Unveiling the Systematic Errors Committed by Machine Learning Models: A User-Friendly Investigation
Introduction:
In this blog post, we introduce Domino, a new approach for discovering systematic errors made by machine learning models. We discuss the importance of identifying slices, which are sets of data samples that share a common characteristic, on which models underperform. Understanding these slices can help practitioners make informed decisions, improve model robustness, and avoid safety or fairness consequences. However, some underperforming slices are hard to find, as they are not easily annotated or extracted from input data. To address this challenge, we present Domino, a novel method that leverages cross-modal embeddings to identify and describe underperforming slices. We demonstrate the effectiveness of Domino through an evaluation framework across various slice types, tasks, and datasets. With Domino, practitioners can gain valuable insights into their models’ performance and make better-informed decisions.
Full Article: Unveiling the Systematic Errors Committed by Machine Learning Models: A User-Friendly Investigation
The Introduction of Domino: Discovering Systematic Errors in Machine Learning Models
Machine learning models that achieve high accuracy in overall performance may still make systematic errors on specific subsets of data known as slices. A slice is a set of data samples that share a common characteristic or subject. Identifying these slices is crucial for model evaluation and decision-making in various fields, including safety-critical areas like medicine. In a diagnostic model, for example, if it underperforms on younger patients, it should not be deployed at a pediatric hospital. Additionally, knowing the underperforming slices can help improve and debug models, ensuring their robustness and reliability. However, some underperforming slices, referred to as “hidden” slices, are difficult to detect due to the lack of annotated metadata or easily extractable information.
Automatic Identification of Underperforming Slices with Domino
To address the challenge of identifying hidden underperforming data slices, a new method called Domino has been introduced. Domino is designed to discover coherent and underperforming slices using cross-modal embeddings. Cross-modal representation learning approaches allow for embedding both images and text in the same latent space, resulting in semantically meaningful representations. This enables Domino to generate natural language descriptions of the identified slices, providing practitioners with a better understanding of the commonalities among the examples within each slice.
The Three-Step Procedure of Domino
Domino follows a three-step procedure:
1. Embed: Domino encodes validation images and accompanying text using a cross-modal encoder, which creates a shared embedding space. This step leverages publicly available encoders specific to different domains such as natural images, natural videos, medical images, and amino acid sequences.
2. Slice: Using an error-aware mixture model, Domino identifies regions in the embedding space where errors are concentrated. These regions correspond to the underperforming slices.
3. Describe: To aid practitioners in comprehending the characteristics of each slice, Domino generates natural language descriptions by surfacing the text closest to the slice in the embedding space.
Auditing a Popular Classifier with Domino
To demonstrate the effectiveness of Domino, an off-the-shelf classifier called ResNet18, pretrained on ImageNet, was audited. The model’s ability to detect cars was examined, particularly focusing on identifying any interesting slices on which the model underperformed. Domino successfully discovered slices related to photos of cars taken from the inside and photos of racecars, which are rare subclasses of the target class. Depending on the use case, more training examples can be added to improve the model’s performance on these slices.
Evaluation of Slice Discovery Methods
In designing Domino, inspiration was drawn from other exciting slice discovery methods, including The Spotlight, GEORGE, and MultiAccuracy Boost. These methods share similar steps with Domino, but use different embeddings and slicing algorithms. However, previous evaluations of slice discovery methods have been predominantly qualitative, relying on practitioners’ judgment. The need for a quantitative approach to estimate the failure rate of an SDM (slice discovery method) became apparent. To address this, 1,235 deep classifiers were trained under specific constraints to underperform on predefined slices across three domains. This approach allowed for the estimation of the failure rate of SDMs.
Conclusion
The introduction of Domino offers a novel approach for discovering systematic errors in machine learning models. By leveraging cross-modal embeddings, Domino can effectively identify coherent, underperforming data slices and generate natural language descriptions for practitioners. The evaluation framework implemented in Domino allows for quantitative assessment of the method’s performance. This advancement in slice discovery methods is crucial for enhancing the robustness, reliability, and safety of machine learning models across various domains.
Summary: Unveiling the Systematic Errors Committed by Machine Learning Models: A User-Friendly Investigation
In this blog post, we introduce a new approach called Domino for identifying systematic errors made by machine learning models. We also discuss the importance of evaluating these errors and how it can impact model deployment in critical settings like medicine. A “slice” refers to a set of data samples that share a common characteristic. Models may underperform on specific slices, which can have safety or fairness consequences. However, identifying these slices can be challenging, especially when the commonality is not easily extracted from the data. We present Domino, a method that uses cross-modal representations to discover and describe underperforming slices. We also evaluate Domino and compare it to other existing methods.
