Outline of Four Computational Models
Outline of Topics
The following is an outline of an article which deals with different decision strategies in neuropsychology. Please consult the following reference for a more detailed look:
Long, D., Graesser, A., & Long, C.J. (1988). Four computational models for investigating neuropsychological decision-making, in The Cognitive Approaches to Neuropsychology, Edited by Williams & Long, Plenum, 3-26.
Expert knowledge-based systems
Exemplar-based reasoning models
Neuropsychological decision-making - investigated from perspective of 4 models:
Neuropsychologists initially interview, observe behavior, and test to obtain information about symptoms, levels of cognitive and emotional functioning, as well as occupational and situational factors.
A. Why use Multivariate Analyses?
B. We have conducted several obverse factor analyses.
- Most familiar
- Techniques - multiple regression, factor analysis, discriminant analysis
- Obverse Factor Analysis - goal to cluster similar and dissimilar subjects Interested in how subjects are associated with latent factors (type Ss)
- Six major assumptions
- There are a set of indicators (I1..In)
- There are a set of patients (P1..Pn)
- Each indicator has a value and are weighted equally. (if an indicator is associated for one patient then it will be for all)
- Patients are selected on basis of a set of P a priori dysfunctions(D1..Dn) (dysfunctions are mutually exclusive)
- Matrix of correlation's computed for all possible pairs of patients (patients with similar profiles will be highly correlated)
- Set of latent factors (F1..Fn) is statistically discovered (patients load on factors to extent they are similar)
- Factors interpreted in two ways
- Do patients who load on a common factor share the same dysfunction? -- group classification
- Do patients who load on same factor share a common subset of indicators? -- profile similarity
- Patients classified based on medical indicators (EEG, CTscan, Surgery).
- Found seven-factor solution, but was not a good fit to data.
- Factor loadings were low (.01-.06)
- The left, right and generalized patients were not homogeneous.
- Groups were recomposed and a fit was obtained with seven factors for the left but not the right or generalized.
- Again concluded - groups were composed of patients who were not similar.
- Several reasons the results did no suggest a strong association between patients
- Patient sampling
- high functioning patients - low couldn't complete
- high level - may reflect individual rather than group differences
- missing data can be a serious problem in multivariate analyses
- Sensitivity of tests
- Tests were not equally sensitive or not to specific functions
- when include inconsistent indicator increase unexplained variance
- Important data does not lend itself to quantification (symptoms, history)
- Need to carefully consider which indicators to include in analyses.
- Multivariate analysis is useful as a discovery method or as a confirmatory technique but is limited to use with quantified variables and with variables which do not interact in antagonistic or asymmetrical ways.
C. We have turned to expert systems in order to investigate decision process.
Expert Knowledge Based Systems
An intelligent interactive computer system that uses a large body of knowledge and a set of inference procedures to make decisions that require significant human expertise (Barr, Cohen, & Feigenbaum, 1981)
- System consists of set of facts, rules, and heuristic control processes
- Facts - static entities such as object properties, attributes, relations
- Rules - usually expressed in form of IF ____ THEN ____ statements
- Heuristic control processes are way rules are combined to operate on facts and generate new facts.
- Knowledge engineer's (designer of system) job is to extract the knowledge and reasoning strategies from the expert so that they can be translated into rules or information structures.
- Expert can be trained to think aloud while problem solving.
- Expert can be questioned about how conclusions are reached
- Expert system for neuropsychology decision making features:
- There is a set of relevant dysfunction indicators (I1..In)
- There is a Set of dysfunctions (D1..Dn)
- There is a knowledge base consisting of a set of rules.
Rules - expressed in form of IF ____ THEN ____ statements.
- Expert systems have inference procedures.
- Rules chained together to allow system to consider a set of hypothesis and attempt to confirm each hypothesis in the set.
- Confirming a hypothesis involved implementing decision rules
- The system begins by attempting to confirm a rule by searching through knowledge base.
- This process continues until the system can confirm one or more rules and converge on a solution.
- Decisions are products of a dialogue between computer and user. (User inputs values of relevant indicators as requested)
- Begins with inputting of start values.
- System searches knowledge base to generate whatever additional indicators are needed and requests information from user.
- When the system converges on a dysfunction, the dysfunction may have certainty factors associated with it. (confidence in decision)
B. We designed a small prototype expert system to discover information about neuropsychological decision-making.
- Ideal expert system would divide neuropsychological services into 3 tasks
- Describe neurological dysfunctions,
- Identify neurological dysfunction,
- Make recommendations for treatment.
- We concentrated on the identification task first.
- Symptoms - case history
- Situational factors - case history
- Emotional distress - MMPI, Cornell
- Cognitive deficits - neuropsychological tests
C. Seven Advantages of Expert System over multivariate analyses
- Only request or compute values of relevant indicators
- all values are not considered equal
- Can handle very complex interactions among indicators
- can handle variables-interact antagonistically or asymmetrically
- cognitive dysfunction - evaluated differently with sig. emotional
- Can identify important qualitative information that can be scaled
- Low frequency events are better represented - visual field cuts
- would be excluded from multivariate methods
- System can trace reasoning by virtue of its symbolic representation
- rules are presented in natural language and easily interpreted
- multivariate - computational rules are statistical
- System can query the user for additional information or recommend the administration of additional tests.
- System can converge on a solution in spite of missing data.
D. Expert system can be important clinical support tool.
- If expert system arrives at the same conclusion, the neuropsychologist can feel confident in the decision.
- If it arrives at a different one, they can trace the systems reasoning
Exemplar-based Reasoning Models
A. Model that produces decisions by matching a target problem to all previous specific cases or episodes, called exemplars.
- System makes a good match or indicates no good matches in memory.
- called reasoning from memory
- Similarity-based reasoning model - one which assess specific memory episodes in several operations (Stanfill & Waltz, 1986)
- First- count No. of times features occur in association with decision (how often errors on aphasia occur with left temporal lesion)
- Second- use feature counts to produce metrics. -- dissimilarity metric.
- Also metrics for weighting features
- Metrics to calculate value similarity
B. Exemplar-based model have several advantages
- reasoning from memory may be an important strategy used by NP
- because information is never lost in the system - important tool.
- Can form Ho: based on a single case in memory.
- Can account for low frequency information.
Connectionist Models or Parallel distributed processing model
Computational power of the parallel processing computer had made research involving these models more feasible. Particularly suited to pattern recognition problems.
- Connectionist models are composed of a large number of units.
- First - set of input units (dysfunction indicators)
- Second- set of output units (dysfunction classification)
- Third - set of hidden units - represent latent features that are predictive of the dysfunction.
- Units have activations associated with them (a real No. indicating the extent the specific input is evaluated.
- Information is represented by a pattern of activation across input, output, and hidden units.
- All units are connected by links
- Each link has a weight associated with it.
- Weights change as the system learns the domain representation.
- Learning involves adjusting connection weights to new values after presentation of input and output pairs.
- The model has learned when it consistently produces the correct output in response to a pattern of inputs.
- Don't know whether this model provides a viable model of neuropsychological decision making.
- knowledge acquisition process is automatic.
- performance does not depend upon the expertise of knowledge engineer.
- Can account for high-order interactions.