Radu Luchian(ov): Portfolio: CogSci
Dissertation Proposal

This document was part of my application to
the Carleton University PhD in Cognitive Science program;
my current research plan may be different.

Proposal for Ph.D. dissertation topic
Radu Luchianov
February 7, 2000

I plan to continue my research in Machine Learning, with the final goal of creating a psychologically valid implementation of a high level model of concept acquisition and processing - the MneMonic net. I have already spent close to six years thinking on this topic and implementing version after version, beginning with an Eliza-like, rule-based discourse parser/generator, then considering the physiologically valid connectionist approach, and finally supplementing the weak areas of that approach (volatility, slow 'learning curve', difficult isolation of representations) with the corresponding strong aspects of the symbolic/semantic approach. I hope that the scientific environment of Carleton's program in Cognitive Science -- with so many people interested in learning: language and knowledge acquisition -- will allow me to complete the implementation I am looking on.

The uses of such a tool are widespread from teaching to research, to documentation analysis, natural language programming, human-machine interaction, and so on. I don't think I have to tell a cognitive scientist much about the Why. So I'll go ahead with the How.

The concept of mnemonic devices is documented as early in time as the Greek philosopher/scientists and orators. They are complex analogies made between a simple, everyday domain (as the placement of the rooms in one's house) and a more complex, even abstract domain (as the order of ideas to be presented in a discourse). The general purpose these mental devices were (and still are) fulfilling is multifold: to improve one's long term memory, to help understand the relationships and interactions within a newly observed phenomenon, to simplify problem solving in general. At a lower level of description, mnemonic devices quicken retrieval and act as a powerful heuristic in (analogical) matching. This double power of the mnemonic devices is the main reason why I started studying the phenomena behind them and, ultimately, why I called the node of the concept network I developed (MneMonic). I implemented the general structure of a mnemonic device as a set of 'MneMonic nodes' (the node which is the basis of the concept network.)

At the implementation level, each node is an instance of a general class (MneMonic); each node takes only the slots and methods necessary for its activity. This node-creation mechanism is essential for the development and the performance of the network - and is in need of further development. Nodes are either active or inactive. Activity is meant here in a pseudo-connectionist way: it 'moves' along links. When the activity coming towards an inactive node passes a threshold, the node is activated. When the internal activity of an active node drops under another threshold, the node is deactivated. I keep the hysteresis positive, like that of a nervous cell. While active, each node 'gives away' activity along its links. Links can be unidirectional or bi-directional and at either end can be 'excitatory' or 'inhibitory,' but not at the same time. They are either direct links (activity passes through them simply weighed up or weighed down, depending on the sign at the far end), or functional nodes, which process the passed activity according to the rules given by the specific functional node.

The network has several 'modes' of operation:

The input to the network is a 'subject' string of text, a 'source' string, a validity value and a special sensory value. Not all input channels are used in every mode of operation. The 'subject' text is used to create leaf nodes (words) or add activation to them if they already exist. The 'source' string is a description of the source of the 'subject' string of text and is referenced 'as-is' in a special slot of each node. The sensory value is used to enforce or inhibit the most active node. The output is another string of text generated by the most active 'expressive' leaves of the most active node. Each node of the network can participate in 'higher' structures: sentences, paragraphs or even larger bodies of text, depending on the level of the sensory value at input time. These structures stand for facts, situations, hypotheses, falsities (facts known to be false), depending on the validity value at input time and depending on the sensory value at output time. Some nodes can become specialized, functional for the other nodes and the network in general, like the slot names in a concept frame paradigm (is, who, what, where, like).

Simple fact retrieval is guaranteed (complete with its reference) by the storage procedure. Further research and experimentation is necessary in order to achieve more consistent results in the more cognitive processes which I am targeting: generalization, deduction and analogical mapping are sometimes achieved through interaction among nodes higher in the hierarchy and functional nodes, but not consistently enough. Also I would like to experiment with leaf nodes at the lexeme level. In the current state, the network responds with too many errors (of discourse while reporting and of structure mapping while applying functional rules) due to the lack of accurate automatic connection among words from the same family (work, works, working, workable). My attempts at 'leaking activation' (partly activating each word that starts with the target letters) failed in the case of prefixing (load, reload) and put the network in a hyperactive state. Another problem looming ahead of me, which I still have to address is with homonyms and words with multiple meanings; using different ontology-markers as feature slots for the nodes extends the knowledge base exponentially.

An example experimental part which I plan to use as a validation of the model as a problem-solving device (chosen because I consider it to be the most complex cognitive activity), can be run separately on a set of human subjects (a group of students) and on different versions of the MneMonic implementation, in several phases:

  1. . preparation of the different versions of MneMonic: different copies of the implementation will evolve in different ways, by interaction with different human subjects on themes chosen by the latter; this will generate slightly different states in the MneMonic nets which I hope will account for different 'backgrounds' and 'states of mind' with which human subjects come to any real life situation;
  2. . psychological experiment: an experiment with human subjects: the subjects are presented with a situation and asked to find a solution (Pre-test:subjects are asked a series of questions to determine whether they know what's necessary for solving a target problem; Experiment: subjects read the text of the problem and try to find an answer under the 'thinking aloud' paradigm.)
  3. . simulation experiment: each of the versions of MneMonic will be 'tutored', if necessary, with the terms involved in the problem-texts, then the same texts as per human subjects are submitted to them; a tracing routine can be used to get data similar to the 'thinking aloud' data in human subjects.
  4. . the analysis of the data resulting from the second part of the experiment can be used to determine whether the mechanisms used by the implementation generate behavior comparable with the mechanisms used by human subjects, and if not, can still be used to modify or augment the set of mechanisms described in the MneMonic model.

[ Home ][ Current Projects ][ Portfolio ][ Pastime ][ Pseudonyms ][ CV ]