Psych 129 - Sensory processes

Distributed coding and adaptation

The representation of information

• Beyond simply characterizing the response properties of neurons, such as measuring their receptive fields, we would like to know how information is represented by the ensemble activity of neurons within the brain.
  
  
• One hypothesis that has been proposed, for example, is that of grandmother neurons. The idea here is that when you look at your grandmother, there is one neuron at some higher level of the brain (for example, in IT) that fires in response to that stimulus, and that single neuron’s activity is the representation of your grandmother. If you look at a different object, then there is another neuron that fires in response to that object, and so forth. One neuron per object in the world. The advantage of such a representational scheme is that it makes it very easy to form associations and to mediate appropriate behavior. The disadvantage is that you need as many neurons as there are objects (which is alot), and if the neuron which codes your grandmother dies, then grandma goes bye-bye. In addition, it begs the question of how you learn new objects - do you have to keep a pool of undedicated neurons around so that they may be assigned to new objects?
  
  
• At the other end of the extreme are dense, distributed codes, similar to the ascii code coming from your computer keyboard. Here, the representational capacity of a relatively small number of neurons is enormous - you can represent as many things as there are distinct combinations of activity. The problem though is that it becomes very difficulty to form associations, as the information pertaining to the presence of a particular object is tied up in the activities of many units.
  
  
• The happy medium between these two extremes is sparse, distributed coding. Here, information about objects is distributed among multiple neurons, but the number of neurons that are active in response to any given object is relatively small. In such a distributed code then, you cannot look at the response of any particular neuron, but rather you must look at the activity across an entire ensemble before deciding whether a particular object is present in the world.
  
  
• One way of studying the nature of distributed coding within the brain, without invasively recording from neurons, is through adaptation experiments. This is one of the most important and widely used experimental methods of psychophysics.
  
  
  

• Adaptation effects are thought to be caused by the habituation of neurons to sensory stimuli. Habituation basically means that the neuron “poops out” after repeated stimulation. For example, if a neuron initially fires vigorously in response to a vertical edge in an image, continued stimulation of the neuron will eventually cause it to be only weakly excited by the edge.
  
  
• Adaptation can be used to reveal perceptual mechanisms (i.e., neural processing) at work within the visual system. As neurons become habituated, they no longer do their job properly, and the effect of this may be seen by measuring what aspects of perception have changed as a result. If neurons tend to be narrowly tuned to a particular stimulus dimension, then adaptation will affect only a narrow range of stimuli. Conversely, if neurons tend to be broadly tuned to a particular stimulus dimension, then adaptation will affect a much broader range of stimuli. Thus, by observing the range of stimuli that are affected by adaptation to a particular stimulus, it is possible to infer the width of tuning of neural mechanisms within the brain.
  
  
• For example, in the so-called tilt illusion (shown in figure 8.29, p. 236 of the text - note that in some textbooks the caption is wrong though), adapting to a grating pattern tilted to the left will cause a perfectly vertical grating pattern to look tilted to the right when viewed immediately afterwards. It is believed that the reason for this is that orientation-tuned neurons whose optimal orientation (i.e., the orientation that yields the maximum response from the cell) is near that of the adaptation grating become habituated, and so they have a weaker than normal response when viewing the vertical grating. The lack of response in these cells will cause the population of neuron activity in the cortex to look different to higher centers of the brain reading out the responses of these cells. The difference is such that the peak in activity among the neurons will be skewed toward the cells whose optimal orientation is opposite that of the adaptation grating (since they were least habituated). Thus, the higher centers reading out this population activity will infer that the vertical test grating is actually tilted to the right. The amount of illusory tilt induced may be used to infer the width of orientation tuning in cortical cells.
  
  
• The same logic may be extended to spatial-frequency selectivity. Adapting to a grating at one spatial-frequency will cause a grating at a slightly different spatial-frequency to look different (i.e., lines spaced further apart or closer together) than it actually is. This is demonstrated in figure 3.24 and explained on pp. 87-88 of the text.