VisAlign Dataset Benchmark

• VisAlign Dataset is a benchmark measuring AI–human visual alignment in image classification by comparing model outputs with empirical human judgments.
• It features clearly defined classes and diverse image scenarios including in-class, out-of-class, and uncertain conditions for comprehensive evaluation.
• The dataset employs distance-based and reliability metrics, such as Hellinger distance and RS_c, to assess model performance and safety in open-set recognition tasks.

VisAlign is a benchmark dataset designed to measure the degree of alignment between AI models and humans in visual perception, specifically within the context of image classification. It operationalizes "AI–human visual alignment" as the agreement between a model's classification—including its option to abstain from classifying—and the empirical distribution of human perceptual judgments. The dataset serves both as a proxy for evaluating AI safety in classification systems and as a rigorous testing ground for open-set recognition and abstention-capable vision models (Lee et al., 2023).

1. Motivation and Design Objectives

VisAlign addresses the challenge of evaluating how closely black-box deep learning models mirror human perceptual decisions, a central concern for AI alignment and safety. Since most large-scale models lack explicit controls for user-intended behavior, their similarity to human choices serves as a practical proxy for safe operation.

The benchmark's core design objectives are:

• Clear class definitions: All target classes are chosen to be unambiguous for human annotators.
• Class familiarity: Only classes well-known to the average human are included.
• Comprehensive scenario coverage: The dataset spans in-class (familiar), out-of-class (foreign/novel), and intermediate ("uncertain") visual recognition scenarios.
• Gold-standard human labels: Every sample, including highly degraded examples, is provided with a label grounded in empirical human judgments.

2. Dataset Structure and Composition

VisAlign is organized into one train set and open/closed test sets, each comprising carefully curated image groups. The train set contains 12,500 images (10 animal classes × 1,250 images each) drawn from ImageNet-21K, with a 10% validation split.

Test sets (900 images per split) are structured as follows:

Group	Category (Count)	Content Description
Must-Act	1 (100), 2 (100), 3 (100)	In-class; original, implausible backgrounds, and adversarial imgs
Must-Abstain	4 (100), 5 (100), 6 (100), 7 (100)	Out-of-class: foreign objects, hybrids, near-misses, stylized
Uncertain	8 (200)	Random crops + synthetic corruptions covering multiple severities

Synthetic samples—hybrid creatures and implausible backgrounds—are generated by Stable Diffusion and MagicMix. Adversarial examples deploy FGSM perturbations. Out-of-class images include selections from ImageNet-1K, Describable Textures, Caltech-101, DomainNet, and ImageNet-R, with stylized domains adding to distributional diversity. The "uncertain" group samples are derived by applying the ImageNet-C corruption suite and random cropping to ImageNet-21K.

3. Annotation Methodology and Quality Controls

Must-Act and Must-Abstain images receive "indisputable" labels, with authors verifying single-object content for clarity. For Uncertain images, the annotation process is fundamentally crowd-sourced: each sample is labeled by 134 independent Mechanical Turk workers, who select from ten animal classes or an explicit abstain category ("None of the 10, uncertain, or unrecognizable").

Quality control measures include:

• Known-distractor images embedded in each batch to identify inattentive annotators.
• Filtering: Any annotator failing to abstain on clear distractors is removed.
• Intra-annotator consistency: κ = 0.91 (duplicate image checks).
• Inter-annotator agreement: Fleiss’ κ = 0.80 (substantial agreement).

Sample size for Uncertain group labeling is determined by survey sampling theory, targeting a 0.05 half-width for 95% CI on $p \approx 1/11$, which calculates to $n \geq 132.2$, rounded up to 134 annotators per image. Reliability is further confirmed with Cronbach’s α = 0.88, evidencing good internal consistency.

4. Alignment Metrics and Scoring Formulations

VisAlign introduces rigorous alignment scoring grounded in both probabilistic proximity and decision reliability.

• Distance-based alignment: For each image, with $P$ the empirical human label distribution and $Q$ the model output ($11$-dim., 10 classes + abstain), alignment is quantified via the Hellinger distance:

$$h(P, Q) = \frac1{\sqrt2} \|\sqrt{P} - \sqrt{Q}\|_2 = \frac1{\sqrt2} \sum_{i=1}^{11} (\sqrt{p_i} - \sqrt{q_i})^2,$$

where $h \in [0,1]$ and smaller values denote superior AI–human alignment.

• Reliability with abstention: For a penalty parameter $c \geq 0$ (the cost of a misclassification), the per-sample score $RS_c(x)$ is case-based:
  • Must-Act: correct $+1$, wrong $-c$, abstained $0$
  • Must-Abstain: abstain $+1$, mistaken class $-c$
  • Uncertain: partitioned by a threshold $\lambda$ (default 0.5) on human agreement; assigned to Must-Act or Must-Abstain formula accordingly.

Total reliability is $RS_c = \sum_{x\in\text{test}} RS_c(x)$. Large $c$ renders any mistake highly penalizing, reflecting "fail-safe" priorities.

5. Model Families, Abstention Approaches, and Evaluation Protocol

Five architectures, each ≈300M parameters, are benchmarked: ViT, Swin Transformer, DenseNet, ConvNeXt, and MLP-Mixer. Models are trained on the curated train set and evaluated on the open-test split, using batch size 16 and a learning rate decaying from $1 \times 10^{-5}$ to $1 \times 10^{-6}$.

Seven to eight abstention strategies are systematically tested:

1. Softmax Probability (SP): $1-$ entropy of the softmax output.
2. Adjusted SP/ODIN: Softmax with temperature scaling, input-perturbed.
3. Mahalanobis Detector: Feature-space Gaussian distance [lee2018simple].
4. KNN: k-nearest neighbor feature distance [sun2022out].
5. TAPUDD: GMM clustering plus Mahalanobis [dua2023task].
6. OpenMax: Weibull-calibrated class mean activations [bendale2016towards].
7. MC-Dropout: Uncertainty from dropout at inference [gal2016dropout].
8. Deep Ensemble: Ensemble variance [lakshminarayanan2017simple].

Performance is reported as mean ± std Hellinger distance (per category) and $RS_c$ for $c \in \{0, 450, 900\}$, across five model initializations.

6. Empirical Insights and Key Results

Experiments reveal:

• Architecture vs. abstention: Model architecture exerts less influence than abstention strategy; differences among ViT, Swin, DenseNet, ConvNeXt, and MLP-Mixer are small.
• Must-Act scenarios: Distance-based abstention methods (Mahalanobis, KNN, TAPUDD) achieve lowest (best) Hellinger alignment.
• Must-Abstain group: SP and ASP, which abstain aggressively, more closely match human judgments on clear out-of-class examples.
• Uncertain category: All methods struggle (Hellinger $\approx 0.50$–$0.60$). KNN-based strategies attain the best aggregate human alignment across all groups.
• Reliability alignment: A strong negative correlation exists between mean Hellinger distance and $RS_{900}$, implying that Hellinger alignment is a valid proxy for reliability under high risk-aversion.
• Pre-training and unsupervised learning: ImageNet-1K pre-training benefits transformer models more than CNN or MLP baselines, while self-supervised approaches (SimCLR, BYOL) reveal non-universal, architecture-dependent effects. Deep Ensembles consistently enhance alignment.

These findings underscore the centrality of abstention mechanism selection and the current difficulty of robust human-like abstention under visual uncertainty (Lee et al., 2023).

7. Applications, Limitations, and Future Directions

Applications of VisAlign include pre-deployment model benchmarking for AI alignment and safety; for example, as a filter in sensitive domains such as medical imaging or autonomous vehicles. The distance-based alignment metrics present a method-agnostic tool for calibration and abstention tuning in open-set or high-stakes perception tasks.

Stated limitations:

• Restriction to 10 familiar mammal classes constrains generality.
• Uncertainty is explored only via synthetic corruptions and cropping, not via "in-the-wild" sources like lighting, occlusion, or motion blur.
• The continuous spectrum of uncertainty is imperfectly mapped by finite corruption intensities.
• The cost and logistics of acquiring 134 annotations per Uncertain image limit dataset scalability.

Planned extensions and research avenues include broader class coverage, incorporation of real-world uncertain scenarios (e.g., adverse weather conditions), adaptation for detection or segmentation tasks, fusion of distance- and probability-based abstention signals, and exploration of ethical/bias-alignment components in other domains (Lee et al., 2023).