A novel uncertainty-aware liquid neural network for noise-resilient time series forecasting and classification

While Liquid Neural Networks (LNN) are promising for modeling dynamic systems, there is no internal mechanism that quantifies the uncertainty of a prediction. This can produce overly confident outputs, especially when operating in noisy or uncertain environments. One potential issue that might be highlighted with LNNs is that their highly flexible connectivity leads to overfitting on the training data. This is targeted by the present work, which introduces the uncertainty-aware LNN framework, the UA-LNN, by considering Monte Carlo dropout for quantifying the uncertainty of LNNs. The proposed UA-LNN enhances the stochasticity of both training and inference, hence allowing for more reliable predictions by modeling output uncertainty. We applied the UA-LNN in the two tasks of time series forecasting and multi-class classification, where we showed its performance on a wide range of datasets and under different noise conditions. The proposed UA-LNN has shown the best results, outperforming the benchmarks of standard LNN, Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP) models in terms of R2, RMSE, and MAE consistently. Additionally, for performance metrics such as accuracy, precision, recall, and F1 score, the results showed improvement over LSTM and MLP models in multi-classification tasks. More importantly, under heavy noise, the UA-LNN maintained superior performance, while demonstrating enhanced classification capabilities across many datasets with challenging tasks, such as arrhythmia detection and cancer classification.