Oct 26, 2017 · newsletter

Bias Mitigation Using the Copyright Doctrine of Fair Use

Pirating a copyrighted song, video, or e-book to listen to the song, watch the movie, or read the book is an infringement of copyright (which can be severely fined). So how about pirating a song, video, or e-book to train machine learning models?

NYU Teaching and Research Fellow Amanda Levendowski proposes a legal approach to reducing bias in machine learning models. Biased data leads to biased models, she argues, and use of existing public domain data, most of which is over 70 years old, introduces biases from a time before, e.g., the civil rights movement or the feminist movement of the mid-20th century. The copyright doctrine of fair use can reduce bias by allowing wider access to copyrighted training data - an interesting and novel proposal.

Word embeddings (numerical representations of language) are biased. While words like “he” vs. “she” or “wife” vs. “husband” are gendered words and should fall on opposite ends on the “gender axis” (x). Words like “brilliant” should not (image taken from Bolukbasi et al.).

Fair use is a legal doctrine that operates as a defense to copyright infringement. This century-old exception essentially gives a get-out-of-jail free card to copiers who would otherwise be liable for copyright infringement. Fair use doctrine permits the copying of copyrighted material on the grounds that the type of copying is beneficial to the public and not unreasonably harmful to the copyright holder.

Courts have not yet ruled on fair use in the machine learning context, though it seems likely that they will need to soon. And once courts have ruled on a few such cases, those rulings will set a precedent for subsequent similar situations. If the precedent allows fair use, machine learning researchers will have the freedom to use copyrighted material with little fear of infringement liability.

Levendowski argues: (1) use of copyrighted materials to train machine learning models should be considered fair use and (2) the resulting availability of these copyrighted materials as training data will help mitigate bias in the models trained with that data.

Levendowski steps through each factor in the legal test for fair use and makes good arguments for why machine learning model training should be fair use. The strongest of these points is that the use is transformative, i.e., it is not used for its primary purpose. For example, a copyrighted music recording was made to be sold and listened to by humans, perhaps over the radio or on a smartphone. Using that same recording to train a model would be a very different use, and one that advances our understanding of music. Courts have held that this weighs in favor of fair use. Also notable is the argument that the copyright owners are not harmed by the use. Using the recording in the example above does not prevent the copyright holder from selling or licensing the recording.

A neural net model trained on romance novels generates captions for images; fair use might remove bias, but it surely entertains (we recommend you check out the authors’ alternate model trained on Taylor Swift lyrics).

We find Levendowski’s fair use analysis persuasive from a legal standpoint, but the benefits, though real, are overstated. Applying fair use may reduce bias, but it would be very unlikely to fix it (as suggested by the title How Copyright Law Can Fix AI’s Implicit Bias Problem). There’s no reason to believe, for example, that a set of recent textbooks would contain any less bias than Wikipedia (data used already during model training).

Bias has many origins, some rooted in legal and social practices. To reduce bias in machine learning models, we need to change these practices. We hope that the federal courts, which will inevitably be faced with these copyright infringement lawsuits, will consult and heed Levendowski’s analysis.

Read more

Oct 26, 2017 · newsletter
Oct 26, 2017 · post

Latest posts

Sep 22, 2021 · post

Automatic Summarization from TextRank to Transformers

by Melanie Beck · Automatic summarization is a task in which a machine distills a large amount of data into a subset (the summary) that retains the most relevant and important information from the whole. While traditionally applied to text, automatic summarization can include other formats such as images or audio. In this article we’ll cover the main approaches to automatic text summarization, talk about what makes for a good summary, and introduce Summarize. – a summarization prototype we built that showcases several automatic summarization techniques. more
Sep 21, 2021 · post

Extractive Summarization with Sentence-BERT

by Victor Dibia · In extractive summarization, the task is to identify a subset of text (e.g., sentences) from a document that can then be assembled into a summary. Overall, we can treat extractive summarization as a recommendation problem. That is, given a query, recommend a set of sentences that are relevant. The query here is the document, relevance is a measure of whether a given sentence belongs in the document summary. How we go about obtaining this measure of relevance varies (a common dilemma for any recommendation system). more
Sep 20, 2021 · post

How (and when) to enable early stopping for Gensim's Word2Vec

by Melanie Beck · The Gensim library is a staple of the NLP stack. While it primarily focuses on topic modeling and similarity for documents, it also supports several word embedding algorithms, including what is likely the best-known implementation of Word2Vec. Word embedding models like Word2Vec use unlabeled data to learn vector representations for each token in a corpus. These embeddings can then be used as features in myriad downstream tasks such as classification, clustering, or recommendation systems. more
Jul 7, 2021 · post

Exploring Multi-Objective Hyperparameter Optimization

By Chris and Melanie. The machine learning life cycle is more than data + model = API. We know there is a wealth of subtlety and finesse involved in data cleaning and feature engineering. In the same vein, there is more to model-building than feeding data in and reading off a prediction. ML model building requires thoughtfulness both in terms of which metric to optimize for a given problem, and how best to optimize your model for that metric! more
Jun 9, 2021 ·

Deep Metric Learning for Signature Verification

By Victor and Andrew. TLDR; This post provides an overview of metric learning loss functions (constrastive, triplet, quadruplet, and group loss), and results from applying contrastive and triplet loss to the task of signature verification. A complete list of the posts in this series is outlined below: Pretrained Models as Baselines for Signature Verification -- Part 1: Deep Learning for Automatic Offline Signature Verification: An Introduction Part 2: Pretrained Models as Baselines for Signature Verification Part 3: Deep Metric Learning for Signature Verification In our previous blog post, we discussed how pretrained models can serve as strong baselines for the task of signature verification. more
May 27, 2021 · post

Pretrained Models as a Strong Baseline for Automatic Signature Verification

By Victor and Andrew. Figure 1. Baseline approach for automatic signature verification using pretrained models TLDR; This post describes how pretrained image classification models can be used as strong baselines for the task of signature verification. The full list of posts in the series is outlined below: Pretrained Models as Baselines for Signature Verification -- Part 1: Deep Learning for Automatic Offline Signature Verification: An Introduction Part 2: Pretrained Models as Baselines for Signature Verification Part 3: Deep Metric Learning for Signature Verification As discussed in our introductory blog post, offline signature verification is a biometric verification task that aims to discriminate between genuine and forged samples of handwritten signatures. more

Popular posts

Oct 30, 2019 · newsletter
Exciting Applications of Graph Neural Networks
Nov 14, 2018 · post
Federated learning: distributed machine learning with data locality and privacy
Apr 10, 2018 · post
PyTorch for Recommenders 101
Oct 4, 2017 · post
First Look: Using Three.js for 2D Data Visualization
Aug 22, 2016 · whitepaper
Under the Hood of the Variational Autoencoder (in Prose and Code)
Feb 24, 2016 · post
"Hello world" in Keras (or, Scikit-learn versus Keras)


In-depth guides to specific machine learning capabilities


Machine learning prototypes and interactive notebooks


A usable library for question answering on large datasets.

Explain BERT for Question Answering Models

Tensorflow 2.0 notebook to explain and visualize a HuggingFace BERT for Question Answering model.

NLP for Question Answering

Ongoing posts and code documenting the process of building a question answering model.

Interpretability Revisited: SHAP and LIME

Explore how to use LIME and SHAP for interpretability.


Cloudera Fast Forward is an applied machine learning reseach group.
Cloudera   Blog   Twitter