AISLEX: Approximate individual sample learning entropy with JAX

Abstract

We present AISLEX, an online anomaly detection module based on the Learning Entropy algorithm, a novel machine learning-based information measure that quantifies the learning effort of neural networks. AISLEX detects anomalous data samples when the learning entropy value is high. The module is designed to be readily usable, with both NumPy and JAX backends, making it suitable for various application fields. The NumPy backend is optimized for devices running Python3, prioritizing limited memory and CPU usage. In contrast, the JAX backend is optimized for fast execution on CPUs, GPUs, and TPUs but requires more computational resources. AISLEX also provides extensive implementation examples in Jupyter notebooks, utilizing in-parameter-linear-nonlinear neural architectures selected for their low data requirements, computational simplicity, convergence analyzability, and dynamical stability.

Keywords

• Anomaly detection
• Learning entropy
• JAX
• Machine learning weight evaluation