I have recently crossed paths with David Bryson M.D. and through our discussions he brought to my attention an article he felt complements my work
- Physical: Regulate, Repose
- Decisional: Record, Report
- Perceptual: Relate, Revise
and which with his permission I am reproducing here. Dr. Bryson is currently working on a concept he intends to present during the 2009 celebration of the 200th anniversary of Darwin’s publication of Evolution of the Species. –Grant Czerepak
BEHAVIORAL ANALYSIS OF MAMMALIAN SLEEP AND LEARNING
Perspectives in Biology & Medicine, Autumn Issue,1969, pages 71-79
University of Chicago Press
David Bryson, M.D. and Stephen Schacher, M.D.
Theories of mammalian learning and theories of mammalian sleep have developed with scant interaction. This article states that mammalian sleep and learning are related fundamentally. Our approach is theoretical; no new experimental data are presented. Aspects of human behavior which are distinct from those of mammals in general will not be considered here.
Much of the mature mammal’s behavior is devoid of any significant decision making, due to the familiarity of most of the inputs with which the mature mammal deals. Familiar inputs are associated with the execution of routine behavioral acts, and during such periods mammalian record keeping is insignificant. In contrast, the record of the mammal’s activities which is kept is directly related to episodes of decision making. A decision is made when the mammal is confronted with an unfamiliar input, and this input, the resulting decision, and the outcome of the decision are recorded. For example, a hungry mammal may suspect that some unfamiliar substance is edible, may decide to sample it, and may find the taste unacceptable. This input, this decision, and this consequence are recorded.
The decisional record thus grows as the waking state progresses, at a rate related to the frequency of unfamiliar inputs. The decision record may be accessed as soon as a decision and its consequences have occurred, enabling a mammal to learn within a single waking period. Should some unfamiliar input recur, the current need for decision making may be reduced considerably if the previous decision for the same input had and acceptable consequence. If not, the probability of invocking this decision again is decreased, and decision making yields another decision and perhaps a more acceptable consequence.
Learning of this type is inadequate in one important respect. Access to the decisional record is limited to one decisional cluster (unfamiliar input, decision, consequence) at a time; decisional clusters are not themselves compared. For those inputs which do recur, the original problem of input unfamiliarity is improved empirically on a case-by-case basis. However, improved decision making for each case still fails to exploit any generic relationships which may exist between the individual decisional clusters.
During sleep, the information in the decisional record is utilized for a further purpose. The individual decisional clusters are combined into informational sets, the combinatorial rule being to group those unfamiliar inputs in which the decisions and consequences were similar. For example, suppose a mammal in an unfamiliar environment has spent various portions of the current waking state in trying to establish reliable landmarks for navigational purposes. During sleep, those environmental features which were selected to be–and in fact turned out to be–reliable landmarks are grouped into one informational set. Inductive analysis of this set of unfamiliar inputs may reveal recurrent similarities in their characteristics, a finding with implications for the input classification (perceptual) system. If so, the perceptual system is revised as the mammal sleeps. Sleep thus enables the mammal to classify unfamiliar inputs having the same significance as the same perception. Thus previously unfamiliar inputs subsequently are perceived with increased familiarity.
Two categories of mammalian learning are thus proposed: decisional learning, a waking activity in which output selection for the same input is improved; and perceptual learning, a sleeping activity in which the functional reorganization of input classes reduces the number of decisional involvements required previously. Decisional learning is akin to “stimulus-response” learning; perceptual learning is akin to “gestalt” learning. All perceptual learning is the result of recent decisional learning. When the inputs associated with a series of decisions with inadequate consequences can be related and thus lead to perceputal learning, a considerable behavioral advancement may occur from one waking state to the next.
The first section of the article develops a model of the informational operations between initial input (attention) and final output (manifest behavior). The model describes how decisional learning and perceptual learning are related to the overall control of behavior. The setting of the model is ethological. The second section confronts learning research in the laboratory setting. The research section is referenced; the model section is not.
The model is presented at the level of a simple behavioral analysis. Our prototype mammal is always in one of two output modes when interacting with the environment: decision making, or decision execution. During decision making, the mammal is obviously inspecting or choosing; during decison execution, the mammal is obviously operating on the environment as part of some goal-directed act. Decison making and decison execution are to us thus a matter of seconds rather than milliseconds.
Two sequential processes are invovled on the input side of mammalian behavior: attention (input sampling) and perception (input classification). Attention and perception are components of both decision making and decison execution. The sequence of attention-perception-output-attention-perception-output may thus represent either ongoing decision making or ongoing decision execution.
We classify input into three types: potential input, attended input, and perceived intput. The process of perception converts attended input into perceived input.
Potential input is the totality of the physical environment to which the mammal theoretically may attend at any time. Potential input is thus a range of input information, reflecting the theoretical limits of the mammal’s sensory equipment. Attended input is that constellation of physical variables (light, sound, chemical concentration, etc.) to which the mammal actually attends at any time. The dimensions of potential input and attended input are always physical. During decison execution, the dimensions of perceived input becomes those of the utility and survival of a particular mammal at a particular time.
During decision execution (peeling a fruit, copulating, grooming), the mammal actively attends to its own behavioral output and in so doing excludes the remainder of potential input. Attended inputs here are thus an extremely narrow and gradually shifting sample of potential input. As long as attended inputs remain within prescribed tolerances, decision execution continues toward completion with no interruption for decision making.
Decision execution may be interrupted by sudden changes in potential input (as sudden noise or shadow). Attention automatically is diverted from behavioral monitoring. If the environmental change, now represented in attended input, is perceived as obviously irrelevant to the mammal (as a familiar “false alarm”), attention is returned to consummating the current goal. If the change is perceived as obviously relevant (a familiar opportunity or danger of greater significance than the current activity), the mammal quickly decides upon the more appropriate behavioral act and devotes its attention to the execution of this decision rather than the previous decision.
When the relevance of attended input is not immediately apparent, the mammal switches to the decision making mode. During decision making, attention is determined by a complex interaction between the mammal and its environment. The mammal attends to increasingly diverse aspects of potential input, making active searches for suspected environmental features and also being passively drawn to unexpected environmental features, and continues to do so until it discovers or is confronted with some environmental feature of obvious relevance. The mammal then returns to the decision-execution mode.
Every attended input is transformed into a perceived input. The product of perception is a perceived input of variable familiarity. Perceived inputs of high familiarity are perceived as action oriented (output oriented) and are associated with ongoing decision execution or the initiation of decision execution. Perceived inputs of low familiarity are perceived as feature oriented (input oriented) and are associated with ongoing decision making or the initiation of decision making.
During decision execution, two mammals may attend to similar features of the environment but have quite different perceptions. For example, suppose two mammals are attending to certain physical features of an edible plant which only grows near an open water source. Attended input A of mammal A is very similar to attended input B of mammal B. However, perceived input A is high in a dimension reflecting the accessibility of food, which mammal A happens to be looking for, while perceived input B is high in a dimension reflecting the accessibility of water, which mammal B happens to be looking for. During decision execution, perception reflects the current functiona requirements of the mammal. During decison making, because of the difficulty in classifying unfamiliar attended inputs, perception reflects the more constant structural features of physical reality.
On the output side of the model, every perceived input is transformed into an output signal. During execution, the output signal furthers the behavioral action represented in the precediing perceived input. During decison making, the output signal is related to obtaining more information relevant to the precedinng perceived input. During both output modes, the output signal directs attention in relation to the formation of the next attended input.
Within a single waking state, decision making results in decisional learning. Decisional learning can directly reduce the duration of subsequent decision-making episodes, and can indirectly reduce the number of subsequent decision-making episodes by improving the quality of decision making. During sleep, decisional learning results in perceptual learning. Perceptual learning can directly improve decision execution by improving input classification, thereby reducing the number of unnecessary interruptions. Perceptual learning sets the tolerances for variation in inputs at a functional level. Improved decison execution thus directly reduces the number of decision-making episodes.
Wide variations exist for the relative role of perceptual learning among different mammalian classes. Rodents are born with a number of input classes already fixed, and relatively few input classes are added as life proceeds. Rodents are nearly always in the decision-execution mode, since they have so few behavioral acts to decide between, and since their tolerance for input variation during decision execution is relatively broad. Most of the input classes of primates are learned, and the complex behavior of an adult primate requires a considerable repetoire of pereceptual categories. Primates are often in the decision making mode, since they have so many behavioral acts to decide between, and since their tolerance for input variation during decision execution is relatively narrow.
Such variation is related to the phenomena of play and curiosity. These are common features of primate behavior, especially during development, and are virtually absent from rodent behavior. We consider play and curiosity as phenomena which maximize decison making. Increased decision making leads to increased decisional learning, decisional learning to perceptual learnining, and perceptual learning to improved decision execution. As a result of play and curiosity, the behavioral repetoire of a mammal is increased, and thereby future waking-state activities which now have a goal-fulfilling function may be performed more effectively.
Implications for Research
Our model suggests that the unusual interaction of decisional learning and perceptual learning results in gradual, quantitative improvements in mammalian performance. For behavioral acts of which the mammal is already capable, improved output selection can simulate revised input classification before the latter occurs. Therefore sleep is not required for quantitative improvement in performance.
In order to demonstrate the role of sleep in mammalian learning criterion performance should represent a qualitative improvement over current performance. Such performance corresponds to the “problem-solving” class of previous experiments. While perceptual learning usually results in the modification of a previous input class, problem-solving behavior requires the formation of a new input class corresponding to the problem solution. We thus consider problem solving as a special case of perceptual learning. Since a new input class is required, decisional learning cannot simulate perceptual learning, allowing the role of sleep to be tested.
That sleep is not required for quantitative improvements in performance is obvious from experiments showing improved performance from the first trial outward . We analyze such performance as improved output selection for the same perceived input (decisional learning). That sleep is required for qualitative improvements in performance is not obvious from previous research. To our knowledge, no experiment in mammalian problem solving has studied sleep as a performance variable. The following analyses of previous experiments are therefore inferential.
Chimps which did not use a hoe to reach for food placed outside their cages did so “three days later,” after intervening ad lib experience with sticks . We suggest that primate curiosity resulted in the formation of a new input class in which sticks became perceived as a functional extension of the arm. A de novo decision execution thereby was made possible when the chimps were reconfronted with the original situation.
The “learning to learn” experiments with primates all take more than one training session ; improvement within a single training session is negligible . We suggest that such experiments can also be explained as a special case of perceptual learning, with the problem solution representing a new input classification.
The perception of “size constancy” (learning that stimulus properties are distance invariant) takes about ten days for rats previously reared in darkness . While not a classic problem-solving paradigm, we suggest that here again the formation of a new perceptual category was based on recent decisional learning, and the intervening sleep was a necessity.
Dogs and cats were trained to move around and behind a screen on the side to which food disappeared . The problem here is to learn that the same food continues to exist after visual representation of the food is no longer a part of attended input. No animal achieved problem solution (moving to the side of the food, sometimes to the right, sometimes to the left) on the first of daily training sessions. Furthermore, in the data as published, problem solution tended to appear on the first trial of a daily session (presumably the only trial after intervening sleep). We suggest that intervening sleep was, in fact, required for criterion performance. Before solution, both dogs and cats exhibited “position habits,” that is, always choosing the same route (for some always the left, for others always the right), resulting in 50 percent rewarded trials for chance alone. We suggest that high chance-based reward rates tend to preserve a decisional basis for behavior, and thus that decreasing the incidence of chance-based payoff will accelerate the appearance of problem solution.
Because sleep has been neglected variable in problem-solving experiments, the first trial of a daily training session usually has had two features which must be dissociated: it is usually only trial which follows sleep, and it is usually the trial which ends the longest intertrial interval. When waking time and sleeping time are equated, we predict that normal sleep is more favorable to problem solution than is any waking activity of equal duration, including continued exposure to a problem. Furthermore, we predict that in problem-solving experiments, as the above examples, criterion performance cannot be established in the first training session (no intervening sleep) regardless of the number of trials within the first training sessions. However, compression of trials within a few training session, as opposed to the same number of trials distributed over many training sessions, should favor the apperance of problem solution. Compression of relevant trials favors inductive operations in the following sleep state.
All decisional learning is time based (the cotemporality of unfamiliar input, decision and consequence within a decisional cluster). In the informational rearrangements of sleep, temporal relationships within a decisional cluster are reduced, and categorical relationships between decisional clusters are enhanced (perceptual learning). Since our model assigns decisional learning to a given waking state, we cannot rigorously approach the effect of sleep on decisional learning. We do suggest that the more a performance represents a unique temporal association devoid of significant implications for input classification, the more that intervening sleep will revert this performance toward the prelearning level. We therefore suggest that time-based performance is better retained by distributing a given trial over many training sessions, as opposed to compressing the same number of trials within a few training sessions.
A general finding for all mammals is a decrease in sleep as life proceeds. In terms of our model, increasing comprehensiveness of input classification increases input familiarity, increased input familiarity reduces decision making, reduced decision making reduces new record keeping, and reduced record keeping reduces the informational rearrangements of sleep. We thus suggest that an unfamiliar environment should cause increased sleep, regardless of the age of the mammal. In particular, we suggest that “sensory deprivation” is familiar to a newborn and thus should cause decreased subsequent sleep, while sensory deprivation is unfamiliar to an adult mammal and thus should cause increased subsequent sleep.
Mammalian sleep is composed of alternating periods of low and high brain metabolism . We speculate that sleep is a recurrent ABABAB phenomenon in which an A period (low metabolism) is a necessary preparation for the following B period (high metabolism). If true, specifically depriving a mammal of B periods (rapid eye movement sleep deprivation) should interfere with recent learning more than total sleep deprivation, even though the latter is of greater clock time.
We repeat what we think is the most important implication of our model for future research. Problem-solving learning is a special case of perceptual learning. We predict that normal sleep is more favorable to problem solution in mammals than is any waking activity of equal duration, including continued exposure to the problem.
- E.E. Smite. Psycol. Bull., 69:77, 1968.
- H.G. Birch. J. Comp. Psychol., 38:367, 1945.
- H.F. Harlow. Psychol. Rev., 56:51, 1949.
- D.R. Meyer. J. Comp. Physiol. Psychol., 44:162, 1951.
- D.P. Heller. J. Comp. Physiol. Psychol., 65:336, 1958.
- L.V. Kruschinsky. In: N. Wiener and J.P. Schade (eds.) Cybernetics of the nervous system, p. 280. Amsterdam: Elsevier, 1965.
- E. Hartmann. The biology of dreaming. p. 19. Springfield, Ill.: Thomas, 1967
- E. Aserinsky and N. Kleitman. Science, 118:273, 1953.