Linking Learning and Consciousness: the Self-Organizing Consciousness(SOC) Model

자료: http://www.u-bourgogne.fr/LEAD/people/perruchet/pdf2066Ko.pdf

Chapter 3.1 of: 

The Unity of Consciousness: Binding, Integration, and Dissociation

Ed. by Axel Cleeremans

2003, Oxford University Press

*** 

Introduction

There are two main ways of thinking about learning in contemporary cognitive sciences. In the historically earlier mode of thought, people acquire knowledge about situations through the application of specialized algorithms whose funcitioning follows certain logical or rational principles. In the second conception, people become sensitive to the regularities embedded in the material thanks to non-specific, and predominantly associative, mechanisms. These two conceptions are typically implemented in classic(formal, symbolic) models and in neural network modeling, respectively. Our starting point is that both conceptions of learning share a similar view with regard to the relation between learning and consciousness. Indeed, in the two conceptions, most if not all of the information manipulation and computation involved during a training episode takes place at an unconscious level. Learner's phenomenal experience has no relation with the hypothetical mechanisms of learning, whether based on symbolic or connectionist architectures.

Reiprocally, the literature on consciousness usually makes no reference to learning. Consciousness is usually construed as making accessible the end-product of the unconscious computations involved  during learning, with conscious thought taking no active part in the genuine learning process. The thesis developed in this chapter is that linking the notions of learning and consciousness provides us with an approach that may bring about a deep-seated revision of our understanding of the two concepts.

In the first section, we illustrate the above claims about the current divorce between learning and consciousness through a specific example. Then, using the same example, we introduce our alternative conception, in which learning is a naturally by-product of on-line phenomenal experience. Subsequent sections are devoted to generalizing our proposal. Our integrative conception is more parsimonious, we argue, because it avoids the needs for two independent machneries: one, fully uinconscious, devoted to learning, and the other generating phenomenal experience and devoid of any function in the learning processs. Of sepcial interest are the implications of our view with regard to the unity of consciousness. To anticipate our conclusions, instead of being a to-be-explained phenomenon without evident  utility, the unity of consciousness appears to be a fundamental concept for development and learning.

Two prevalent approaches to learning

The situation selected to illustrate our proposal concerns the discovery of words from a non-segmented speech flow. This ability is worthy of interest in itself, in so far as the formation of words can be construed as a necessary prerequisite for learning about other aspects of language, such as its syntactic structure. Moreover, this situation may be taken as representative of a very large class of non-linguistic problems, such as perceiving objects in the visual environment. In a formal way, this situation taps our ability to build internal representations that are isomorphic to the world structure, when this structure is no salient in the sensory input.

Language acquisition initially proceeds from auditory input, and linguistic utterances usually consist of sentences linking several words without clear physical boundaries. The question thus arises: how do infants become able to segment a continuous speech stream into words? Recent psycholiguistic research has identified a number of prosodic and phonological cues that may potentially help infants, but they provide only probabilistic information. The importance of prosodic and phonological cues in word discovery is further questioned by recent experimental studies showing that these cues are not necessary. For instance, Saffran et al.(1996) used an 

artifical language consisting of six trisyllabic words, such as ^tutibu^ and ^dutaba^. The words were read by a speech synthesizer in random order in immediate succession, without pauses or any other prosodic cues. Thus the participants heard a continuous series of syllables without any word boundary cues. In the following phase, they were asked to perform a forced choice test in which they had to indicate which of two items sounded more like word from the artificial language. One of the items was a word from the artificial language, whereas the other was a new combination of three syllables belonging to the language. Participants performed significantly better than would be expected by chance. This and other studies (e.g., Saffran st all. 1997) offer impressive support for the hypothesis that people are able to learn words forming a continuous speech stream without any prosodic or phonological cues for word boundaries.

The symbolic model of word segmentation developed by Brent and Cartwright(1996) offers a first way to accont for this ability. The authors construe segmentation as an optimization problem. The principle of the method is akin to establishing a list of all the possible segmentations of a give utterance (although the authors used computational tools which prevented the program from proceeding in this way). The choice between possible segmentations is then made in order to fulfil a number of criteria. These criteria are threefold(according to the somewhat simplified presentation by Brent(1996): min imize the number of novel words, minimize the sum of the lengths of the novel words, and maximize the product of the relative frequencies of all the words. The process of optimization is performed thanks to a statistical inference method, called the "minimum representation (or description) length" method. This method has been applied with some success for parsing phonetic transcripts of child-directed speech into words.

The performance observed in the Saffran et al. experiments can also be accounted for by connectionist models. Most of the connectionist models which address the word segmentation issue rely on the simple recurrent networok, or SRN, initially proposed by Elman(e.g., 1990; see also Cleeremans 1993). An SRN is a network which is designed to learn to predict the next event of a sequence. ........ Therefore, an SRN provides information useful for the parsing of a continuous speech flow into words, although words can only be revealed using a subsequent cluster analysis (for more recent models, see Aslin et al. 1996; Christiansen et al. 1998)

Prevalent approaches and consciousness

Let us consider now the conscious experience of the learner. Certainly this experience may differ from one subject to another. However, following description presumably captures the gist of everyone's experience. During the initial exposure to a new language, the same phenomenon presumably occurrs as with any unstructured sequence of items: the material is chunked into parts composed of a few perceptual primitives. These chunks form the discrete content of successive attentional focuses, and are typically short lived. As a confirmation of this assumption, if the speech flow is stopped at a given moment, a subject may be able to recall the most recent set of two or three syllables, but no more. ......

Another view

We now intend to show that seriously considering the subjective experience of the learner makes the unconscious algorithmic computations involved in symbolic or connectionist models of learning useless. Indeed, the content of subjective experience is endowed with a remarkable property, namely its self-organizing nature.

Let us return to the word segmentation experiment presented above. Imagine that the initial chunks of a few syllables composing the conscious experience of the learner, instead of being an epiphenomenon doomed to fulfill the subjective scene while the serious learning mechanisms are operating in the deep unconscious, are in fact the prefiguation of what will become, a few minutes later, the words of the language. Of more probably, they are part words, or straddle over word boundaries. How can these randomly drawn structurally irrelevant chunks, become the relevant words?

A response is provided by the application of the three following principles, which have been confirmed by a huge amount of experimental studies. 

First, the components of any conscious percept tend to be retrieved as a whole, and perceived again as a whole when these components, or a part of them, are displayed again in the nominal stimulus.....

Second, the phenomenon of unitization described above is short-lived. If there is no subsequent opportunity to perceive the components of a given percept, the properties described above tend to disappear, as a consequence of both natural decay and interference from the processing of similar material. However, unitization is strengthened by repetition.

The third principle is that unitization allows the new representation to become a component within a further conscious percept, and hence, due to the recurrence of the first principle, allows the formation of another, more complex representation. This means that if "tuti" has become a new perceptual unit due to its repeated perception, "tutibudu" may be a further conscious percept. If "tutibudu" reoccurrs frequently enough in the corpus, it will evolve as a new stable representation. ....

Concept of self-organizing consciousness

The above demonstration directly introduces the concept of self-organizing consciousness(SOC): the content of consciousness self-organizes, that is to say it becomes increasingly isomorphic to the world as a consequence of the interaction between its own properties and the properties of the world. 

The SOC framework is primarily a theory of learning. This theory of learning shares many aspects with the conventional associative theory of learning and memory. The basic phenomenon that allows extraction of the structure of the material is that the unitization resulting from the simultaneous perception of individuated components is strengthened through repitition and vanishes through forgetting and interference. The only point of departure with regard to the conventional theory is our focus on ^conscious^ percept. Indeed, we assume that conscious representation are the very stuff on which associative process operate. However, even of this principle is not commonly stated as such, it is in keeping with a widespread conception according to which associative learning and memory are nothing other than the by-products of ^attentional^ processing. For instance, many authors.... agree in claiming that associative learning is an automatic process that associates all the components that are present in the attentional focus at a given point(...). If one subscribes to the view that what constitutes the content of the phenomenal experience at a given moment is what is attended to at this moment, and vice versa(...), it then appears that our framework is akin to the contemporary conceptions of associative learning and memory. The novelty, in fact, is not in the principles involved, but in the dmonstration of their power in extracting the world structure.

However, the SOC framework goes far beyond the field of learning as it is usually defined. More exactly, by giving to any conscious experience a role in the formation of subsequent conscious experience, it opens up a new conception of the relations between the conscious and the unconscious.

In the mainstream of cognitive approaches, the function assigned to consciousness consists in making certain parts of cognitive functioning accessible. To quote a recent overview by Baars(1998), "Many proposals about brain organization and consciousness rflect a single underlying theme that can be labeled the 'theater metaphor.' In these views the overall function of consciousness is to provide very widespread access to unconscious brain regions." And elsewhere in the same paper: "A classical metaphor for consciousness has been 'bright spot'cast by a spotlight on the stage of a darker theater....Nearly all current hypothesis about consciousness and selective attention can be viewed as variants of this fundamental idea." Being conscious is seeing with the mind's eye. It is worth stressing that the very notion of access implies that consciousness is stripped of any major role. The state and operations that are accessed would still exist even if access were not possible. The unconscious mental life is fully autonomous. ....

※ Key ideas are yet to be unfolded from this point on...(this reader's comment for a later reading )

***

Is the SOC framework advantagesous with regard to the conventional approaches of cognition

...

About generalization

...

Generalization: the bottom-up approach

...

Generalization: the top-down approach

...

  - syntactic processing

  - Problem solving

  - Incubation

  - Decision making

  - ...