EventsThe 3rd International Online Conference on Biomimetics
Published
This submission belongs to the session S4. Bioinspired Sensorics, Information Processing and Control of the event The 3rd International Online Conference on Biomimetics
Published date
19 Nov, 2026
Academic Editor
author-avatarAndrew Adamatzky
Citation
Sudip Barua, Reverse-Time Learning in Neural Networks: Time Direction Is Not a Constraint for Information Processing, in Proceedings of The 3rd International Online Conference on Biomimetics, 24 November–26 November 2026, MDPI: Basel, Switzerland
Share
Email
Facebook
Twitter
LinkedIn

Reverse-Time Learning in Neural Networks: Time Direction Is Not a Constraint for Information Processing

1. "Independent Researcher"
Abstract

Nature processes information bidirectionally through spike-timing-dependent plasticity, where pre-before-post spikes induce long-term potentiation and post-before-pre spikes induce long-term depression. This anti-Hebbian mechanism suggests that biological neural networks are not constrained by temporal causality. Yet modern artificial neural networks remain limited by a 75-year-old assumption that learning requires temporal causality from past to future only. Using a simple continuous-time neural network architecture inspired by biological anti-Hebbian plasticity, this paper challenges that assumption by demonstrating that neural networks can learn reverse-time mappings, predicting past events from future data, across diverse real-world domains including finance, healthcare, climate, energy, transportation, technology, and social media. Key findings include battery voltage prediction achieving 98.96% reverse-time correlation, which actually surpasses forward-time performance of 85.49%. Electricity demand forecasting achieves 94.25% correlation, website traffic prediction achieves 90.90%, temperature forecasting achieves 90.14%, forex exchange rate achieves 86.58%, and heart rate monitoring achieves 27.99% correlation. Notably, reverse-time learning outperformed forward-time learning on three datasets: battery voltage (98.96% vs 85.49%), stock price (53.94% vs 44.13%), and heart rate (27.99% vs 18.23%). Statistical testing reveals no significant difference between forward and reverse time predictions across all domains, with an average difference of only 5.25% and p-value greater than 0.05, strongly suggesting that time direction is not a fundamental constraint for neural information processing. These results demonstrate that reverse-time learning is practical on real-world data with immediate applications in battery health monitoring, energy forecasting, anomaly detection, and continuous learning systems where traditional phase-separated training is impractical.

Keywords
Reverse-time learning
anti-Hebbian plasticity
continuous-time neural networks
temporal causality
time series prediction
bidirectional learning
spike-timing-dependent plasticity
neuromorphic computing
retrodictive forecasting
time reversal
Machine Learning Discovery of Mechanical Efficiency Principles in Bioinspired Lattice Architectures: An Explainable Artificial Intelligence Framework
Biomimetic desing of an electrochemical sensor utilizing TPhPFe³⁺/Al₂O₃ smart materials for detection H2O2 and C2H5OH