How neural networks learn to classify chaotic time series

Area of Science:

• Dynamical Systems
• Machine Learning
• Time Series Analysis

Background:

• Analyzing neural networks for time series classification is crucial.
• Large Kernel Convolutional Neural Networks (LKCNNs) show superior performance in regular vs. chaotic time series classification.
• LKCNNs outperform traditional methods like Lyapunov exponent calculation.

Purpose of the Study:

• To analyze the internal mechanisms of LKCNNs used for time series classification.
• To understand why some LKCNNs achieve high accuracy while others perform poorly.
• To identify criteria for predicting poor performance in trained LKCNNs.

Main Methods:

• Investigated LKCNNs trained on regular and chaotic time series data.
• Analyzed network activations to understand classification strategies.
• Compared performance of models initialized with random weights.

Main Results:

• LKCNNs utilize qualitative features of input sequences for classification.
• Trained LKCNN models fall into two distinct performance groups: high (0.94 accuracy) and low (0.72 accuracy).
• Low-performing models display periodic activations, similar to randomly initialized networks.

Conclusions:

• Periodic activations in LKCNNs correlate with low classification accuracy.
• These activation patterns offer a potential method for a priori identification of poorly performing trained weights.
• Understanding LKCNN internal dynamics can improve model reliability in time series analysis.