~/projects/cnn-face-recognition $ cat README.md
How CNNs and humans recognise faces
Do neural networks recognise faces the way we do? My bachelor's thesis.
- Literature synthesis
- Cognitive psychology
- Computational vision
// the problem
CNNs reach human-level scores on face recognition benchmarks. But a high score doesn't tell you whether they process faces the way humans do.
// what I built
A literature review comparing CNNs and humans on four levels: performance, internal representations, the errors each makes, and their biases. Title: “Facing the Future: How CNNs and Humans Recognize Faces, and Where They Fall Short.”
// result
Graded 9/10. In short, CNNs and humans reach similar scores but get there in different ways. Humans rely on familiarity and holistic processing, networks are vulnerable to adversarial changes, and both carry demographic biases.
// what I took away
If you want to compare brains and networks, the mistakes they make tell you more than their accuracy. That question is a big reason why I'm doing IT & Cognition now.