Introduction
In deep learning, "invariance" refers to a model’s ability to produce the same internal representation for different inputs—for instance, recognizing a dog regardless of lighting or background. For artificial intelligence to be safe and interpretable, its internal invariances should ideally align with human perception; the model should ignore the same "irrelevant" details that humans do. However, the research community has been divided on whether deep neural networks (DNNs) naturally develop human-like vision or rely on features entirely alien to us. This paper addresses this conflict by demonstrating that the perceived alignment of a model is highly sensitive to the evaluation methodology used.
Methodology: Representation Inversion
The authors evaluate alignment using Representation Inversion. Given a target image, they start with a random noise "seed" and use gradient descent to find an Identically Represented Input (IRI)—an image that the model perceives as identical to the target at a specific layer.
To test the limits of this alignment, the study categorizes IRI generation into three types:
- Regularizer-free: Finding an IRI without any constraints.
- Human-leaning: Adding constraints (like Total Variation) that force the generated image to look "natural" to humans.
- Adversarial: A novel approach that intentionally searches for an IRI that is as perceptually different from the target as possible (using LPIPS distance).
The Measurement Trap
A core finding of the paper is that previous claims of human-aligned representations were often artifacts of "human-leaning" regularizers. As shown in the figure above, when a standard model is tested with a human-leaning regularizer, the resulting IRI looks somewhat recognizable. However, in a regularizer-free setting, the same model's IRI is indistinguishable from noise. This suggests that the regularizer "cleans up" the output, creating an illusion of alignment that the model's internal features do not actually possess.
Key Insights and Results
The study identifies a specific "recipe" that consistently leads to better human-machine alignment across benchmarks like CIFAR and ImageNet:
- Standard Models Fail: Standard ResNet18 models show nearly 0% alignment when tested without regularizers, relying almost entirely on non-human features.
- Robustness is Essential: Adversarially trained (AT) models—those trained to resist $\ell_2$ perturbations—show a significant jump in alignment. For example, a robust ResNet18 reaches ~63% alignment on CIFAR10.
- Architecture Matters: Residual connections (ResNets) align significantly better than VGG-style architectures, even when both are robustly trained. This suggests that skip connections may be fundamental to learning human-like hierarchies.
- Self-Supervised Learning (SSL) Superiority: Models trained with contrastive losses (like SimCLR) combined with adversarial augmentation often achieve the highest alignment scores, sometimes surpassing supervised robust models without requiring any labels.
Conclusion
While adversarial training and residual architectures significantly improve alignment, the authors note that even the best models still harbor non-human invariances detectable by the "adversarial regularizer." This work suggests that "alignment" should become a standard evaluation criterion alongside accuracy and robustness. Future research is needed to determine how these metrics correlate with out-of-distribution generalization and broader model interpretability.
Summary updated