Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning

The Platonic Representation Hypothesis posits that behind different modalities of data (what we sense or detect), there exists a universal, modality-independent representation of reality. Inspired by this hypothesis, we treat each neuron as a system, where we can detect the neuron's multi-segment activity data under different peripheral conditions. We believe that, similar to the Platonic idea, there exists a time-invariant representation behind different segments of the same neuron, which reflects the intrinsic properties of the neuron's system. Intrinsic properties include the molecular profiles, location within brain regions and morphological structure, etc. The optimization objective for obtaining intrinsic neuronal representations should meet two criteria: (I) the representations of recording segments from the same neuron must exhibit higher similarity compared to those from different neurons; (II) the representations should generalize effectively to out-ofdomain data. To this end, we propose the NeurPIR (Neuron Platonic Intrinsic Representation) framework, which leverages contrastive learning by treating segments from the same neuron as positive pairs and those from different neurons as negative pairs. In the implementation, we adopt VICReg, which only uses positive pairs while indirectly separating dissimilar samples through regularization terms. To validate the efficacy of our method, we first conducted tests on simulated neuronal population dynamics data generated by the Izhikevich model. The results confirmed that our approach accurately captured the neuron types as defined by the preset hyperparameters. Subsequently, we applied our method to two real -world neuron dynamics datasets, which included neuron type annotations derived from spatial transcriptomics and the location of each neuron within brain regions. The representations learned from our model not only accurately predicted neuron types and locations but also demonstrated robustness when tested on out-of-domain data (data from unseen animals). This finding underscores the potential of our approach in furthering the understanding of neuronal systems and offers valuable insights for future neuroscience research. Code is available at https://github.com/ww20hust/NeurPIR . :