Abstract
Neural decoding—the process of predicting behavioral or cognitive states from brain activity—often suffers from severe class imbalance, where common events overshadow rare but critical ones. This study introduces the DPP-VAE, a Variational Auto-encoder that incorporates a Determinantal Point Process (DPP) prior to encourage diversity in latent representations. By preventing the model from over-allocating representational power to majority classes, the DPP-VAE significantly improves the decoding of infrequent neural patterns. We demonstrate its efficacy on hippocampal recordings from rats performing an odor sequence task, where the model successfully captures rare "sequence replay" events.
Introduction
A fundamental challenge in neuroscience is the "class imbalance problem." In many experimental paradigms, subjects spend significantly more time in certain states than others. For example, in a sequential memory task involving odors A through E, a rat is far more likely to experience early odors (A, B) than later ones (E) due to trial terminations or errors. Standard machine learning models, such as Variational Auto-encoders (VAEs), tend to optimize for these majority classes, leading to "lazy" latent representations that fail to distinguish between rare, yet biologically significant, events.
Methodology: The DPP-VAE
The proposed model addresses this by modifying the standard VAE framework. While a traditional VAE assumes a standard normal distribution for the latent space, the DPP-VAE utilizes a k-DPP prior.
A Determinantal Point Process (DPP) is a probabilistic model that characterizes repulsion between points; it assigns higher probability to subsets of items that are diverse. By integrating a DPP into the latent space, the model is mathematically discouraged from clustering all latent variables around the majority class. This "diversity-encouraging" mechanism ensures that rare events (minority classes) occupy distinct, non-redundant regions of the latent space, making them easier to decode linearly. While the normalization term for continuous DPPs is computationally challenging to evaluate directly, the authors utilize approximation bounds to maintain feasibility.
Experimental Setup
The model was evaluated using electrophysiological recordings from the CA1 region of the rat hippocampus. Rats were trained to identify whether odors presented in a sequence (A-B-C-D-E) were "in sequence" (InSeq) or "out of sequence" (OutSeq).
As shown in Figure 1, the task structure naturally creates an imbalanced dataset, as the frequency of later odors in the sequence is lower than that of the initial odors.
Results and Discussion
The DPP-VAE was first validated on a synthetic imbalanced MNIST dataset. At a 1:1000 imbalance ratio, the DPP-VAE increased the representation of the minority class in generated samples by nearly 20-fold compared to a standard VAE. In the neural decoding task, the DPP-VAE showed marked improvements in identifying the rarest odors in the sequence.
As detailed in Figure 5, the F1 score for the rarest odor (E) improved from 0.103 to 0.133, while odor D saw a jump from 0.172 to 0.278. Beyond raw metrics, the model enabled the visualization of "sequence replay." During the presentation of a single odor, the latent representation was observed to shift toward the next odor in the sequence before it was even presented, providing low-dimensional evidence of hippocampal anticipation.
Conclusion
The DPP-VAE offers a robust solution to the class imbalance problem in neural decoding. By treating diversity as a geometric property of the latent space, the model preserves the nuances of rare neural events. This approach not only improves decoding accuracy but also facilitates the discovery of complex biological phenomena, such as sequential anticipation, that are often lost in the noise of imbalanced datasets.
Summary updated