Evaluating a Model's Performance Discrepancy
Based on the description of the positional encoding method in the case study, evaluate the engineer's observations. Is the observed performance pattern (poor at absolute position tasks, strong at relative position tasks) consistent with the properties of the described encoding method? Justify your reasoning.
0
1
Tags
Ch.2 Generative Models - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Evaluation in Bloom's Taxonomy
Cognitive Psychology
Psychology
Social Science
Empirical Science
Science
Related
An attention mechanism incorporates positional information by applying a unique rotation to each query and key vector based on its absolute position in a sequence. The attention score between a query from position 't' and a key from position 's' is then computed. A key property of this rotation is that the dot product between the rotated query and key vectors is a function of the original vectors and the difference in their positions (t-s). Based on this information, what can be concluded about the attention scores produced by this mechanism?
Analysis of a Positional Encoding Method
Evaluating a Model's Performance Discrepancy