For decades, neuroscientists and computer scientists faced a "mapping problem." If you want to know if a Deep Neural Network (DNN) processes an image the same way a human brain does, how do you compare them? A brain is a wetware mesh of billions of neurons; a DNN is a silicon-based stack of linear algebra operations. There is no obvious way to map "Neuron A" in the visual cortex to "Weight B" in a convolutional layer.
The breakthrough, championed by researchers like Robert M. Mok and Katherine R. Storrs, is to stop looking at the units themselves and start looking at the relationships between the things they represent. This approach, known as Representational Similarity Analysis (RSA), provides a "universal translator" for intelligence.
The Relational Intuition
The core intuition of RSA is that the identity of a neural response matters less than its context relative to other responses.
Imagine showing a human and an AI two pictures: a Golden Retriever and a Labrador. In both systems, these images trigger a specific pattern of activity. Now show them a third picture: a toaster.
In a well-aligned system, the "distance" between the two dog representations should be small, while the distance between either dog and the toaster should be large. By measuring these distances across hundreds of stimuli, we create a Representational Dissimilarity Matrix (RDM). As Julia Berezutskaya explains, if the RDM of a brain region matches the RDM of an AI layer, we can conclude they share a common computational strategy, even if their internal "hardware" is completely different.
Hierarchical Alignment
One of the most compelling pieces of evidence for this approach is the discovery of hierarchical gradients. In Brains on Beats, Umut Güçlü and colleagues showed that DNNs trained to tag music exhibit a structure that mirrors the human auditory cortex. Early layers of the AI align with the primary auditory cortex (processing simple frequencies), while deeper layers align with the superior temporal gyrus (processing complex musical concepts).
This hierarchical "handshake" isn't limited to sound. Jinge Wang demonstrated that face-recognition AI and primate brains use nearly identical "region-based" coding to identify individuals. Similarly, Cai Wingfield showed that we can actually improve AI performance by "reverse-engineering" its architecture to better match the bottleneck structures found in human speech processing.
Why Does Alignment Emerge?
If biological and artificial systems are so different, why do their representations align? The research suggests that alignment is driven by task constraints.
In Invariant recognition drives neural representations of action sequences, Andrea Tacchetti and Leyla Isik argue that the physical requirement of "invariance"—recognizing an action regardless of the camera angle—forces any successful system (biological or artificial) toward the same mathematical solution.
Furthermore, the type of learning matters. Grace W. Lindsay found that Reinforcement Learning (RL) produces representations that match the mouse visual cortex better than standard supervised learning. This suggests that our brains aren't just passive filters; they are shaped by goal-directed behavior. This is echoed by George L. Malcolm, who notes that the brain dynamically re-weights these representations based on what the observer is trying to achieve.
Beyond Simple Categorization
Modern applications of RSA are moving beyond simple object recognition into more abstract domains:
- Navigational Affordances: Clemens G. Bartnik and Iris Groen used RSA to show that the brain has a specific map for "what can I do here?" (e.g., walking vs. swimming), a feature better captured by multimodal models like GPT-4 than by pure vision models.
- Generative AI: Lorenz Linhardt applied RSA to Latent Diffusion Models (like Stable Diffusion), finding that the most "human-like" understanding of objects occurs in the intermediate up-sampling layers, not the compressed bottleneck.
- Social Decision Making: Qianying Wu used structural similarity to show that our complex prosocial decisions (like helping a stranger) can be modeled as a cost-benefit tradeoff that aligns with semantic descriptions of the scenarios.
The Yardstick for Intelligence
Ultimately, RSA allows us to use AI as a "yardstick" for the brain. As Jianghong Shi and Michael A. Buice demonstrated in mouse models, we can use task-specific DNNs to unveil the functional properties of cortical regions that were previously mysterious. By comparing the relational structures of both systems, we are moving toward a unified "cognitive computational neuroscience" that treats biological and artificial intelligence as two sides of the same coin.