Neural Heterogeneity Enables Adaptive Encoding of Time Sequences
Biological systems represent time from microseconds to years. An important gap in our knowledge concerns the mechanisms for encoding time intervals of hundreds of milliseconds to minutes that matter for tasks like navigation, communication, storage, recall, and prediction of stimulus patterns. A recently identified mechanism in fish thalamic neurons addresses this gap. Representation of intervals between events uses the ubiquitous property of neural fatigue, where firing adaptation sets in quickly during an event. The recovery from fatigue by the next stimulus is a monotonous function of time elapsed. Here we develop a full theory for the representation of intervals, allowing for recovery time scales and sensitivity to past stimuli to vary across cells. Our Bayesian framework combines parametrically heterogeneous stochastic dynamical modeling with interval priors to predict available timing information independent of actual decoding mechanism. A compromise is found between optimally encoding the latest time interval and previous ones, crucial for spatial navigation. Cellular heterogeneity is actually necessary to represent interval sequences, a novel computational role for experimentally observed heterogeneity. This biophysical adaptation-based timing memory shapes spatiotemporal information for efficient storage and recall in target recurrent networks.
Code (0)
등록된 구현이 없습니다.
Similar Papers 제목 키워드 기반
Q-LocalAdam: Memory-Efficient Client-Side Adaptive Optimization for Edge Federated Learning
Federated learning on edge devices must cope with non-IID client data and tight memory budgets. Adaptive optimizers like Adam stabilize training under data heterogeneity but require storing full-precision momentum and va…
Federated LearningKairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models
Inherent temporal heterogeneity, such as varying sampling densities and periodic structures, has posed substantial challenges in zero-shot generalization for Time Series Foundation Models (TSFMs). Existing TSFMs predomin…
Zero-shot GeneralizationSkeleton-to-Image Encoding: Enabling Skeleton Representation Learning via Vision-Pretrained Models
Recent advances in large-scale pretrained vision models have demonstrated impressive capabilities across a wide range of downstream tasks, including cross-modal and multi-modal scenarios. However, their direct applicatio…
Representation LearningAction RecognitionHeterogeneity of Interaction Strengths and Its Consequences on Ecological Systems
Ecosystems are formed by networks of species and their interactions. Traditional models of such interactions assume a constant interaction strength between a given pair of species. However, there is often significant tra…
Spatio-Temporal Self-Supervised Learning for Traffic Flow Prediction
Robust prediction of citywide traffic flows at different time periods plays a crucial role in intelligent transportation systems. While previous work has made great efforts to model spatio-temporal correlations, existing…
AttributePredictionRobust Traffic PredictionSelf-Supervised Learning+2