TL;DR
Neural representations of sensory information aren't static, even when behavior is stable. This paper shows that this "representational drift" in mouse visual cortex has a specific geometric signature, which closely matches the drift observed in artificial neural networks when "dropout" noise is applied. This suggests a potential mechanism for drift in the brain and hints at its computational benefits, like preventing overfitting.
What Problem Does This Paper Address?
The brain's neural circuits are constantly changing, yet our ability to perform tasks and recall memories remains remarkably stable. A key puzzle is how neural representations of sensory information can change over time – a phenomenon called "representational drift" – without degrading performance. This paper aims to understand the geometric properties of this drift in biological neural networks and compare it to artificial neural networks (ANNs) to uncover potential underlying mechanisms and computational advantages.
Methods
The authors employ a geometric approach to characterize representational drift. They analyze neural activity from two-photon calcium imaging in mice and compare it to ANNs. Key techniques include:
- Principal Component Analysis (PCA): Used to define the "variational space" of stimulus representations, capturing the primary directions of variance.
- Drift Vector: Defined as the difference between mean response vectors of a stimulus group across different time points.
- Geometric Metrics: Quantifying drift magnitude, angle relative to principal components, variational space overlap, and classifier performance.
- Artificial Neural Networks (ANNs): Simple convolutional neural networks (CNNs) were trained and then subjected to various types of noise (additive node, additive gradient, node dropout, weight dropout, multiplicative node) to induce feature space drift.
- Linear Support Vector Classifiers (SVCs): Used to assess the separability of stimulus representations over time.
Datasets
- Allen Brain Observatory (Passive Data): In vivo two-photon calcium imaging data from mice passively viewing visual stimuli.
- Allen Visual Behavior Two Photon Dataset (Behavioral Data): In vivo two-photon calcium imaging data from mice performing an image change detection task.
- CIFAR-10 Dataset: Used to train convolutional neural networks.
Key Insights and Ideas
- Drift is Geometrically Structured: Representational drift isn't random; it follows specific geometric rules. It tends to occur along directions that already have high variance within a stimulus group, leading to a turnover of active neurons.
- Dropout as a Potential Mechanism: The geometric characteristics of drift in ANNs induced by "node dropout" noise closely mirror the drift observed in biological visual cortex. This suggests that a similar mechanism, perhaps a random suppression of neuronal activity, could underlie biological representational drift.
- Computational Advantage: The similarity to dropout hints that representational drift might be a computationally advantageous strategy for the brain, potentially helping to prevent overfitting and improve generalization, much like it does in ANNs.
- Performance Stability Despite Drift: Despite significant changes in neural representations, linear classifiers trained to distinguish stimuli maintain high accuracy across days, suggesting that the drift preserves the essential information for task performance.
Key Results
- Passive Data: Representational drift was observed in mice passively viewing natural movies. Drift occurred along directions of high variance and showed a bias towards representation turnover.
- Behavioral Data: Similar drift characteristics were found in mice performing an image change detection task, regardless of whether the stimuli were familiar or novel.
- ANNs: Node dropout, and to a lesser extent weight dropout and multiplicative node noise, induced drift in ANNs that closely matched the geometric properties observed in the biological data.
- Classifier Performance: In both biological and artificial systems, linear classifiers trained on one time point could accurately classify data from another time point, even with significant representational drift.
- Drift Magnitude: The magnitude of drift was comparable to the magnitude of the mean response vector, indicating substantial changes over days.
Limitations and Future Work
- Timescale: The study focused on drift over days. Further investigation into shorter timescales and how drift accumulates is suggested.
- Brain Areas: The study was limited to the primary visual cortex. Future work could explore if these geometric characteristics are present in other brain areas and neuron types.
- "Coding Null Space": The stochastic yet non-uniform nature of drift suggests it might occur along a specific manifold or "coding null space," which warrants further investigation.
- More Realistic Models: Injecting equivalent noise into more biologically realistic network models, such as spiking neural networks, is proposed to see if the observed trends continue.
- Behavioral Dependence: While efforts were made to control for behavioral influences, the complex interplay between behavior and neural representations remains an area for deeper exploration.
Summary updated