Abstract
Trustworthy machine learning (TML) under imperfect data has recently attracted significant attention in data-centric machine learning and artificial intelligence. Real-world data often deviates from the ideal independent and identically distributed (i.i.d.) assumption, manifesting in three primary forms of imperfection: label-level (noisy labels), feature-level (adversarial examples), and distribution-level (out-of-distribution data). This paper synthesizes a decade of research into these challenges, presenting algorithmic solutions such as SIGUA, Co-teaching, and specialized OOD detection methods. Furthermore, it outlines emerging frontiers in TML, including the security of foundation models, federated learning, and causal reasoning.
Introduction
Standard Machine Learning (ML) paradigms are built upon the i.i.d. assumption, which posits that training and testing data are drawn from the same underlying distribution. In practical applications, this assumption frequently collapses. Data collected from the real world is inherently "imperfect," leading to models that are brittle, easily fooled, or prone to confident errors. To achieve true Trustworthy Machine Learning (TML), systems must be designed to explicitly detect, mitigate, or adapt to these imperfections rather than simply minimizing empirical risk on a potentially flawed dataset.
This research categorizes data imperfections into three distinct layers: the label, the feature, and the distribution. By addressing each layer, we can move toward AI systems that are robust enough for deployment in critical environments.
Addressing Label-Level Imperfection: Noisy Labels
Training data often contains incorrect labels due to human error, automated tagging glitches, or sensor noise. Standard deep learning models tend to "memorize" these labels, leading to poor generalization. Several strategies have been developed to counter this:
- SIGUA (Stochastic Integrated Gradient Underweighted Ascent): This approach utilizes a dual-gradient strategy. While standard gradient descent is applied to likely "clean" samples, gradient ascent is applied to identified noisy samples. This effectively "unlearns" the harmful information contained in incorrect labels.
- Co-teaching: This paradigm employs two neural networks trained simultaneously. Each network identifies "small-loss" samples—which are statistically more likely to be clean—and feeds them to its peer for training. This cross-pollination reduces the accumulation of error that occurs when a single model filters its own noise.
- Masking: By incorporating human cognition priors to constrain the "noise transition matrix," researchers can better estimate how labels are being flipped and correct for these transitions during the learning process.
Defending Against Feature-Level Imperfection: Adversarial Examples
Adversarial examples are inputs modified by small, intentional perturbations that are often imperceptible to humans but cause models to make catastrophic errors.
Research suggests that adversarial vulnerability often stems from models latching onto non-causal "style" features or spurious correlations. By treating adversarial robustness as a causal problem, we can attempt to align the distributions of natural and adversarial examples. By eliminating spurious correlation paths and forcing the model to focus on the causal "content" that determines a label, the model becomes significantly harder to deceive. Additionally, tools like SAMMD (Semantic-Aware MMD) use deep kernels to improve the statistical detection of these subtle adversarial perturbations.
Detecting Distribution-Level Imperfection: Out-of-Distribution (OOD) Data
Models deployed in the wild inevitably encounter data that differs significantly from their training set. Detecting these OOD samples is critical for safety.
- Watermarking and Reprogramming: Rather than the high cost of retraining, models can be "reprogrammed" by adding a static pattern (watermark) to inputs. This enhances the model's intrinsic ability to flag OOD samples.
- DivOE (Diversified Outlier Exposure): This strategy uses a limited set of auxiliary outliers to "extrapolate" and predict the characteristics of unseen OOD data.
- NegLabel: For Vision-Language Models (like CLIP), this method leverages vast text corpora to create "negative labels," helping the model better define the boundaries of its own knowledge.
Key Insights and Future Directions
A central theme in TML is the Memorization Effect: deep networks tend to learn simple, clean patterns early in training and only memorize noise later. Exploiting this temporal window allows for effective noise filtering. Furthermore, the shift from retraining to reprogramming offers a more scalable path for adapting foundation models to new environments.
Despite these advancements, several frontiers remain:
- Trustworthy Foundation Models: Large-scale models are vulnerable to "jailbreak" attacks and multi-modal noise where text and images are misaligned.
- Federated Learning: In decentralized settings, the inability to inspect raw data makes detecting malicious "poisoning" or unintentional noise significantly more difficult.
- Causal Understanding: Achieving true safety requires a mechanistic understanding of why models reason the way they do, moving beyond pattern matching toward causal inference.
Conclusion
Building trustworthy AI requires a transition from idealistic data assumptions to a framework that anticipates imperfection. By integrating causal reasoning, adaptive gradient methods, and robust outlier exposure, we can develop ML systems that are not only accurate but resilient to the complexities of the real world.
Summary updated