Zero-Shot Across-Subject Neural Decoding via Neuron-Level Representations

Neural decoders for spike-sorted electrophysiology data are highly accurate but fail to transfer across subjects in the absence of finetuning data. A key obstacle is that neural recordings lack a shared input space: neurons in different recordings have no correspondence. Existing neural decoder pre-training methods work around this challenge by introducing population-specific parameters in their architectures. While this approach allows for multi-subject pre-training, these decoders have to rely on supervised finetuning when transferring to new subjects, limiting their applicability to situations where such labeled calibration data is available. We introduce ZEN, which replaces these learned population-specific parameters with self-supervised and semantic representations inferred from activity by a Neuron-Level Encoder. Training a neural activity decoder on top of this representation space yields a model that applies to unseen animals zero-shot, without needing any fine-tuning or gradient-based updates. Our results demonstrate cross-subject zero-shot decoding from spike-sorted electrophysiology recordings for the first time. Additionally, when further finetuned on the target recordings, ZEN shows state-of-the-art label efficiency.