www.lesswrong.com/posts/5jbD2WQDy4yWHTtwL/a-lack-of-introspective-ability-is-not...
1 correction found
the only way we've ever figured out how to teach a computer to do it is by generating a new, also uninterpretable de novo algorithm through deep learning.
This overstates the history of face recognition. Computers were recognizing faces decades before deep learning, using classical methods such as geometric-feature systems and eigenfaces/PCA-based approaches.
Full reasoning
This sentence is contradicted by the published history of computer face recognition.
- A 1991 MIT paper, "Eigenfaces for Recognition," describes a computer system that can "recognize the person" from facial images using a PCA/eigenfaces method — long before the modern deep-learning era.
- A 1994 MIT Media Lab publication reports "experiments with eigenfaces for recognition, verification, and interactive search" on a large face database.
- OpenCV's official face-recognition documentation also describes several non-deep-learning face-recognition methods, including Eigenfaces, Fisherfaces, and Local Binary Patterns Histograms (LBPH), and notes that early automated systems used geometric facial marker points.
- NIST's Face Recognition Vendor Test 2006 evaluated face-recognition algorithms in 2006 and reported large-scale performance results years before deep learning became the dominant approach.
So the incorrect part is the universal claim that the only way we've ever taught computers face recognition is through deep learning. Deep learning is one major approach, but not the only one, and not the first one historically.
4 sources
- Eigenfaces for Recognition | Journal of Cognitive Neuroscience | MIT Press
We have developed a near-real-time computer system that can locate and track a subject's head, and then recognize the person by comparing characteristics of the face to those of known individuals.
- View-Based and Modular Eigenspaces for Face Recognition - MIT Media Lab
In this work we describe experiments with eigenfaces for recognition, verification, and interactive search in a large-scale face database.
- OpenCV: Face Recognition with OpenCV
The currently available algorithms are: Eigenfaces ... Fisherfaces ... Local Binary Patterns Histograms ... One of the first automated face recognition systems ... used marker points (position of eyes, ears, nose, ...) to build a feature vector.
- Face Recognition Vendor Test 2006 and Iris Challenge Evaluation 2006 Large-Scale Results | NIST
The Face Recognition Vendor Test (FRVT) 2006 and Iris Challenge Evaluation (ICE) 2006 are independent U.S. Government evaluations of face and iris recognition performance.