Value-laden Disciplinary Shifts in Machine Learning

TL;DR

The paper argues that progress in ML is shaped by value-laden disciplinary dynamics, not purely objective metrics. It introduces a framework where model-types organize research, with their dominance tied to evolving evaluation criteria and prerequisites that encode social values. The ImageNet case illustrates how data-rich and compute-rich environments, plus selective metrics, steer the field toward deep learning, often at the expense of concerns like privacy and environmental impact. The authors advocate disciplinary-level reforms (e.g., Green AI) and inclusive deliberation on who sets goals and standards, bridging philosophy of science with practical governance in data-driven research.

Abstract

As machine learning models are increasingly used for high-stakes decision making, scholars have sought to intervene to ensure that such models do not encode undesirable social and political values. However, little attention thus far has been given to how values influence the machine learning discipline as a whole. How do values influence what the discipline focuses on and the way it develops? If undesirable values are at play at the level of the discipline, then intervening on particular models will not suffice to address the problem. Instead, interventions at the disciplinary-level are required. This paper analyzes the discipline of machine learning through the lens of philosophy of science. We develop a conceptual framework to evaluate the process through which types of machine learning models (e.g. neural networks, support vector machines, graphical models) become predominant. The rise and fall of model-types is often framed as objective progress. However, such disciplinary shifts are more nuanced. First, we argue that the rise of a model-type is self-reinforcing--it influences the way model-types are evaluated. For example, the rise of deep learning was entangled with a greater focus on evaluations in compute-rich and data-rich environments. Second, the way model-types are evaluated encodes loaded social and political values. For example, a greater focus on evaluations in compute-rich and data-rich environments encodes values about centralization of power, privacy, and environmental concerns.

Figures (2)

• Figure 1: A toy rendition of two different ways that model performance can improve with data. Model A performs better than Model B in a low-data regimen, but Model B performs better in a high-data regime.
• Figure 2: A figure (reproduced from OpenAI amodei_hernandez_2018) depicting the change in amount of computational power used to train key results in machine learning. The amount of compute power is measured on the $y$-axis in petaflop/s-days, which translates to the number of operations (e.g. add, multiply) performed during model training divided by roughly $10^{20}$. Before 2012, the amount of compute power follows Moore's law and doubles at a 2-year rate. However, after 2012, the amount of compute power used increases much more rapidly -- doubling every 3.4 months. As we can see, 2012, the year that marked the rise of deep learning, also marked a drastic shift in the reliance on compute power.