TL;DR
This paper investigates how the geometric structure of neural representations for visual concepts in the human brain and deep neural networks (DNNs) relates to classification accuracy. They found that while both brains and DNNs use geometric properties to improve classification, they employ distinct strategies: the brain primarily relies on increasing the "geometric Signal" (distance between concept representations), whereas DNNs initially leverage increasing "Dimensionality" (number of independent variation directions).
What Problem Does This Paper Address?
Accurately classifying visual concepts (like objects or scenes) is a fundamental task for the visual system. While it's known that the "disentanglement" of neural representations for different concepts is crucial for this, it's unclear how this disentanglement is achieved across different brain regions or how it compares to artificial systems like deep neural networks (DNNs). Specifically, the authors want to understand what geometric properties of "concept manifolds" (sets of neural activity patterns representing a concept) vary across the human visual cortex and how these variations contribute to classification accuracy. They also aim to compare these strategies to those employed by DNNs, which are often seen as models of brain function.
Methods
The core of the study relies on a theoretical framework developed by Sorscher et al. [6] that links specific geometric properties of concept manifolds to the performance of "few-shot" linear classifiers (classifiers trained on very few examples). These properties include:
- Geometric Signal: The distance between the centers of concept manifolds.
- Bias: The relative radius of one manifold compared to another.
- Dimensionality: The number of independent directions of variation within a manifold.
- Noise-Signal Overlap: The alignment of manifold directions with the vector connecting their centers.
- Noise-Noise Overlap: The alignment between the principal directions of different manifolds.
The authors used fMRI data from the Natural Scenes Dataset (NSD) and also employed a brain-optimized deep neural network (GNet) trained on this data to generate synthetic, "noiseless" brain activity patterns. They then calculated these geometric properties for concept pairs across various visual cortex regions of interest (ROIs) and across layers of DNNs. They used "geometric SNR" (Signal-to-Noise Ratio) as a proxy for disentanglement and classification accuracy.
Datasets
- Natural Scenes Dataset (NSD) [17]: A large fMRI dataset containing blood-oxygenation-level dependent (BOLD) activity from eight subjects viewing thousands of natural scenes. This dataset was used for both estimating brain activity patterns and training the GNet encoding model.
- Common Objects in Context (COCO) dataset [18]: Used for defining visual concepts (super-categories like "appliance," "vehicle," "food," etc.) for manifold analysis.
- ImageNet database [27]: Used for a broader analysis of concept manifolds with a larger set of object categories.
Key Insights and Ideas
- Brain's "Signal" Strategy: The human visual cortex appears to improve classification accuracy primarily by increasing the "geometric Signal" – making the representations of different concepts further apart. This strategy is observed consistently across visual areas.
- DNNs' Initial "Dimension" Strategy: In contrast, early layers of DNNs achieve improved classification by increasing "Dimensionality" – allowing for more complex variations within a concept's representation. This strategy is often coupled with a decrease in geometric Signal.
- Antagonistic Covariation: A key finding is the strong negative correlation between geometric Signal and Dimensionality. As one increases, the other tends to decrease. This "antagonistic" relationship means that changes in one property can counteract the effects of changes in the other on classification accuracy.
- "Non-functional" Variation: In higher visual areas, despite significant variation in manifold geometry, the overall classification accuracy (geometric SNR) doesn't change much. This suggests that some geometric variations might be "non-functional" in terms of improving disentanglement.
- Overlap is Predictable: The "combined geometric overlap" term is highly predictable from Signal and Dimensionality, suggesting these two properties are the primary drivers of manifold geometry.
Key Results
- Variation in classification accuracy across human visual cortex is driven by a variation in geometric Signal.
- Variation in classification accuracy across early and mid DNN layers is driven by an increase in Dimensionality.
- Geometric Signal and Dimensionality are strongly, negatively correlated across both brain regions and DNN layers.
- The geometric SNR predicted by the Sorscher et al. theory accurately captured empirical few-shot classification error in both fMRI data and the encoding model's output.
- In higher visual cortex, the antagonistic covariation between Signal and Dimensionality limits the impact of geometric variation on classification error.
- DNNs switch strategies: early layers use the "Dimension" strategy, while later layers adopt the "Signal" strategy, similar to the brain.
Limitations and Future Work
- Species and Measurement Differences: The authors acknowledge potential discrepancies with previous studies due to differences in species (monkey vs. human), brain regions studied, and measurement techniques (spiking vs. fMRI). They note that fMRI likely underestimates true geometric dimensionality.
- Trial-to-Trial Variability: While the encoding model helps mitigate trial-to-trial variability, it's still a factor that can influence estimates. The authors suggest that the brain's "Signal" strategy might be more robust in the presence of high trial-to-trial variability.
- Understanding DNN Strategy: The reason why training induces a "Dimension" strategy in DNNs remains unclear.
- Future Work: The authors suggest further investigation into how visual cortex controls dimensionality and exploring the functional implications of different geometric strategies on classification decision-making. They also pose the question of whether local learning rules and trial-to-trial variability could induce more brain-like geometry in DNNs.
Summary updated