism that took advantage of the information carrying parts of the image.
Barlow initially thought that the main benefit to be derived from exploiting redundancy was economy. Later work showed that brains physiology and anatomy is not suited to the use of compressed economical representations. Instead the brain uses principle of redundancy exploitation. Accommodation of sensory discharges to constant stimuli, light and dark adaptation and lateral inhibition use the principle. In 1959 Hubel and Wiesel discovered that orientational selectivity of cortical neurons also exploits redundancy in images.
Barlow (1961) suggested that economy is brought by reducing the impulses in neurons rather that reducing the number of neurons involved. Barlow (1972) established that small numbers of neurons selected from a large population represented the sensory input in a distributed form. The elements of the distributed form are called the ‘cardinal cells’ and signal the occurrence of messages which would be useful to learn about.
In 1989 Barlow spoke about the associated structure of sensory messages where the elements that encode environmental regularities are statistically independent. He proposed factorial coding for this. At higher levels in the brain there are vastly more channel each active at a lower rate. The coding becomes sparse but redundancy seems to be increased. In distributed representation, multiple stimuli activate the same neurons and hence errors occur. To reduce the loss in efficiency, many more neurons are required so as to increase redundancy. One reason that the cortex has so many neurons might be that to learn associations, input patterns have to be identified and their frequency has to be estimated. Hence sensory redundancy is important because knowledge of regularities in environment is advantageous say in making predictions. Redundancy in representation reduces the extent to which elements are active at any time.
Watanabe pointed at the similarity between inductive inference and recoding to reduce redundancy. Inductive inference uses sophisticated regularities to produce component descriptions of data much like what is done in minimum descriptor length approach which says that the shortest computer code to produce a sequence of data is the least redundant representation of data.
In 1998 Haterten and van der Schraff ran independent component analysis on images. They showed that the predicted receptive fields matched the ones determined experimentally in some if not all properties. Hence it is very possible that the V1 neurons are adaptive to regularities in natural images.
Now Shepard’s work is discussed in light of the arguments formulated above.
To understand internalizing as an evolutionary adaptation, it is not sufficient only to copy regularity internally but it must also be turned to some advantage so that the internalization can survive. Shepard exemplifies by diurnal rhythms where the rhythm itself is internalized. But generally the mechanisms required to gain advantages are likely to be very complicated and not very obvious.
Subjects seeing an object in two places experience it moving along a path between the points and according to Shepard the path followed is close to that dictated by Chasle’s rule which essentially states that every motion is composed of a translation and a rotation about an axis. The problem here is that the object does not necessarily has to move along the path corresponding to simultaneous rotational and translation (Todoroui 200). Secondly, accuracy of subject’s judgments is not established to distinguish Shepard’s predictions from those types that occur frequently in images of moving objects and the redundancy exploitation hypothesis predicts. An even more fundamental problem is that the neural basis for subjective experience of moving objects is not known. Shepard’s claim assumed true in a specific framework raises the question that how could interpolated representations of static positions be formed early in the movement before the object had appeared at its second position. This model ignores modern neurophysiology of sensory systems and hence is not convincing. It is also known that brain has some ways of representing image information which makes it useless to apply kinematical geometry rules.
The advantages that can be derived from evolutionarily adapted mechanisms were shown by van Hatern and van der Shroff (1998). The results suggest that cortical neurons selectively respond to spatial patterns and spatio-temporal patterns of moving features in images of moving objects. Such neurons could act as matched filters for these patterns and be optimal for detecting them at low SNR’s or early in the course of movement. They can also signal spatio-temporal patterns in early stages and also interpolate motion. Shepard does not suggest any of these possible benefits.
Sinha and Poggio (1996) described experiments which show that subjective experience is heavily influenced by mathematical perspective transformation rules, recent experience of a particular motion and a tendency to consider as rigid the motions that occur together. Shepard on the contrary proposed that internalized geometric rules lead directly to subjective experience. Shepard assumes that perceptions embody whole kinematic geometry regularities. But the experiments by Sinha and Poggio show that laws of perspective transforms are used. This fits the earlier ideas presented in this regard.
A hierarchy of operations occurs int the visual areas of the brain and strong evidence shows that at last in the early ones redundancy is really exploited. Hence the idea of internalization of kinematic geometry does not seem very plausible.