Basic emotions correspond to bodily signals. Many psychologists think that there are only a few basic emotions, and that most emotions are combinations of these few. Here we advance a hypothesis that the number of principally different emotions is near infinite. We consider emotions as mental states with hedonic content, indicating satisfaction and dissatisfaction. Our hypothesis is that a large number of emotions are related to the knowledge instinct (KI, or a need for knowledge). Contradictions between knowledge and bodily motivations, between various elements of knowledge are known as cognitive dissonances. We suggest that specific emotions are involved with cognitive dissonances. The number of cognitive dissonances is combinatorial in terms of elements of knowledge. Correspondingly, the number of these knowledge-related emotions is very large. We report experimental results indicating that emotions of cognitive dissonance exist. We propose that these emotions are different from basic emotions in principle, and outline future research directions toward showing that their number is large.

People learn the meaning of words in ambiguous contexts with many possible words for any referent and many referents for any word. Cross-situational learning is an approach to solve this word-to-world mapping problem based on the idea that a learner can determine the meaning of a word by finding something in common across all observed uses of that word. Here we investigate the performance of a simplified variant of the general-purpose Neural Modeling Fields (NMF) categorization mechanism to infer the correct word-referent mapping in a cross-situational learning scenario that simulates experiments with adult subjects. We study two learning situations: the batch-mode learning in which the processing of data requires the memorization of all training examples, and the online learning in which the data processing occurs concomitantly with the exhibition of the examples. A training example consists of a picture of a number of objects accompanied by the utterance of the same number of words. We show that the equations derived to describe the batch-mode learning situation can also be applied to the more realistic online learning situation The resulting online algorithm yields predictions which are both qualitatively and quantitatively in agreement with the experimental results.

Cross-situational learning is based on the idea that a learner can determine the meaning of a word by finding something in common across all observed uses of that word. Although cross-situational learning is usually modeled through stochastic guessing games in which the input data vary erratically with time (or rounds of the game), here we investigate the possibility of applying the deterministic neural modeling fields (NMF) categorization mechanism to infer the correct object-word mapping. Two different representations of the input data were considered. The first is termed object-word representation because it takes as inputs all possible object-word pairs and weighs them by their frequencies of occurrence in the stochastic guessing game. A re-interpretation of the problem within the perspective of learning with noise indicates that the cross-situational scenario produces a too low signal-to-noise ratio, explaining thus the failure of NMF to infer the correct object-word mapping. The second representation, termed context-word, takes as inputs all the objects that are in the pupil's visual field (context) when a word is uttered by the teacher. In this case we show that use of two levels of hierarchy of NMF allows the inference of the correct object-word mapping.

Movement seems to be the key ingredient to understanding how children perceive (and hence name) objects in their environment. This insight, which is based on Spelkepsilas experiments on language acquisition by children, is akin to Aristotlepsilas conception of object as a ldquocontinuous thingrdquo that has one and only one movement. Here we test this idea with a computer experiment in which the perceptual system of the individual is modeled by the Neural Modeling Fields categorization mechanism, and the environment is composed of two complex objects that can move with respect to each other. Rather remarkably, we find that movement indeed makes possible the differentiation of objects which the individual would deem indistinguishable if motionless.

We propose a categorization task scenario to study language acquisition in which an agent receives linguistic input from an external teacher, in addition to sensory stimuli from the objects that make up the environment. The agent is endowed with the modeling field theory (MFT) categorization mechanism, which enables it to identify overlapping categories from the exposition to hundreds of examples. Rather remarkably, we find that the agent with language is capable of differentiating object features that it could not distinguish without language. In this sense, the linguistic stimuli prompt the agent to redefine and refine the discrimination capacity of its sensory channels.

Evolutionary language games have proved a useful tool to study the evolution of communication codes in communities of agents that interact among themselves by transmitting and interpreting a fixed repertoire of signals. Most studies have focused on the emergence of Saussurean codes (i.e., codes characterized by an arbitrary one-to-one correspondence between meanings and signals). In this contribution, we argue that the standard evolutionary language game framework cannot explain the emergence of compositional codes-communication codes that preserve neighborhood relationships by mapping similar signals into similar meanings-even though use of those codes would result in a much higher payoff in the case that signals are noisy. We introduce an alternative evolutionary setting in which the meanings are assimilated sequentially and show that the gradual building of the meaning-signal mapping leads to the emergence of mappings with the desired compositional property.

The emergence of communication is studied in a scenario where agents endowed with distinct object-meaning mappings learn from scratch signal-meaning associations (i.e., communication codes) that allow them to identify the objects in their environment. Meanings are created through the Modeling Field Theory categorization mechanism, and learning is based on two variants of the obverter procedure, in which the agents may or may not receive feedback about the success of the communication episodes. We show that in the unsupervised learning scheme the agents fail to develop ideal communication codes, whereas success is guaranteed in the supervised scheme provided the size of the repertoire of signals is sufficiently large, though only a few signal are actually used in the code. Thus the mere ability to produce and observe different signals bears on the quality of the evolved communication codes.

In order that communication can take place there must be something to be communicated. This basic stage of language evolution is the symbol grounding problem which addresses the issue of how physical signs acquire meaning. It is the symbols (e.g., words) associated to those meanings that are communicated by language. Here we show how the combination of the modeling field theory and the Akaike information criterion can be used to find the true number of objects in an environment. We demonstrate creation of suitable representations and meanings for those objects and discuss the possible role of language in improving these representations.

We address the issue of the precision required by an N-input threshold element in order to implement a linearly separable mapping. In distinction with previous work we require only the ability to correctly implement the mapping of P randomly chosen training examples, as opposed to the complete boolean mapping. Our results are obtained within the statistical mechanics approach and are thus average case results as opposed to the worst case analyses in the computational learning theory literature. We show that as long as the fraction P/N is finite, then with probability close to 1 as N →∞ a finite number of bits suffice to implement the mapping. This should be compared to the worst case analysis which requires O(N log N) bits. We also calculate the ability of the constrained network to predict novel examples and compare their predictions to those of an unconstrained network. Finally, we address the issue of the performance of the finite-precision network in the face of noisy training examples.