NeuroScope

A framework for informed training of multi-layer perceptron, diagnosing training issues, accelerating learning and rapid prototyping.

NeuroScope is a Python frameworl designed to make AI interpretability accessible. It allows researchers to visualize the internal state of neural networks during training.

Features:

  • Activation Histograms: Visualize how neuron activations are distributed across layers.
  • Gradient Flow Analysis: Track gradient norms to detect vanishing or exploding gradients.
  • Training Dynamics: Watch how weights evolve over time with animated histograms.

The framework is built with only NumPy and Matplotlib.