Skip to main content

Convolutional neural networks

From a single 3x3 filter slid over pixels to a ResNet50 fine-tuned on a waste-sorting dataset and read through Grad-CAM: this course covers the CNN machinery that still powers most computer vision in production.

Course duration: 7h​

What you will learn​

  • Compute convolution outputs and parameter counts by hand, before trusting any framework
  • Choose stride, padding and dilation to control resolution and the receptive field
  • Read and reproduce the historical architectures — LeNet, AlexNet, VGG, ResNet, Inception, MobileNet — and know when each one still makes sense
  • Augment an image dataset without corrupting the labels
  • Transfer a pretrained backbone through a two-phase fine-tuning procedure without breaking batch normalisation
  • Interpret a trained model with Grad-CAM and detect when it looks at the background instead of the object

Prerequisites​

  • Deep learning fundamentals (course 07): backpropagation, optimisers, overfitting
  • One deep learning framework (course 08 for Keras, course 09 for PyTorch)
  • Comfort with NumPy arrays and image tensors

Course modules​

  1. Convolution, filters and feature maps
  2. Stride, padding and receptive field
  3. Pooling and spatial downsampling
  4. LeNet and AlexNet: the first successes
  5. VGG and stacking small filters
  6. ResNet and residual connections
  7. Inception, MobileNet and compute efficiency
  8. Data augmentation for images
  9. Transfer learning and progressive fine-tuning
  10. Interpretation: Grad-CAM and saliency maps

The red thread​

Modules 1 to 3 build a small CNN from scratch on CIFAR-10 (10 classes, 32 by 32 colour images), so that every layer type is written and measured before a framework hides it. Modules 4 to 6 replace that small network with LeNet, then a VGG-style block, then a reduced ResNet, on the same data. Modules 7 to 9 switch to a realistic case study: sorting photographs of household waste into six classes with a ResNet50 pretrained on ImageNet, then fine-tuned. Module 10 opens the model and asks whether it looked at the object or at the sky.

The framework is Keras, with short PyTorch equivalents in call-outs whenever the two differ in a way that matters.

Assessment and certificate​

The course ends with a 40-question exam covering every module. On success, a certificate of completion is issued; its number can be verified by any third party on the platform.

Free courses, by contrast, end with a 5-question quiz and a certificate preview, without certification.