Part IV · Ch. 03 — CNN (Convolutional Network)

draft

Connectionist · Vision · Convolution. Detects spatial patterns through hierarchical filters; it was the architecture that started the deep learning revolution. Card: ../02-types-of-ai/02-connectionist.kmd.

🎨 Figure F-IV.3.0 — Hierarchy of features. Brief: an image passing through layers that detect edges → textures → parts → object; activation maps beside each layer.

CNN — hierarchy of features

1. Definition and short history

A network that applies convolutions to extract local features with translation equivariance. Lineage: LeNet (1989) → AlexNet (2012, the spark) → VGG/ResNet → still standard on edge and for specific tasks (see Part III, era 5).

2. Foundations

  • Group theory / symmetries — translation equivariance is the key prior.
  • Optics / signal processing (DSP) — convolution, filters, frequency.
  • Linear algebra and calculus — convolution as a linear operation; backprop.
  • Neuroscience — inspired by the visual cortex (receptive fields, Hubel & Wiesel).

3. Algorithms and architectures

  • Convolutional layer — sliding filters + activation (ReLU).
  • Pooling — reduces resolution, gives local invariance.
  • Depth + residual — ResNet (skip connections) trains hundreds of

    layers.

  • Variants — depthwise (MobileNet), detection (YOLO, Faster R-CNN),

    segmentation (U-Net).

4. Inputs

  • Hardware: GPU (training); very efficient on edge/NPU (inference).
  • Data: labeled images (ImageNet-like), augmentation.
  • Data structures: 4D tensors (NCHW), activation maps.
  • Systems: PyTorch/TF; data augmentation; ONNX for edge.

5. Specialized life cycle

Stage Specialization
0 Problem Classification? detection? segmentation? edge constraint?
1 Data Labeled images, augmentation (flip, crop, color jitter)
2 EDA Class balance, label quality, visual biases
3 Modeling Backbone (ResNetEfficientNetMobileNet), task head
4 Training SGD/Adam, batch norm, transfer learning from pre-training
5 Evaluation Accuracy/top-5, mAP (detection), IoU (segmentation)
5.5 Acceptance Robustness to perturbation, tests on target distribution, edge latency
6 Production Quantization/pruning for NPU; low-latency inference
7 Monitoring Visual drift (lighting, camera), accuracy degradation
8 Retraining New field data, fine-tuning
9 Governance Privacy (facial recognition), demographic bias, surveillance

6. Capabilities, modes and modalities

Spatial/visual: classification, detection, segmentation, recognition; the basis of vision in embedded and medical systems.

7. Limits, risks and ethics

Fragile to adversarial examples and domain shift; sensitive uses (surveillance, facial recognition, demographic bias). ViT surpasses CNNs at large scale.

8. State of the art and examples

ResNetEfficientNetConvNeXt; YOLO (real-time detection); U-Net (medical). Convergence with Transformers (conv-attention hybrids); CNNs persist where edge efficiency matters.