Demo 05

Deep Learning And CNNs: How do machines learn local visual features?

CNNs break image understanding into local receptive fields and hierarchical features. This chapter uses a small grid to show how a kernel slides and produces a feature map.

Teaching interaction

CNN Kernel Explorer

How do machines identify local features in an image?

-101
-101
-101
3
Current window response: 3

Weighted sum 3 ÷ 1 = 3

Scan the upper-left window

Edge detection highlights changes in brightness. The kernel first inspects a small region, not the entire image.

Learning goals
  • Understand how a kernel scans an image through local windows instead of reading the whole image at once.
  • Observe how different kernels produce different feature-map responses.
  • Explain why local receptive fields and parameter sharing suit visual feature extraction.
Simplification note

This demo illustrates convolution on a two-dimensional grid and does not train a real CNN. The values only build intuition for local receptive fields and feature maps.

Observation guide

Local features can combine into higher-level visual understanding

The earlier problem was that hand-crafted visual features generalized poorly. CNNs addressed local pattern detection and parameter sharing, but they did not automatically solve data requirements, robustness, or interpretability. Their follow-on influence includes deep vision models, residual networks, and Vision Transformers.

References

Concept Check

Test The Core Intuition With One Question

This check supports reflection and never blocks the next chapter. Records stay on this device.

What is the key mechanism when one convolution kernel slides across an image?