💻 computer science

Synthetic Designed Experiments for Diagnosing Vision Model Failure

This paper proposes Synthetic Designed Experiments for Representational Sufficiency (SDRS), a framework that applies statistical Design of Experiments to systematically audit vision models for specific failure modes (coverage gaps and spurious dependencies) and prescribe targeted synthetic data to efficiently rectify them.

Original authors: Krisanu Sarkar

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Krisanu Sarkar

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 trying to teach a robot to recognize different shapes. Currently, most people do this by throwing a massive amount of "fake" pictures at the robot, hoping that somewhere in that mountain of images, the robot will finally learn what it's missing. It's like trying to find a specific needle in a haystack by just throwing more hay at it.

This paper proposes a smarter way: Synthetic Designed Experiments (SDRS). Instead of guessing, the authors treat the robot like a patient in a lab and the image generator like a precise medical scanner.

Here is how it works, broken down into simple steps:

1. The Old Way: Throwing Darts Blindfolded

Right now, when people create fake images for robots, they usually just pick random combinations of things (lighting, background, angle, size). They hope this "random sampling" covers all the mistakes the robot might make.

The Problem: If the robot is already good at 95% of things, 95% of those random images are a waste of time. The robot learns nothing new from them.

2. The New Way: The "Doctor's Exam"

The authors suggest we stop guessing and start diagnosing. They treat the image generator as a machine with knobs (like "lighting," "background," or "angle").

Step A: The Structured Test (The "Designed Experiment")
Instead of making thousands of random images, the system creates a very small, carefully planned set of images. Think of it like a doctor giving a patient a specific set of tests to check their heart, lungs, and liver separately, rather than just waiting to see if they get sick.

They use a statistical trick called ANOVA (Analysis of Variance). In plain English, this is a way of measuring exactly which knob on the image generator causes the robot to get confused.

Step B: The Diagnosis (Finding the "Gaps")
The system looks at the robot's mistakes and sorts them into two specific buckets:

Bucket 1: The "I've Never Seen This" Problem (Type I Gap).
- Analogy: The robot is great at recognizing a red car, but it has never seen a blue car. It fails because it lacks experience with that specific color.
- The Fix: Generate more images with that missing color.
Bucket 2: The "Cheating" Problem (Type II Gap).
- Analogy: The robot is actually cheating. It thinks "if the background is green, it must be a car." It isn't looking at the car; it's looking at the grass. This is a "spurious shortcut."
- The Fix: Show the robot a car on a green background and a car on a red background to teach it that the background doesn't matter.

Step C: The Prescription
Once the doctor knows exactly what's wrong, they prescribe a tiny, targeted dose of medicine.

If the robot is missing experience (Bucket 1), they generate images to fill that gap.
If the robot is cheating (Bucket 2), they generate "counterfactual" pairs (two images that are identical except for the one thing the robot is cheating on) to force the robot to stop cheating.

3. What They Found (The Results)

The authors tested this on three different scenarios:

The Shape Test: They created a robot that was bad at recognizing shapes because it was "cheating" by looking at the position of the shape instead of the shape itself.
- Result: The diagnosis found the cheat. After the targeted fix, the robot's accuracy jumped from 49.9% to 79.0%.
The Segmentation Test: They tested a robot trying to cut out objects from complex backgrounds. The robot was cheating by looking at the complexity of the background instead of the object.
- Result: The diagnosis found the shortcut. The robot's performance improved from 0.948 to 0.998 (almost perfect).
The "Broken Machine" Test: They tested the system on a generator that was secretly broken (where changing the "style" of an image also accidentally changed its "size").
- Result: The system detected this hidden "leak" or entanglement, proving it can even tell if the image generator itself is flawed.

4. A Surprising Discovery: The "Whac-A-Mole" Effect

The authors found something interesting. When they tried to fix one type of cheating (e.g., making the robot ignore the background), the robot sometimes started cheating on a different thing (e.g., it started relying too much on the lighting).

They call this "Sensitivity Transfer."
Analogy: It's like fixing a leak in a boat by plugging one hole, only to find water rushing in through a different hole you didn't notice. This suggests that while the diagnosis is perfect, the cure (how we train the robot) needs to be more careful in the future.

The Bottom Line

This paper argues that we shouldn't just dump random fake data on AI. Instead, we should use scientific experiments to find exactly what the AI doesn't know or where it is cheating, and then generate only the specific images needed to fix those exact problems. It turns the process from "guessing and hoping" into "diagnosing and curing."

1. Problem Statement

Current pipelines for generating synthetic data in computer vision operate on an open-loop paradigm. They typically rely on heuristics (e.g., random domain randomization, text prompts, or difficulty scores) to generate large volumes of images, treating synthetic data as a cheap substitute for real data.

Key Issues Identified:

Inefficiency: When a model already handles 95% of a visual distribution, randomly generated data provides zero learning signal, wasting computational resources.
Lack of Diagnosis: Existing methods do not diagnose why a model fails or which specific scene factors (e.g., lighting, occlusion, viewpoint) drive those failures.
Misuse of Controllability: Synthetic generators (3D engines, diffusion models) possess unique, controllable parameters for scene factors, but this property is underutilized for targeted model auditing.

2. Methodology: Synthetic Designed Experiments for Representational Sufficiency (SDRS)

The authors propose SDRS, a framework that applies the statistical theory of Design of Experiments (DoE) to synthetic data generation. It treats the downstream vision model as a "black-box system" and the synthetic generator as an "experimental apparatus."

The framework operates in four distinct phases:

Phase 1: Designed Experiment

Instead of random sampling, SDRS uses fractional factorial designs to generate a small, structured set of synthetic images.

For $k$ factors, a Resolution IV design requires only $2^{k-p}$ images (e.g., 8 images for 5 factors) rather than the full combinatorial space ( $2^k$ ).
This ensures efficient probing of the model's sensitivity to specific factors while preserving the ability to estimate main effects.

Phase 2: Representational Audit via ANOVA

The generated probe images are passed through the downstream model ( $f_w$ ).

Metric: Task loss ( $\ell$ ) is computed for each image.
Analysis: A One-way ANOVA is performed on the grouped losses for each factor.
Output: An F-statistic ( $F_j$ ) is calculated for each factor to determine if the model's performance variance is significantly dependent on that specific factor. Holm–Bonferroni correction is applied to control for multiple comparisons.
Distinction: Unlike linear probing (which measures feature extractability), this measures prediction-level dependence, which is more relevant for failure diagnosis.

Phase 3: Gap Diagnosis

The audit results are cross-referenced with the task structure to classify failures into two actionable types:

Type I Gaps (Coverage Failures): The model fails significantly on underrepresented or unseen levels of a factor (e.g., missing orientations).
Type II Gaps (Spurious Shortcuts): The model relies heavily on a nuisance factor (e.g., background complexity) that correlates with the target but is not semantically relevant.
Decision Rule: If a factor shows a coverage drop, it is Type I. If it shows significant nuisance dependence without a coverage drop, it is Type II.

Phase 4: Targeted Prescription

Based on the diagnosis, specific synthetic data is generated to address the gaps:

For Type I: Generate diverse samples covering the deficient factor levels.
For Type II: Generate counterfactual pairs $(x_a, x_b)$ where only the nuisance factor changes, and apply an invariance penalty to the representation to force the model to ignore that factor.
Objective: The total loss combines task loss on real and Type I data, plus the invariance loss on Type II counterfactuals.

3. Key Contributions

Principled Diagnostic Tool: Introduces an ANOVA-based audit that decomposes model failure into per-factor contributions, transforming the synthetic generator into a structured evaluation instrument.
Unifying Framework: Demonstrates that existing approaches (Domain Randomization, Counterfactual Augmentation, Active Learning) are special, often suboptimal, cases of the SDRS framework.
Identification of "Sensitivity Transfer": The authors identify a critical open problem where suppressing one nuisance factor via invariance penalties can inadvertently amplify sensitivity to other nuisance factors (representational reallocation).

4. Experimental Results

The framework was validated across three experiments:

Experiment 1: Controlled Diagnostic (dSprites)

Setup: Shape classification with planted biases (shortcut: shape correlated with position; coverage gap: missing orientations).
Result: The audit correctly identified the position shortcut (Type II) and orientation gap (Type I).
Performance: Targeted data improved held-out accuracy from 49.9% to 79.0% (+29.1 points).
Observation: Removing the shortcut reduced the F-statistic for position from 2.1 to 1.0 (non-significant).

Experiment 2: Dense Segmentation (Procedural Scenes)

Setup: 128x128 RGB scenes with factors like lighting, background complexity, and occlusion.
Result: The audit detected a dominant shortcut on background complexity (Type II) and an occlusion coverage gap (Type I).
Performance: mIoU improved from 0.948 to 0.998.
Insight: While random sampling also reached high mIoU in this saturated benchmark, SDRS provided the diagnostic insight (identifying why the model failed) which is crucial for harder benchmarks.

Experiment 3: Generator Entanglement Detection

Setup: Comparing a perfect generator vs. an entangled one (where "style" implicitly changes object size).
Result: The ANOVA audit detected the leakage: the F-statistic for "style" increased by +4.7, while "size" decreased by -7.9, successfully identifying the generator's flaw without ground-truth labels.

5. Significance and Discussion

Shift in Paradigm: SDRS argues that synthetic data pipelines should be designed as structured experiments rather than random sampling processes.
Diagnostic over Aggregate Metrics: The primary value is not just the final accuracy gain, but the decomposition of errors. It tells practitioners exactly what to fix (e.g., "add more occlusion samples" vs. "remove background bias").
Sensitivity Transfer: A major finding is that invariance penalties can cause sensitivity transfer, where suppressing one dependency amplifies another. This suggests that future correction losses need to be more holistic to prevent this trade-off.
Future Work: The paper highlights the need for better correction losses and validation on diffusion-based generators, which have weaker factor separation than procedural 3D engines.

In conclusion, SDRS provides a rigorous, statistically grounded method to audit vision models, diagnose specific representational failures, and prescribe targeted synthetic data to efficiently close those gaps.