LearnAISign in
Phase 5 · 10 lessons · 30 h · about 3 weeks

👁 Computer Vision & CNNs

Teaching machines to see — and the architecture ideas that shaped all of deep learning

Lessons

5.1Convolution: the operation and its mathLocal dot products, shared everywhere⚡ demo · 3 h5.2CNN building blocksPooling, 1×1 convs, depthwise, dilated and transposed convolutions⚡ demo · 2 h 30 min5.3Architectures: LeNet → AlexNet → VGG → Inception → ResNet → EfficientNet → ConvNeXtEach landmark network contributed one durable idea3 h5.4Transfer learning and training CNNs in practiceStart from pretrained weights, then fine-tune with a modern recipe3 h5.5Object detectionWhere is it, and what is it — IoU, anchors, NMS, mAP and DETR3 h5.6SegmentationEvery pixel gets a label — FCN, U-Net, Mask R-CNN and Segment Anything2 h 30 min5.7Vision TransformersPatches as tokens — and what happens when you remove the convolutional prior2 h 30 min5.8Self-supervised and contrastive learningLearning representations without labels — SimCLR, MoCo, BYOL, DINO, MAE3 h5.9Interpretability, robustness and adversarial examplesWhat is the model looking at, and how easily can it be fooled?2 h 30 min5.10Project: train and deploy an image classifierFrom a folder of images to a public web demo, with error analysis5 h

About this phase

Vision is where deep learning first proved itself (ImageNet, 2012) and where many ideas you will meet again were born: weight sharing, residual connections, batch normalization, transfer learning, contrastive pretraining. This phase starts with the convolution operation and its math, builds up the architectures from LeNet to ConvNeXt, then moves to detection, segmentation, Vision Transformers and self-supervised learning. You finish by training and deploying a real classifier. Along the way you learn to count parameters and FLOPs, read an architecture table, fine-tune a pretrained backbone properly, and interrogate a model with Grad-CAM and adversarial examples — habits that transfer directly to transformers and LLMs.

By the end you can

  • Compute convolution outputs, output sizes, receptive fields, parameter counts and FLOPs by hand
  • Explain what each landmark architecture contributed (AlexNet → ResNet → EfficientNet → ConvNeXt) and why residual connections train
  • Fine-tune a pretrained backbone with a modern recipe (augmentation, discriminative LRs, mixed precision) on your own data
  • Describe detection and segmentation pipelines, and compute IoU, NMS results, mAP and Dice by hand
  • Explain Vision Transformers and contrastive/masked self-supervised learning, including the InfoNCE loss
  • Audit a vision model with Grad-CAM, adversarial examples and shift benchmarks, then deploy it as a web demo