DNNs May Determine Major Properties of Their Outputs Early, with Timing Possibly Driven by Bias

TL;DR

This paper shows that deep neural networks often determine their outputs early in inference, influenced by inherent biases, which has implications for bias mitigation and understanding model decision processes.

Contribution

It introduces the concept of early-stage decision-making in DNNs driven by biases, supported by diffusion model case studies, offering new insights into inference dynamics.

Findings

DNNs make early decisions influenced by biases

Bias type and extent affect early inference stages

Implications for bias mitigation and model interpretability

Abstract

This paper argues that deep neural networks (DNNs) mostly determine their outputs during the early stages of inference, where biases inherent in the model play a crucial role in shaping this process. We draw a parallel between this phenomenon and human decision-making, which often relies on fast, intuitive heuristics. Using diffusion models (DMs) as a case study, we demonstrate that DNNs often make early-stage decision-making influenced by the type and extent of bias in their design and training. Our findings offer a new perspective on bias mitigation, efficient inference, and the interpretation of machine learning systems. By identifying the temporal dynamics of decision-making in DNNs, this paper aims to inspire further discussion and research within the machine learning community.