CNN Binary Lab
Tiny convolutional neural network for two-class image classification with filter and feature-map visualization.
This page trains a compact CNN: Conv(3x3) -> ReLU -> Flatten -> Dense -> ReLU -> Dense -> Softmax.
Input images are converted to grayscale and resized to 32x32, so each sample is a 1024-dimensional vector before convolution.
Binary labels are mapped to class probabilities: P(class 1) and P(class 2).
The network is optimized with cross-entropy over two classes:
$$ L = -\frac{1}{N}\sum_{i=1}^{N}\sum_{c=1}^{2} y_{ic}\log(\hat{y}_{ic}) $$
Use a lower learning rate for stable convergence, or a higher learning rate for faster but noisier updates.
- Load demo cat/dog images or upload custom files into each class bucket.
- Initialize weights, run a few epochs, and monitor loss and accuracy.
- Check filter values and feature maps to understand what the first conv layer captures.
- Upload a test image and inspect class probabilities.
Dataset: 0
Epoch: 0
Loss: -
Accuracy: -
Training Images
Uploaded images are resized to 32x32 grayscale before training.
Loading images...
Conv Filters (Realtime)
Prediction
-