Learning is ... - a relatively permanent change in behavior and/or - - PowerPoint PPT Presentation

• 11th International Conference on Intelligent Tutoring Systems

14 – 18 June, 2012 in Crete, Greece

Invited Address by

Norbert M. Seel

University of Freiburg, Germany seel@ezw.uni-freiburg.de

Learning is ...

• a relatively permanent change in behavior and/or in mental associations with

specific experiences.

• a response to environmental requirements
• different from biological maturation,

which, however, is a fundamental basis for learning.

The focus of this presentation is on

• 1. The biological and evolutionary constraints on learning
• 2. Human learning
• 3. Artificial learning

Biological and Evolutionary Constraints on Learning – 1 Learning is biologically predisposed. Modern views on learning and memory accept the notion of biological and evolutionary constraints ---- the formation of association is not uniform across all stimuli. Definition (Domjan, 2012): A biological or evolutionary constraint on learning is a limitation on classical

• r instrumental conditioning that is observed despite the use of procedures

that would be expected to produce successful learning.

Biological and Evolutionary Constraints on Learning – 2 Biological constraints on learning result from the genetic make-up of a given species and can affect sensory, behavioral and cognitive abilities of an animal. There are two branches of research, which point to the fact that genotypes can affect the relationship between behaviour and environment:  Studies on species-specific behaviors  Studies on taste-aversion learning

Biological and Evolutionary Constraints on Learning – 3

Species-specific behaviors

Keller and Marian Breland (1951) attempted to condition a raccoon to put a coin into a slot for food reinforcement, as a part of an entertainment display. As training progressed, the raccoon became increasingly reluctant to let go of the coin and let it drop into the slot. When required to put two coins into the slot, the Raccoon rubbed the coins together instead of releasing them into the slot. Attempts to train a pig to put a coin into a piggy bank produced a similar failure. The pig would root the coin on the ground rather than put it into the slot.

Biological and Evolutionary Constraints on Learning – 4

Taste-aversion learning

Another major constraint on learning was discovered by John Garcia in the course of his work on the biological effects of x-irradiation. Garcia et al. (1974) found that foot-shock (i. e. mild electrical shock) will readily condition an aversion to an audiovisual cue but not a taste cue in rats. In contrast, illness induced by irradiation or a drug injection will condition an aversion to a taste but not to an audiovisual stimulus.

Biological and Evolutionary Constraints on Learning – 5 Constraints on learning were also found in avoidance conditioning. For example, rats are more likely to learn to freeze, jump out of a box, or run in a wheel to avoid shock than they are to rear or press a lever. In pigeons, key pecking is much more difficult to condition as an instrumental shock-avoidance response than is pressing a treadle. In humans, even without conditioning – cautiousness towards unknown food, unlearnt unwillingness to taste new food, or even familiar food in a new environ- ment.

Biological and Evolutionary Constraints on Learning – 6 The biological constraints on learning contradict the behavior systems theory which assumes that evolution has shaped behavior into functionally organized units that enable organisms to successfully accomplish biologically important tasks such as feeding, defense, and reproduction. Learning procedures are superimposed on the behavior system that is activated at the time of training. For example, food deprivation and food reinforcement activate the feeding system. Behavior systems theory focuses on the outcomes of that evolutionary process rather than the ecological context in which evolution occurs.

Biological and Evolutionary Constraints on Learning – 7 An alternative approach, proposed by Domjan (2008), emphasizes that evolutionary selection cannot be divorced from the ecological context in which the selection takes place ... .... looking for biological influences on learning by examining how learning takes place in an organism’s natural habitat. Evolution has significantly shaped and constrained what humans and

• ther animals learn about and how

learning occurs. This in turn has led to abandonment

• f the principle of equipotentiality.

Influences of cognitive processes on learning thinking, remembering, perceiving, using a language Social learning – or observational learning – Examples: Animal social learning and innovation Primate social cognition Rule learning – rules are instructions about how to behave in certain situations; they are discriminating stimuli, in fact. Sammy, the Beagle dog ....

Influences of cognitive processes on learning thinking, remembering, perceiving, using a language Cognitive maps – (Tolman) Organisms learn the general topography

• f their environment, even if not reinfor-

ced. Animals create cognitive maps of their environment in order to learn. Insight learning – (W. Köhler) Many problems consist of inner relations, which sometimes need restructuring for the problem to be solved. Insight is thus defined as a sudden understanding of the relations between the elements.

Human Learning – 1 Humans and other animals have three essential abilities for processing information and acting successfully in various environ- ments.

• 1. They are very good at pattern matching

in order to quickly settle on an interpretation

• f an input pattern.
• 2. Humans are very good at modeling their worlds

with the aim to anticipate new states of affairs.

• 3. Humans are good at manipulating their environ--

ments with the aim to create external represen- tations (i.e. a version of (hu)man-the-tool-user.

Human Learning – 2 These basic capabilities presuppose a cognitive system with two modules or sets of units (Rumelhart et al., 1986):  an interpretation network – which is concerned with producing appropriate responses to any input from the external world.

The interpretation network receives input from the world and reaches a relaxed mental state by producing relevant cognitive responses.

 an model of the world – which is concerned with producing an interpretation of what would happen if we did that with a particular external representation.

Human Learning – 3 This cognitive architecture corresponds to Piaget’s epistemology that cognition is regulated by the continuous interaction between assimilation and accommodation, which aims at adjusting the mind to meet the necessities of the external world .

Human Learning – 4 Assimilation is dependent on the activation of schemas, which allow new information to be integrated into existing cognitive structures. Schemas are slot-filler structures that serve central cognitive functions, such as integrating information into cognitive structures, regulating attention, making inferences in the process of acquiring knowledge, and reconstructing it from memory “Without a schema to which an event can be assimilated, learning is slow and uncertain” (Anderson, 1984, p, 5).

Human Learning – 5 Accommodation aims at restructuring knowledge ... .... by means of accretion, tuning, or the reorganization of schemas and their content (Norman & Rumelhart, 1978). . If an adjustment of a schema is not possible, i.e., if the accretion, tuning, and/or reorganization of a schema fails –

• r if no schema can be activated at all –

the learner either can abandon the cognitive processing or must invest some mental effort to develop a mental model.

Human Learning – 6 As long as the information being processed can be assimilated promptly into cognitive structures and as long as schemas can be modified by means of accretion, tuning, and reorganization, there is no need to construct a mental model.

Cumulative Learning – 1 Learning of humans and other animals is cumulative by nature. Learning systems, human or artificial, accumulate knowledge and abilities that serve as building blocks for subsequent cognitive development. Cumulative learning deals with the gradual development of knowledge and skills that improve over time. In both psychology and artificial intelligence, such layered or sequential learning is considered to be an essential cognitive capacity ... ... in acquiring useful aggregations and abstractions that are conductive to intelligent behavior and in producing foundations for further cognitive development.

Cumulative Learning – 2 “There has been remarkably little study of learning -- real learning, the learning of complex topics, the learning that takes months, even years to accomplish. Elsewhere I have estimated that experts at a task may spend 5,000 hours acquiring their skills: that is not such a long time; it is 2 1/2 years of full-time study, 40 hours a week, 50 weeks a year. Not much time to become a professional tennis player, or computer programmer, or linguist. What goes on during that time? Whatever it is, it is slow,

• continuous. No magic dose of knowledge in form of pill or lecture. Just a lot of slow,

continual exposure to the topic, probably accompanied by several bouts of restructuring

• f the underlying mental representations, reconceptualizations of the concepts, plus

many hours of accumulation of large quantities of facts [...] Very little effort gets spent at studying what it would take to accomplish this, perhaps because there is the implicit realization that the task is harder than it might seem. [...] And so the study and understanding of the learning process remains at a miniscule level. Pity" (Norman, 1981, p. 284).

Cumulative Learning – 3 PISA shocked the world (i. e. Germany) in 2000 .... In one of the follow-up PISA-studies, PISA-I-Plus, the longitudinal change of learning results was measured across two grades (at two measurement points). In mathematics, the students (15 and years old) had improved their learning with 60 %, in science only with 44 % .... but 20 % had forgotten what they had learned before. Contrarily, an extensive meta-analysis by Semb & Ellis (1994) shows that students retain much of what they have learned in the class- room and, amazingly, they retain in over a really long time. Of course, a loss of retrieval would be observable after some time but forgetting was evidently not comparable with the forgetting curves in psychological learning experiments.

Cumulative Learning – 4 That‘s the curve expected by most people (and by the PISA guys): But a learning „curve“ is far from a straight progression: What we should expect:

Cumulative Learning – 5 Remember: Learning of humans and other animals is cumulative by nature. The primary benefit of CL is that it consolidates the knowledge one has obtained through the experiences, allowing it to be reproduced and exploited for subsequent learning situations through cumulative interaction between prior knowledge and new information. The demographic and ecological success of our species is frequently attributed to

• ur capacity for cumulative cultural learning.

However, it is not yet known how humans combine social and individualized learning to generate effective strategies for learning in a cumulative cultural context.

Incremental Learning – 1 In Informatics and the field of AI, cumulative learning is called incremental learning. Pfeffer (2000), for example, defines a cumulative learning agent as one that learns and reasons as it interacts with the world by using its accumulated knowledge and its observations. The term “incremental learning” is often used for sequential or constructive learning in contrast to batch or epoch learning (Bertsekas and Tsitsiklis 1996). Incremental learning is based on the principle of starting with simple and basic principles before advancing to more complex information.

Incremental Learning – 2 Incremental learning happens in bits and pieces, and successful retention of knowledge is based upon previously attained knowledge. Accordingly, in the field of artificial learning some algorithms for incremental learning have been developed with regard to concept formation and perception of visual categories.

Incremental Learning – 3 Easterlin (1986) divided the process of concept formation in machine learning into three subsequent components:  Aggregation, in which important instances of experiences are grouped into a set of

• aggregates. Experiences are aggregated by the learning system itself for further use

based on their contribution to a successful problem solution and to system perfor- mance.  Characterization, in which a description of the essential information for an aggregate

• f experiences is generated or constructed in terms of characteristics that are useful

to the system based on individual descriptions of each member of the aggregate.  Utilization, in which the concept description is integrated with the performance element of the system and the important aspects of the aggregate are captured. .

Incremental Learning – 4 This process leads to the formation of concepts containing functional information. Then, the agent applies its operators to the subsequent task by mapping objects and

• perators of current states to consequent states, thus producing a generalized schema

with its background knowledge. In Pfeffer’s (2000) Integrated Bayesian Agent Language (IBAL), a learning agent can modify its models based on its collected observations and use them in future situations.

References Only one:

Thanks for your attention.