Abstract: International audience
Introduction: The End of the "Intractability" Excuse
For decades, perception scientists utilized the perceived "intractability" of human-level vision as a shield, justifying slow, methodical research focused on simplified stimuli. Biological vision was long considered a unique, "impossible to match" miracle of nature. Consequently, researchers often focused on "baby steps"—studying elementary patterns like Gabor patches—because complex, naturalistic object recognition seemed too difficult to model computationally.
However, the "deep learning revolution" has fundamentally shifted this paradigm. Deep Neural Networks (DNNs) now routinely outperform humans in standard object recognition benchmarks. This sudden success has rendered the old excuses obsolete, transforming AI from a distant science-fiction dream into a powerful—if biologically imperfect—tool and model for understanding the human brain. The author, Rufin VanRullen, argues that this creates a crisis of purpose: if machines can "see" better than us, the role of the perception scientist must evolve from cataloging simple behaviors to understanding the complex mechanics of high-level vision.
Methodological Bridges: AI as a Sandbox for Neuroscience
Rather than introducing a single new algorithm, this perspective piece synthesizes how modern AI methodologies are bridging the gap between computer science and neuroscience. Key techniques include:
- Generative Modeling (VAE/GAN): Using a 13-layer Variational Auto-Encoder (VAE) trained via a Generative Adversarial Network (GAN) procedure on the CelebA database (over 200,000 celebrity faces), researchers can encode complex visual information into a low-dimensional "latent space." This allows for the manipulation of abstract concepts—like adding a "beard vector" to a face—using simple linear algebra.
- Representational Similarity Analysis (RSA): By comparing the activity patterns in specific layers of a DNN to specific regions in the human brain (such as the IT cortex), scientists have found that these networks process information in a hierarchical, analogous way to biological systems.
- Biological Reconciliations: Efforts are underway to make DNNs more "brain-like" by introducing spiking neurons, oscillatory components, and Recurrent Neural Networks (RNNs) to mimic the massive feedback loops found in nature.
Emergent Phenomena: The Face Inversion Effect
One of the most striking insights from this research is that certain "human" perceptual quirks are actually logical consequences of computational constraints. A prime example is the "Face Inversion Effect." When a network is trained exclusively on upright faces, it fails to process upside-down faces correctly, often "hallucinating" features where they do not belong. This mirrors a well-known human psychological bias, suggesting that our own visual limitations may stem from the way our neural architectures optimize for natural statistics.
Key Results and Neural Correspondence
The paper highlights the "staggering creativity" and accuracy of modern models. Beyond surpassing human performance in ImageNet benchmarks, top-performing DNNs show internal representations remarkably similar to those found in the primate visual cortex (e.g., Cadieu et al., 2014). Furthermore, the ability to manipulate latent spaces suggests that these networks have truly "learned" the underlying structure of the world, rather than just memorizing data points.
Discussion: Limitations and the Path Forward
While DNNs are powerful, VanRullen notes they are not yet perfect brain clones. Most rely on "back-propagation," a learning algorithm unlikely to exist in the brain, and they lack the temporal dynamics (spiking) and top-down feedback of biological systems. Furthermore, replicating a behavior in a machine does not automatically equate to a full understanding of biological function.
The author concludes with a "Wake-up Call" for the field. Perception scientists must move beyond "monastic" focus on over-simplified stimuli and use DNNs to create new experimental designs. By using AI as an "artificial cousin," researchers can probe the limits of the visual system using perfectly controlled, synthesized stimuli, eventually leading to a unified theory of perception that combines the "how" of AI with the "why" of psychology.
Summary updated