Principal component analysis (PCA for short) is a standard multivariate analysis technique whose origin can be traced to Cauchy (1815) for the mathematics and Galton (1877) or Pearson (1901) for the statistical or geometric aspect. In the mid and late 1980's, several researchers independently suggested that PCA could be used to analyze face images (i.e., Abdi, 1988; Sirovich & Kirby, 1987; Turk & Pentland,1991). From the beginning, the PCA model was seen both as a convenient way of analyzing the information in a set of images and as a model of human face recognition. From a psychological point of view, the PCA model insists on the relevance of the statistical properties of faces for human recognition. Because PCA can also be implemented as a statistical learning algorithm (e.g., such as neural networks), it has also been seen as a model of learning. These two interpretations of the PCA model (as a tool and as a model) will serve to structure this talk where I will review the first 20 years of the PCA model, mainly from a psychological point of view (computer vision applications are now too numerous to be reviewed).