Source-linked AI summary
Value-laden Disciplinary Shifts in Machine Learning
Ravit Dotan, Smitha Milli
TL;DR
The paper asks how values shape the machine-learning discipline and develops a philosophy-of-science framework for analyzing the rise of model-types. It argues that model-type shifts and evaluation criteria reinforce one another, while those criteria encode social, political, and environmental values.
Problem
How values shape machine learning’s disciplinary focus and development is a central question addressed by the paper.
Source: §1 · p. 2Method
The paper analyzes model-types through a conceptual, descriptive framework informed by philosophy of science.
Source: §1 · pp. 1–2Results
Takeaways & Limitations
Making disciplinary values explicit allows them to be examined and changed when deemed undesirable.
Source: §5 · p. 9Who should decide which values evaluation criteria further, and who should have a voice, remains an open question.
Source: §5 · p. 9
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
The paper shifts attention from values in individual machine-learning models to values shaping the discipline, arguing that disciplinary shifts are not objective. It outlines a philosophy-of-science framework and argues that model-type adoption and evaluation criteria reinforce one another and encode social, political, and environmental values.
1 INTRODUCTION
- The paper asks how societal, political, environmental, and other values shape machine learning’s focus and development, beyond efforts to address values in individual models.The authors argue that addressing undesirable values may require intervention at the disciplinary level, not only in particular models.Source: §1 · p. 1
- Disciplinary shifts framed as objective progress can conceal implicit values that become defaults, making recognition a prerequisite for intentional choices.Source: §1 · p. 1
- The paper develops a descriptive philosophy-of-science framework in which model-types such as deep learning, graphical models, and support vector machines organize machine-learning research.It compares model-types with Kuhn’s paradigms and Lakatos’s research programmes.Source: §1 · p. 1
- A model-type’s rise is self-reinforcing: it accompanies the rise of evaluation criteria suited to that type, as the deep-learning case illustrates.The authors argue that ImageNet’s influence included a shift toward evaluation in compute-rich and data-rich environments.Source: §1 · p. 1
- Evaluation criteria encode values: compute- and data-rich settings and predictive accuracy can further centralization while hindering privacy and environmental sustainability.Source: §1 · p. 1
- The paper offers a descriptive account of machine learning for philosophers of science and aims to make disciplinary values visible to machine-learning researchers.It presents its framework as a starting point for further analysis and as a way to revisit philosophy-of-science concepts.Source: §1 · p. 2
2 MODEL TYPES AS ORGANIZING AND GUIDING RESEARCH
Model-types organize research and shape its agenda, while commitments to them reinforce the tools, data, and compute that support their success. Their role resembles paradigms and research programmes in some ways, but they are not themselves theories.
2.1 Commitment to model-types
- Researchers commit to model-types as favored approaches whose successes are taken as evidence of potential generalizability, and these commitments organize research activity.Commitments also appear in workshops centered on particular model-types.Source: §2.1 · p. 2
2.2 How a commitment to a model-type influences research
- Commitments guide problem selection, constrain the solutions pursued, and promote the supporting tools and resources that a model-type requires.As many researchers commit to one type, the discipline and resources available to practitioners change.Source: §2.1 · p. 2
2.2.1 Problem selection.
- Because model-types have different strengths, commitment can skew problem selection toward areas where a favored type performs well.Researchers also work to improve that type in areas where it performs less well.Source: §2.2.1 · p. 2
2.2.2 Constraining the search for solutions.
- Commitment leads researchers to pursue improvements within their favored model-type rather than other types or entirely new ones.Figure 1 illustrates that performance rankings can depend on data availability: Model A does better with little data, while Model B does better with much data.Source: §2.2.2 · p. 2, Figure 1 · p. 3, §2.2.3 · p. 3
2.2.3 Prerequisites.
- Deep learning typically needs more data and compute than other model-types, and its growing popularity has promoted data, compute, and specialized technologies.Industry labs developed supporting software and specialized hardware as deep learning grew in popularity.Source: §2.2.3 · p. 3
2.3 Similarities and differences between model-types and traditional concepts
- Model-types resemble paradigms in efforts to expand scope and precision, and resemble research programmes in adjusting within a favored approach when problems arise.Researchers may tune hyperparameters or attribute a problem to missing prerequisites such as data or compute.Source: §2.3 · p. 3
- Unlike paradigms and research programmes, model-types are not themselves theories or straightforwardly composed of hypotheses, prompting reconsideration of traditional philosophy-of-science concepts.The paper introduces “model-type” as a concept for describing machine learning’s disciplinary organization.Source: §2.3 · p. 4
3 COMPARISON BETWEEN MODEL TYPES IS MODEL-TYPE-LADEN
Comparisons between model-types depend on standards that researchers committed to different types may dispute. The authors use deep learning’s rise to argue that ImageNet’s success did not alone establish its superiority and that model-type comparison is methodologically laden.
3 COMPARISON BETWEEN MODEL TYPES IS MODEL-TYPE-LADEN
- Model-type comparisons depend on judgments about important problems, success measures, and background assumptions, so they are not neutral or straightforwardly objective.ImageNet cannot by itself show that deep learning is better when the evaluation conditions favor it.Source: §3 · p. 5, §3.2 · p. 6
3.1 ImageNet
- 16.4% error rate: AlexNet won the 2012 ImageNet challenge, 10% lower than the runner-up, after no deep-learning models were submitted in 2010 or 2011.The best error rate in 2010 and 2011 was 25.8%.Source: §3.1 · p. 5
3.2 Does ImageNet justify the rise of deep learning?
- ImageNet’s win had more disciplinary impact than earlier deep-learning competition successes, and its database was much larger than PASCAL VOC’s.ImageNet’s 2010 database had 1,461,406 images and 1,000 classes, compared with PASCAL VOC’s 19,737 images and 20 classes.Source: §3.2 · p. 5
- ImageNet fostered hopes that deep learning would generalize to other vision tasks and fields, although the cited hypothesis of transfer to other vision tasks had not been systematically tested.The passage notes that ImageNet’s influence spread beyond computer vision.Source: §3.2 · p. 5
- 3.5 months: after 2012, compute used for state-of-the-art deep-learning training doubled at that rate, alongside growing belief in progress through larger datasets.Before 2012, compute doubled every two years, in correspondence with Moore’s law.Source: §3.2 · p. 6
- Deep learning’s rise brought evaluation in data-rich and compute-rich environments, despite privacy and environmental reasons to reject progress measures independent of those resources.The authors conclude that evaluation depends on considerations about which people committed to different model-types may disagree.Source: §3.2 · p. 6
3.3 Incommensurability?
- Machine-learning model-types are methodologically incommensurable because their adherents may disagree about comparison measures, without requiring semantic or perceptual differences.The authors distinguish this from semantic and ontological/perceptual incommensurability in philosophy of science.Source: §3.3 · pp. 6–7
4 COMPARISON BETWEEN MODEL TYPES IS VALUE-LADEN
The paper argues that comparing and prioritizing model-types is value-laden, including through the prerequisites they favor and the criteria used to evaluate them. Deep learning illustrates how compute and data requirements, and choices such as prioritizing accuracy, can advance some values while constraining others.
4 COMPARISON BETWEEN MODEL TYPES IS VALUE-LADEN
- Prioritizing model-types depends on considerations that encode social, political, and other values, so comparison is not neutral.The authors argue that these considerations are also ones on which people committed to different model-types may disagree.Source: §4 · p. 7
4.1 Prerequisites
- Model-type prioritization favors some prerequisites over others, making it value-laden when those prerequisites encode social and political values.The authors introduce deep learning’s compute and data requirements as examples.Source: §4.1 · p. 7
4.1.1 Compute Power.
- Deep learning’s GPU costs can advantage well-resourced companies and deepen dependence on major providers, while its computational demands also carry environmental costs.The authors argue that compute can promote centralization even if GPU costs fall; one large model’s training emissions exceeded those of one car’s average lifetime.Source: §4.1.1 · p. 7
4.1.2 Data.
- Large training datasets can advantage large companies, while collecting sensitive personal data raises privacy and individual-freedom concerns; unequal data availability can also be abused.The paper’s DeepNude example links data requirements to existing differences in what images are available about groups.Source: §4.1.2 · pp. 7–8
4.2 Evaluation Criteria
- Accuracy is not value-neutral: permissible prediction inputs can encode social values, and prioritizing accuracy over other virtues also reflects value choices.The paper notes that accuracy based on racial generalizations and accuracy achieved while avoiding them can carry different values.Source: §4.2 · p. 8
- Choosing accuracy over interpretability, or a model-type that favors one, reflects value priorities; greater interest in fairness and explainability may motivate alternatives to deep learning.Deep learning is described as difficult to scrutinize, while graphical models, decision trees, and support vector machines are typically easier to scrutinize.Source: §4.2 · pp. 8–9
5 CONCLUSION, AND VALUES IN DELIBERATION
Machine-learning progress cannot be reduced to choosing objectively better models: model-type popularity develops alongside evaluation criteria, and comparisons encode social, political, and environmental values. The authors call for making these values visible while leaving open who should decide which values research advances.
5 CONCLUSION, AND VALUES IN DELIBERATION
- The rise of a model-type brings corresponding evaluation criteria, and comparing model-types encodes social, political, and environmental values.The authors reject the simple account that disciplinary shifts merely reflect progress toward better models.Source: §5 · p. 9
- Making disciplinary values explicit allows them to be examined and changed when deemed undesirable.The paper cites Green AI’s proposal to report efficiency measures, such as accuracy relative to computational cost, rather than accuracy alone.Source: §5 · p. 9
- Who should have a say in which values machine learning advances, and which criteria matter for evaluating model-types, remains an open question.The authors raise whether laypersons, among others, should have a voice in these decisions.Source: §5 · p. 9