This paper presents a comparison between Principal Component Analysis (PCA) and Independent Component Analysis (ICA) for face recognition. The authors used the FERET database to evaluate the performance of these methods. PCA is a popular unsupervised method that finds a set of basis images by maximizing the variance of the data. ICA, a generalization of PCA, is used to find statistically independent basis images, which may be more effective for capturing high-order relationships among pixels in face images. The study used two different architectures for ICA: one where images were treated as random variables and pixels as outcomes, and another where pixels were treated as random variables and images as outcomes. The first architecture produced spatially local basis images, while the second produced a factorial face code. Both ICA representations outperformed PCA in recognizing faces across days and changes in expression. A classifier combining the two ICA representations achieved the best performance. ICA is based on the principle of optimal information transfer through sigmoidal neurons. It was applied to the FERET database, and the results showed that ICA representations were more robust to noise and variations in lighting, make-up, and facial expressions. The ICA representations also captured more information about facial identity compared to PCA representations. The study also compared the class discriminability of the ICA and PCA representations. The ICA coefficients consistently had greater class discriminability than the PCA coefficients. The ICA factorial representation (Architecture II) outperformed the PCA representation in recognizing faces across days. The results suggest that ICA is more effective for face recognition than PCA, especially in scenarios involving variations in lighting and expression. The paper concludes that ICA provides a more robust and effective representation for face recognition than PCA, particularly in capturing high-order relationships among pixels. The results support the use of ICA for face recognition tasks, especially in environments with varying lighting and expressions. The study also highlights the importance of considering the architecture of ICA when applying it to face recognition tasks.This paper presents a comparison between Principal Component Analysis (PCA) and Independent Component Analysis (ICA) for face recognition. The authors used the FERET database to evaluate the performance of these methods. PCA is a popular unsupervised method that finds a set of basis images by maximizing the variance of the data. ICA, a generalization of PCA, is used to find statistically independent basis images, which may be more effective for capturing high-order relationships among pixels in face images. The study used two different architectures for ICA: one where images were treated as random variables and pixels as outcomes, and another where pixels were treated as random variables and images as outcomes. The first architecture produced spatially local basis images, while the second produced a factorial face code. Both ICA representations outperformed PCA in recognizing faces across days and changes in expression. A classifier combining the two ICA representations achieved the best performance. ICA is based on the principle of optimal information transfer through sigmoidal neurons. It was applied to the FERET database, and the results showed that ICA representations were more robust to noise and variations in lighting, make-up, and facial expressions. The ICA representations also captured more information about facial identity compared to PCA representations. The study also compared the class discriminability of the ICA and PCA representations. The ICA coefficients consistently had greater class discriminability than the PCA coefficients. The ICA factorial representation (Architecture II) outperformed the PCA representation in recognizing faces across days. The results suggest that ICA is more effective for face recognition than PCA, especially in scenarios involving variations in lighting and expression. The paper concludes that ICA provides a more robust and effective representation for face recognition than PCA, particularly in capturing high-order relationships among pixels. The results support the use of ICA for face recognition tasks, especially in environments with varying lighting and expressions. The study also highlights the importance of considering the architecture of ICA when applying it to face recognition tasks.