Introduction: Owing to recurrent seizures, temporal lobe epilepsy (TLE) patients show neural reorganization, which has especially been studied in language and memory networks (Baciu & Perrone-Bertolotti, 2015; Sidhu et al., 2013). Although these two functions have mainly been studied separately, recent evidence indicates their dynamic interaction (Duff & Brown-Schmidt, 2017). Driven by this neural coupling, we are interested in exploring abnormal patterns of activation within the joint language and episodic memory network in TLE patients over the tasks that engage one or the other function or both simultaneously. However, the problem of exploring complex structures is related to the high dimensionality of data and joint variation between the regions, which, in addition, to sample size, restrict the usage of standard statistical analysis. We will present an application of non-linear manifold learning for fMRI data dimension reduction as one way of addressing these issues. Methods: The GE2REC fMRI protocol (Banjac et al., 2019) was used with 21 healthy adults and 12 left TLE patients to obtain language data through sentence generation (GE), episodic memory data through visual recognition (RECO) and mixed language and memory data through recalling and generating sentences (RA). The ROIs were defined as the regions of Brainnetome atlas (Fan et al., 2016) in which more than 40% of the voxels were activated (p <.001, uncorrected; k > 20) in the control group during the tasks. The beta weights were extracted using the Rex toolbox. In order to explore the deviation of language and memory network activation patterns in TLE patients, the following steps were performed (Tilquin et al, 2019): 1. Dimensionality reduction by using the Umap (McInnes, Healy & Melville, 2018) to convert a set of standardized ROI beta values into a unique point in a manifold subspace representing one participant during one task. 2. Model estimation based on the formula Y=f(x)+ε. Where Y is the control data consisting of values for all ROIs in real space, x is the matching point in the reduced space and ε residuals. The f(x) represents the non-linear kernel regression function between the reduced and natural space. 3. TLE patients analysis by projecting them onto the learned manifold space where the projection point reflects the position that a patient would occupy if they belonged to the control group. The ε depicts the abnormalities in the patient’s dataset in the case when it is greater than the model variability obtained by the leave-one-out procedure. Altered values of the residuals for each ROI were identified by using a Z-score with a p < .001 threshold. Results: The reconstructed manifold differentiated the tasks, with mixed RA task situated between those of language (GE) and memory (RECO) as expected. The GE and RA tasks positioned close to each other, while the RECO task was more distant, which could be due to more similar cognitive processes engaged in the GE and RA tasks. The tasks were more dispersed and superposed for the patients and there was not any patient who projected within the majority of controls for all the tasks. The overlap of tasks in patients seems to correspond to higher involvement of the posterior medial system (Ranganath & Ritchey, 2012) than predicted regardless of the task. It appears that, for the majority of patients, there is a disruption within this system reflected by opposite residuals across the tasks between parahippocampal and posterior superior parietal lobule. The illustrative single case (Figure 2) shows how residuals reflect impairment of the ventral semantical pathway. Conclusion: The manifold learning shows abnormalities within high-dimensional networks. This approach enables connecting each point to the original data in order to better understand the reduced space and provides a measure of deviation. This approach is especially valuable for small sample studies or those comparing single-patient with a control group.