Generalization Bounds from Loss and Prediction Projections

20 concepts9 papers1 proposal

To understand why a machine learning model generalizes to unseen data, information theorists traditionally look at the mutual information between the model's internal weights $W$ and the training dataset $S$, denoted as $I(W; S)$. The intuition is straightforward: if the weights memorize too much specific information about the training set, the model will overfit.

However, when applied to modern deep neural networks, this classical framework breaks down. Because deep networks are heavily overparameterized, the weight space is incredibly high-dimensional. Worse, many standard training algorithms are deterministic; if the mapping from data to weights is a fixed function, the mutual information $I(W; S)$ can become infinite or "vacuous"—providing theoretical limits that are too large to be useful.

A recent wave of research solves this problem through a shared insight: stop looking at the high-dimensional weight space, and instead project the model's state down to low-dimensional functional spaces, such as predictions or losses.

The Intuition: Compression via Data Processing

Why does projecting model information onto lower-dimensional spaces yield tighter, more practical generalization bounds? The answer lies in a fundamental principle of information theory: the Data Processing Inequality (DPI).

The DPI states that if you process a variable, you can only lose information, never gain it. The pipeline of a machine learning model forms a clear data-processing chain:

$$\text{Training Data } (S) \longrightarrow \text{Model Weights } (W) \longrightarrow \text{Predictions } (\hat{Y}) \longrightarrow \text{Losses } (L)$$

By calculating mutual information at the level of predictions ($\hat{Y}$) or scalar losses ($L$) rather than weights ($W$), we are looking at a highly compressed representation of the model's state. This compression automatically filters out irrelevant weight-space variations—such as different permutations of hidden neurons or internal representations that yield identical decision boundaries. Consequently, the information tracked remains tightly focused only on what impacts the final output, rendering the resulting generalization bounds finite, stable, and computationally tractable.

[ High-Dimensional Weight Space (W) ]  --> Contains infinite/irrelevant variations
                │
                ▼ (Data Processing Inequality)
[ Low-Dimensional Prediction Space (Ŷ) ] --> Filters out non-functional weight changes
                │
                ▼ (Further Compression)
[ One-Dimensional Loss Space (L) ]     --> Retains only the essential error metrics

Projecting onto Prediction Space: Functional CMI

The first level of dimensionality reduction focuses on the model's outputs. In Information-theoretic generalization bounds for black-box learning algorithms, Hrayr Harutyunyan et al. introduced Functional Conditional Mutual Information ($f$-CMI). Instead of measuring what the weights know about the data, $f$-CMI measures the dependency between the model's discrete class predictions and the selection of training samples. Because it prioritizes functional behavior over internal states, this mechanism remains valid and finite even for completely deterministic learning algorithms.

Similarly, Mohamad Rida Rammal et al. proposed a leave-one-out conditional mutual information (loo-CMI) framework in On Leave-One-Out Conditional Mutual Information For Generalization. The authors demonstrated that calculating mutual information in the prediction space directly reflects the geometry of the loss landscape, providing non-trivial, practical guarantees for real-world image classification tasks.

Shifting the analytical lens to predictions also allows for more granular diagnostics. For instance, in Class-wise Generalization Error: an Information-Theoretic analysis, Firas Laakom et al. leveraged predictions and loss differentials to estimate generalization performance for specific sub-tasks or demographic groups, explaining why a model might generalize perfectly for one class while failing on another.

Extreme Compression: Loss and Loss-Difference Projections

If predictions compress a million-dimensional weight space down to a few dozen class probabilities, mapping those predictions to a scalar loss compresses the information down to a single dimension.

In A New Family of Generalization Bounds Using Samplewise Evaluated CMI, Fredrik Hellström and Giuseppe Durisi developed a framework based on evaluated Conditional Mutual Information (e-CMI). By analyzing the information encoded strictly within the scalar loss values a model produces on specific data points, the authors derived a "Binary KL" bound that avoids the vacuous results often encountered in high-dimensional weight spaces.

Taking this a step further, Ziqiao Wang and Yongyi Mao introduced the loss-difference technique in Tighter Information-Theoretic Generalization Bounds from Supersamples. Instead of looking at absolute losses, they projected high-dimensional model outputs onto a one-dimensional scalar representing the difference between training and test losses within a "supersample" framework. This projection retains the exact essence of what causes a generalization gap—the performance discrepancy between trained and unseen data—while discarding extraneous details. The authors later extended this 1D loss projection framework to distributed architectures in Generalization in Federated Learning: A Conditional Mutual Information Framework, establishing a formal information-theoretic link between client privacy and generalization.

What happens when a model perfectly fits the training data, entering the "interpolating regime" typical of modern deep learning? In Rethinking Information-theoretic Generalization: Loss Entropy Induced PAC Bounds, Yuxin Dong et al. showed that instead of computing complex mutual information, researchers can look at loss entropy—the randomness in the one-dimensional loss distribution. This reduction to a scalar variable makes generalization bounds exceptionally stable to estimate, even when the training error drops to zero.

Handling Complex Losses and Testing Frameworks

While early information-theoretic bounds relied on "pointwise" losses (where each data point's loss is calculated independently), modern algorithms like contrastive learning use multi-sample, relational losses.

To address this complexity, Yuxin Dong et al. expanded the loss-difference technique in Towards Generalization beyond Pointwise Learning: A Unified Information-theoretic Perspective. By combining a supersample framework with low-dimensional information metrics, the authors proved that the theoretical complexity of analyzing relational learning objectives does not scale exponentially with the number of samples involved in a single loss calculation.

Finally, to make these theoretical insights actionable during training, Yilin Lyu et al. introduced the Recognizable Information Bottleneck (RIB) framework. Instead of attempting to reconstruct high-dimensional data to estimate information retention, RIB frames the problem as a binary recognition task. It measures how well an optimal "tester" can distinguish whether a sample belongs to the training set based purely on its low-dimensional representation, providing a highly efficient, practical upper bound on the generalization gap.

Key Foundational Work

Information-theoretic generalization bounds for black-box learning algorithms (2021) — contributed 5 concept elements to this direction.
A New Family of Generalization Bounds Using Samplewise Evaluated CMI (2022) — contributed 2 concept elements to this direction.
On Leave-One-Out Conditional Mutual Information For Generalization (2022) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_loss_difference_projection_868b1363,
  title = {loss-difference projection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/loss-difference-projection/rdr_868b136301534ed4b12c6723e4577467},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_loss_difference_projection_868b1363_v1,
  title = {loss-difference projection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/loss-difference-projection/rdr_868b136301534ed4b12c6723e4577467/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake