Assessment of weight-error derivative (WED) analysis

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Connectionist models are crucial for understanding cognitive processes.
• Lewandowsky's 1995 WED analysis provided a method for analyzing these models.
• Limitations exist in the original WED analysis for certain applications.

Purpose of the Study:

• To address the insufficiency of the original WED analysis.
• To propose an enhanced method for analyzing connectionist models.
• To improve the interpretability of connectionist model outputs.

Main Methods:

• Review and critique of the original WED (Weak Equivalence Difference) analysis.
• Development of an extended WED analysis incorporating vector norms.
• Comparative analysis of original and extended WED methods.

Main Results:

• The original WED analysis may not capture all relevant information.
• The proposed extension, including vector norms, offers a richer representation.
• Enhanced WED analysis provides deeper insights into connectionist model structures.

Conclusions:

• The extended WED analysis is a valuable improvement for connectionist model research.
• Incorporating vector norms enhances the diagnostic power of WED analysis.
• This refined method facilitates a more thorough understanding of neural network representations.