The North Atlantic sea level pressure (SLP) variability is usually studied with the help of linear methods (e.g. rotated or non-rotated PCA). Some problems in application of linear methods in SLP-related studies may originate from the fact, that climate system is non-linear in its origin. Our study may be regarded as an attempt at non-linear description of North Atlantic SLP variability. We used the database of reanalyzed monthly mean SLP values from NCAR (geographical grid 2.5x2.5° from January 1948 to December 1998). The region of interest was restricted to 90°W - 50°E and 10°N - 80°N. With respect to the fact that the zonal distances between grid points decrease poleward, the data set was transformed into more „regular“ grid with both zonal and meridional distances between the neighbouring grid points corresponding to 5° at the equator. At each „regular“ grid point the SLP anomalies were calculated month by month and averaged into mean seasonal anomalies (overlapping 3-month seasons). Finally, only „winter“ seasons NDJ, DJF and JFM were selected from the data set. These data were then pre-processed by the means of linear non-rotated PCA, which was not used here for pattern recognition but only for dimensionality reduction. The PC1 explains about 45 % of the original SLP variability, PC2 about 14 % and PC3 about 12 %. The scores of PC1 - PC3 were linearly transformed to the interval from -1 to +1 and then non-linearly combined with the help of 5-layers autoassociative neural network (input layer: 3 neurons with linear activation, „encoding“ layer: 3 neurons with hyperbolic tangens activation., „bottleneck“ layer: 1 neuron with linear activation, „decoding“ layer: 3 neurons with hyperbolic tangens activation and output layer: 3 neurons with linear activation). Cross-validated BEP (back error propagation) network training with early stopping and a weak Weigend regularization was used. Random division to training/verification/testing subsets was repeated 20 times, for each division 10 trainings from random initial conditions were carried out. The best network was selected with respect to training, verification and testing errors (networks with low errors of similar magnitude were preferred) and predictor/predictand correlations for all 3 PC scores (the ability of the network to decode the PC scores correctly after their encoding). Trained network was then split into encoding (from the input layer to the „bottleneck“ layer) and decoding (from the „bottleneck“ layer to the output layer) parts. The „scores“ of the non-linear PCA (NLPCA) were calculated from the input data with the help of encoding part of the network. Empirical quantiles of NLPCA „scores“ from 5% to 95% with the step of 5% were then calculated and processed by the decoding part of the network. Scores of PCA1-3 were reconstructed in this way for the empirical quantiles of NLPCA „scores“ mentioned above and after that the output patterns (modelled SLP anomalies) were calculated. Results indicate some non-linear phenomena connected with North Atlantic SLP variability (non-linear transition from the negative to the positive NAO phase, the role of SLP in Scandinavian region, etc.)