The latent low-rank representation (LatLRR) is applied in the field of pattern recognition as a classical unsupervised feature extraction algorithm. However, the dimensions of the features obtained using the algorithm cannot be reduced. Two low-rank matrices are separately learned by the algorithm such that the overall optimality cannot be guaranteed. Furthermore, the algorithm ignores the samples' residuals in the learning process. This study proposes a joint projection learning algorithm based on the LatLRR to address these problems. First, projection and reconstruction matrices are used to approximate the low-rank projection matrix in the LatLRR such that the algorithm can extract discriminative features while reducing the samples' dimensions. Second, the projection, reconstruction, and low-rank matrices are jointly learned by the algorithm such that they can be mutually boosted. The obtained projection can extract more discriminative features. Simultaneously, the samples' residuals in the process of learning are constrained in the algorithm model. Third, the alternating iterative method is used to solve the model. Experiments on multiple datasets show that the algorithm can effectively reduce the samples' dimensions while further improving the discriminative ability.