LeNet5 CNN Visualizer

The handwriting character recognition system exposes the inner layer structure of the LeNet-5 CNN model to provide interactive weight and activation inspection interfaces.

1. Exposed Layer Topology

By adding intermediate value_info tensors to the graph outputs during model loading, the backend extracts the following activations dynamically for every input:

• Input Layer: Grayscale pixel array of shape [28, 28].
• Conv1 Output (ReLU): Channel tensors of shape [8, 28, 28].
• Pool1 Output: Subsampled tensors of shape [8, 14, 14].
• Conv2 Output (ReLU): Channel tensors of shape [16, 14, 14].
• Pool2 Output: Subsampled tensors of shape [16, 4, 4].
• Dense Inputs: Flattened vector of shape [256].
• Logits / Probs: Final predictions of shape [10].

2. Dynamic Weight Mapping

Static weight matrices are converted to standard arrays at startup:

• Conv1 kernels are loaded as shape [8, 1, 5, 5] along with 8 biases.
• Conv2 kernels are shape [16, 8, 5, 5].
• FC connection weights are shape [256, 10]. In the mobile client, this matrix is visualized as a 16x16 interactive grid heat-map, highlighting positive weights in blue and negative weights in red.

LeNet-5 Explainability Output Architecture

Interactive CNN Explainability Output