Evaluating a System Upgrade Strategy
A research lab developed a machine translation system a decade ago. It paired a moderately effective neural network with a highly complex and computationally expensive search algorithm to achieve acceptable translation quality. The lab now has the resources for a major upgrade and is considering two options:
- Option A: Dedicate all resources to designing an even more sophisticated and intricate search algorithm, keeping the original neural network.
- Option B: Dedicate all resources to training a new, exceptionally powerful and large neural network, but replace the complex search algorithm with a much simpler, computationally cheaper one.
Which option represents a more modern and effective strategy for significantly improving the system's performance? Justify your choice by explaining the relationship between the capabilities of the neural network and the requirements of the search component.
0
1
Tags
Ch.5 Inference - 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
Shift in Research Focus Away from Inference with the Rise of Deep Learning
A research team is building a system to generate complex, multi-sentence text. They find that when they replace their original, moderately effective neural network with a new, exceptionally powerful one, they can drastically reduce the complexity of their search algorithm (e.g., considering only the single best next word at each step instead of many alternatives) and still achieve state-of-the-art results. What is the most accurate analysis of this outcome?
Resource Allocation for a Language Generation System
Evaluating a System Upgrade Strategy