Intra- and inter-subject variability of EEG signals make brain-computer interface calibration necessary. This phase decreases the available operational time and can cause fatigue. Our goal is to reduce this time, i.e. the number of EEG trials required to learn the preprocessing, feature extraction and classification parameters. Self-Supervised Learning (SSL) offers new approaches, in particular to increase the number of training examples for a classification task. Thus, fewer trials are needed for calibration. Banville et al. [1] propose the SSL-RP (Relative Positioning) method, which learns a latent space to discriminate whether two EEG segments are close or far apart in time. Numerous segment pairs can be generated from a limited amount of data. The latent space can then be used to extract features and teach a classifier to discriminate segments of kinesthetic motor imagery (KMI) from those of a resting period. We propose to apply the UMAP projection method [2] following an EEGNet projector to form a new latent space (Fig. 1). UMAP well preserves the separability of KMI periods from resting periods, even in a 2D latent space (Fig. 1). We compared on dataset 2a of the BCI competition IV this SSL-RP-UMAP method with other methods that use a latent space to classify, such as EEGNet and CSP+LDA. The classification task discriminates between EEG windows of right hand KMI and resting periods (and another one between left hand KMI and resting periods). Results are averaged across the 9 subjects and for both the left hand KMI vs resting task and the right hand KMI vs resting task. Using only 22 trials for training, for example, EEGNet and SSL-RP test accuracies are over 94%, while CSP+LDA and SSL-RP-UMAP are 89%. The discrimination rate of SSL-RP-UMAP is thus comparable to that of CSP+LDA but lower than EEGNet and SSL-RP. The addition of a UMAP projection, despite the separability of the class features obtained, does not facilitate the recognition of KMI periods.