We navigate a sea of symbolic representations of how machines see the world. These representations have developed into an aesthetic shorthand with machine vision, and are supposedly designed to allow us to gain some understanding of what is going on inside the black box of how a machine sees. Of course, a machine doesn’t ‘see’ in terms of understanding. How we see the world is not the same as how a machine sees the world, and the feature visualisation – from green bounding boxes to meshes over faces – form an aesthetic language of representation of how we see AI seeing (or rather, collecting) the world. Feature visualisations are representations of how a machine sees that are understandable by humans, but as they are not actually how machines see, then what are we understanding?
When talking about computer vision, I’m not talking about computer graphics, and it is not quite captured by the idea of image processing. It is also most definitely not data visualisation. But the representations of how machines see the world, through the help of feature visualisation are somewhat misleading as to what is going on.
With facial recognition technologies – automatic recognition of human faces by comparison to a database – a particular aesthetic shorthand occurs time and time again. A photograph of a human face, presented facing forward with eyes open, without expression, and usually very well lit is overlaid with nodes and lines showing angles and shapes of key features. Giles Bergel, of the Visual Geometry Group at Oxford University calls this a Cabalistic reticule (Kabballah = the ancient Jewish tradition of mystical interpretation of the Bible, first transmitted orally and using esoteric methods (including ciphers) and reitucla = a fine network or net-like structure).
1 Geometric algorithms
Early facial recognition algorithms are still used today in a modified form. They rely on biometrics, measuring facial features such as the distance between a person’s eyes, and turning this two-dimensional image into a set of numbers that describes a particular face. For this reason, this geometric algorithm (or template based) facial recognition method is also known as a feature-based method. The recognition process then compares these vectors to a database of faces which have undergone the same mapping process.
1 The template-based methods can be constructed using statistical tools including SVM (Support Vector Machines), PCA (Principal Component Analysis), LDA (Linear Discriminant Analysis), Kernel methods, Trace Transforms.
2 Neural Networks
TAGS: Feature visualisation; metaphor; facial recognition; operational images;