Introduction
The field of vision science is currently divided by a profound disagreement regarding the role of Deep Neural Networks (DNNs). On one hand, "scientific hype" suggests that DNNs are near-perfect models of the human visual system. On the other, a growing "gloom" perspective—most notably articulated by Bowers et al. (2022)—argues that DNNs fail so fundamentally at psychological tasks, such as Gestalt grouping and global shape recognition, that they are unsuitable as models for human vision.
In their commentary, Wichmann, Kornblith, and Geirhos argue that neither extreme does justice to the current state of the field. Instead of viewing DNNs as a static failure or a finished success, they propose viewing them as a rapidly evolving class of models that currently offer the most promising framework for developing image-computable theories of vision.
The Evolution of Artificial Vision
A central pillar of the "gloom" argument is that DNNs exhibit "fundamental" flaws, such as a reliance on local textures rather than global shapes. However, the authors point out that these critiques often rely on a "snapshot in time." By comparing the performance of early models like AlexNet to modern architectures, it becomes clear that many previously "insurmountable" limitations are being overcome.
For instance, while older convolutional neural networks (CNNs) were notoriously texture-biased, newer models trained on larger datasets or using different architectures (such as Vision Transformers) have begun to exhibit a shape bias much closer to that of humans. Large-scale models, including those with 22 billion parameters, are increasingly matching human error patterns and showing improved robustness against image distortions like noise and blur.
The Necessity of Image-Computability
The authors emphasize that for any model to be truly useful in vision science, it must be image-computable—meaning it can take raw pixels as input and produce a verifiable output. This is a critical advantage of DNNs over traditional psychological models, which often only address isolated components of the visual process and cannot handle the complexity of real-world images.
While DNNs are imperfect, they provide a complete pipeline from pixels to labels. This allows researchers to perform rigorous, head-to-head comparisons between artificial models and human behavior. The authors argue that a model that cannot process images is inherently limited in its ability to be falsified or applied to the complexities of natural vision. Furthermore, for many of the tasks where DNNs currently fail, there are no successful traditional computational models to take their place.
Prediction vs. Explanation
A common critique of DNNs is that they are "black boxes" that predict well but explain nothing. Wichmann and colleagues reject this dichotomy. They argue that in science, "an explanation without prediction cannot be trusted, and a prediction without explanation does not aid understanding." The goal should not be to choose between a predictive model and an explanatory theory, but to use the predictive power of DNNs to refine and ground our theoretical explanations.
Furthermore, the authors suggest that the perceived "hype" is often a byproduct of scientific incentives that reward bold claims over nuanced ones. By separating the engineering successes from the scientific interpretations, we can better appreciate the utility of DNNs without overstating their current human-likeness.
Future Directions: Beyond Discriminative Learning
Acknowledging that DNNs still struggle with complex tasks like combinatorial judgments and human-like sensitivity to Gestalt principles, the authors suggest several paths forward to bridge the gap between biological and artificial vision:
- Moving Beyond Classification: Incorporating causality and generative components (such as predictive coding) may be necessary to move beyond simple discriminative learning.
- Grounding in Physics: Future models may need to incorporate "intuitive physics" and psychology to achieve a more human-level understanding of the world.
- Vision-Science-Inspired Research: The field must transition from pure engineering—optimizing for top-1 accuracy—to research specifically designed to investigate the alignment between artificial and biological systems.
Conclusion
DNNs are not a final answer to the mysteries of human vision, but they are an indispensable tool. By avoiding the extremes of hype and gloom, vision scientists can leverage the rapid progress of machine learning to build more robust, falsifiable, and comprehensive theories of how we see.
Summary updated