🤖 machine learning

Conveyance: A Versatile Framework for Learning in Structured Class Spaces

The paper introduces \textsc{Conveyance}, a versatile classification framework and loss function that leverages graph-like relationships between classes to improve model performance in structured spaces like hierarchical classification and ordinal regression, outperforming or matching specialized baselines without requiring complex joint distributions or manual tuning.

Original authors: Yasser Taha, Grégoire Montavon, Nils Körber

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yasser Taha, Grégoire Montavon, Nils Körber

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to recognize animals. You show it a picture of a cat, and you tell it, "This is a cat."

In the world of standard machine learning, the robot's teacher (the "loss function") usually works like a strict, all-or-nothing judge. If the robot guesses "cat," it gets a gold star. If it guesses "dog," "bird," or even "tiger," it gets a red "X." To this standard judge, guessing "dog" is just as bad as guessing "tiger." It treats every mistake as equally terrible.

But in the real world, mistakes aren't all equal.

If the robot thinks a cat is a tiger, that's a "near miss." They are both cats, they look similar, and they live in the same neighborhood. It's a small error.
If the robot thinks a cat is a truck, that's a huge, absurd error.

The paper introduces a new teaching method called CONVEYANCE. Think of it as upgrading the robot's teacher from a strict judge to a wise mentor who understands the "neighborhood" of the world.

The Core Idea: The "Plausible Neighborhood" Map

Instead of just saying "Right or Wrong," CONVEYANCE lets you give the robot a simple map (called a Boolean Matrix). This map tells the robot: "If you see a cat, it's okay if you think it's a tiger or a lion, because they are neighbors. But it is NOT okay if you think it's a truck."

You don't need to write complex math formulas or probability theories to make this map. You just draw lines connecting things that make sense together.

How It Works: The Two-Step Safety Net

The new method uses a special "scorecard" (a loss function) that checks two things at once, like a security guard with two rules:

Rule A (The Specific Check): "Did you pick the exact right answer?" (This is the standard check).
Rule B (The Neighborhood Check): "Did you at least pick something from the right neighborhood?" (This is the new superpower).

If the robot guesses "Tiger" for a picture of a "Cat," Rule A might give it a small penalty, but Rule B says, "Wait! Tiger is in the Cat neighborhood! That's a good guess!" The robot gets a much better score than if it had guessed "Truck."

This helps the robot learn that the world has structure. It stops treating "Cat" and "Truck" as just two random labels and starts understanding that "Cat" and "Tiger" are close friends.

Why This Matters: Three Real-World Examples

The paper tested this "Wise Mentor" approach on three different types of problems where standard methods usually struggle:

1. The "Noisy Label" Problem (Label Asymmetry)

The Scenario: Imagine a student taking a test where the teacher accidentally writes the wrong answer key for some questions. Maybe the teacher thinks "Cats" are actually "Dogs" 60% of the time.
The Old Way: The robot gets confused. It tries to memorize the wrong answers because the teacher is shouting them the loudest.
CONVEYANCE: The robot looks at the map. It knows, "The teacher says 'Dog', but I know 'Dog' is a neighbor of 'Cat'. I'll bet the teacher made a mistake and it's actually a Cat." It ignores the noise and learns the truth.

2. The "Age Guessing" Problem (Ordinal Regression)

The Scenario: Guessing someone's age. If you guess a 30-year-old is 31, that's a tiny mistake. If you guess they are 60, that's a huge mistake.
The Old Way: Standard methods treat "31" and "60" as just two different wrong answers. They don't care that 31 is closer to 30.
CONVEYANCE: The map tells the robot, "If the answer is 30, then 29, 30, and 31 are all acceptable 'neighborhood' guesses." This teaches the robot to be smooth and continuous, rather than jumping wildly between ages.

3. The "Family Tree" Problem (Hierarchical Classification)

The Scenario: Identifying birds. There are many species of "Warblers." If the robot sees a specific Warbler it has never seen before, but it knows it belongs to the "Warbler" family, it should guess "Warbler."
The Old Way: The robot might get stuck trying to guess the exact species name and fail completely if it hasn't seen that specific bird before.
CONVEYANCE: The map groups all Warblers together. Even if the robot doesn't know the exact species, it knows the "Warbler" family is the right neighborhood. It generalizes better to new, unseen birds.

The Big Takeaway

The authors claim that CONVEYANCE is a "one-size-fits-all" tool. You don't need to build a new, complicated robot for every specific problem. You just need to give it a simple map of what things are related to what.

It's fast: It runs just as quickly as the standard methods.
It's flexible: It works for noisy data, age guessing, and family trees.
It's smart: It understands that in the real world, some mistakes are forgivable, and some are not.

In short, CONVEYANCE stops treating the world like a list of random, unrelated items and starts treating it like a connected map, making the AI smarter, more robust, and less easily confused by mistakes.

Technical Summary: Conveyance

Problem Statement

Despite rapid advancements in deep learning architectures, standard training objectives, particularly cross-entropy (CE), remain largely "structure-agnostic." These losses assign scalar penalties based solely on label matching, treating all misclassifications as interchangeable. This assumption fails in two critical regimes:

Label Asymmetry: Scenarios where the relationship between observed labels and true classes is structurally non-uniform, such as annotation bias, underdiagnosis in medical imaging, or Multiple Instance Learning (MIL) where positive bags contain negative instances.
Semantic and Ordinal Continuity: Domains where categories represent points on a continuum (e.g., age, disease stages, taxonomic hierarchies). In these cases, a "near-miss" error is semantically less costly than a random error, yet standard CE penalizes them equally, leading to overfitting and fragmented latent spaces.

Existing remedies often address these issues in isolation. Soft-label methods rely on symmetric, pre-specified distributions that cannot capture asymmetric or graph-structured relations. Structured Output SVMs and energy-based models offer principled margins but require non-differentiable inference over exponentially large output spaces. Contrastive losses require large batches and sensitive pair mining. No existing framework combines structural expressiveness, discriminative power, and computational simplicity comparable to cross-entropy.

Methodology: Conveyance

The authors propose CONVEYANCE, a classification framework designed for structured class spaces. It consists of two primary components:

1. Knowledge Encoding (The Matrix $Q$ )

Instead of defining complex joint distributions or utility matrices, the user specifies a Boolean matrix $Q \in \{0, 1\}^{C \times C}$ .

$Q_{ct} = 1$ indicates that it is plausible for an instance labeled $t$ to actually belong to class $c$ (or vice versa).
This matrix encodes problem-specific knowledge, such as false-positive risks, taxonomic hierarchies, or ordinal neighborhoods.
From $Q$ , a set of plausible classes $S = \{c \in \mathcal{C} : Q_{ct} = 1\}$ is derived for a given label $t$ .

2. The Conveyance Loss Function

The loss function operationalizes the matrix $Q$ by maximizing two separate margins over distinct class partitions:
$\ell(p, t) = \log \left( 1 + \alpha \cdot \frac{1 - p_t}{p_t} + \beta \cdot \frac{1 - p_S}{p_S} \right)$
Where:

$p_t$ is the probability of the target class.
$p_S = \sum_{s \in S} p_s$ is the total probability of the plausible set.
$\alpha$ and $\beta$ are hyperparameters.

Mechanism:

The first term ( $\alpha$ ) rewards allocating probability to the target class over non-targets (standard odds).
The second term ( $\beta$ ) rewards allocating probability to the plausible set $S$ over the implausible set $N = \mathcal{C} \setminus S$ .
By tuning $\alpha$ and $\beta$ (e.g., $\alpha=0.1, \beta=10$ ), the model prioritizes concentrating mass on the plausible set $S$ , using the specific target $t$ as a tie-breaker.

Theoretical Properties:

Monotonicity: The loss decreases if probability mass is reallocated to increase both $p_t$ and $p_S$ .
Partial Convexity: The loss is convex with respect to the logits of non-plausible classes ( $z_n$ for $n \in N$ ) when plausible logits are held constant.
Efficiency: The loss can be reformulated using Log-Sum-Exp (LSE) operations on logits, ensuring numerical stability and a computational cost identical to standard cross-entropy ( $O(B \cdot C)$ ).

Key Contributions

A Unified Framework: CONVEYANCE provides a single objective function that handles label asymmetry, ordinal regression, and hierarchical classification without requiring architectural changes or generative models.
Tractable Knowledge Encoding: The use of a Boolean matrix $Q$ allows practitioners to encode complex structural priors (graphs, hierarchies) more easily than full probability distributions, while being more expressive than fixed parametric families.
Theoretical Guarantees: The paper establishes the monotonicity and partial convexity of the loss, providing theoretical grounding for stable optimization.
Versatility: The method recovers standard cross-entropy as a limit case ( $\beta=0$ ) but offers significantly greater flexibility for structured spaces.

Experimental Results

The authors evaluated CONVEYANCE across three domains, consistently matching or exceeding specialized state-of-the-art (SOTA) baselines while maintaining the computational cost of cross-entropy.

1. Label Asymmetry (Annotation Bias & MIL)

Synthetic Noise: On CIFAR-10 and CIFAR-100 with column, block, and asymmetric noise, CONVEYANCE outperformed the generative "Masking" framework [17] and standard CE. Notably, on CIFAR-100 with asymmetric noise, it surpassed the clean-data CE baseline by +3.56 percentage points, suggesting the structural prior acts as an implicit regularizer.
Multiple Instance Learning (MIL): On five tabular MIL benchmarks and the Camelyon-16 whole-slide image benchmark, CONVEYANCE achieved SOTA accuracy and AUC. On Camelyon-16, it improved AUC by 1.8 points over the next-best method (CAMIL), despite using no problem-specific architectural specializations (e.g., feature distillation or prototype clustering).

2. Ordinal Regression (Age Estimation)

Evaluated on CACD2000, CLAP2016, and UTKFace.
CONVEYANCE achieved the best Mean Absolute Error (MAE) on all datasets compared to soft-prior methods (Soft Labels, DLDL) and ordinal-specific architectures (ORCNN).
On the noisy CLAP2016 dataset, it reduced MAE by 0.17 over the strongest baseline (DLDL).
Qualitative analysis showed that unlike CE, which fragments the latent space, CONVEYANCE organizes embeddings into a smooth, continuous manifold reflecting the semantic progression of aging.

3. Hierarchical Classification (Zero-Shot Genus Generalization)

Evaluated on CUB-200-2011 (bird species).
In a zero-shot setting where species from specific genera were withheld during training, CONVEYANCE achieved the highest OOD genus accuracy (85.1%), outperforming CE+SupCon (84.1%) and CE+Triplet (80.9%).
The method successfully organized congeneric species into coherent clusters in logit space, demonstrating effective transfer of taxonomic structure to unseen classes.

Significance and Claims

The paper claims that CONVEYANCE offers a unified, differentiable, and architecture-agnostic alternative to the complex inference required by structured output frameworks (like SSVMs). Its primary significance lies in lowering the barrier to incorporating domain structure into supervised learning for tasks where class topology is known but difficult to express via generative modeling.

The authors emphasize that CONVEYANCE achieves these results with minimal hyperparameter tuning (often a single default configuration of $\alpha=0.1, \beta=10$ ) and no increase in computational complexity over standard cross-entropy.

Limitations Acknowledged:

The method relies on the correctness of the input matrix $Q$ ; incorrect or incomplete masks may reinforce spurious assignments.
The hyperparameter $\alpha$ scales with the number of classes, requiring recalibration for large-vocabulary datasets (e.g., ImageNet).
Constructing a meaningful $Q$ for fine-grained or large-scale datasets requires structured domain knowledge or ontologies that may not always be available.

The paper concludes by envisioning future work in deriving $Q$ automatically from auxiliary signals (e.g., annotator disagreement, uncertainty estimates) and extending the framework to sample-level transition masks.