UCLA

Learning representations from human neural activity is constrained by limited supervision, participant-specific sampling, heterogeneous recordings, and distribution shifts across institutions, tasks, and brain states. This dissertation develops representation-learning strategies under these constraints as targets progress from clinical events to episodic content during recall and sleep. Meaningfulness is operational: a representation must relate to its target, remain informative under variation, and be evaluated with evidence distinct from training. The first setting begins with high-frequency oscillation candidates proposed by legacy detectors but lacks exhaustive and consistently defined pathological labels. A self-supervised variational model learns event morphology, clustering discovers weak labels, and a classifier refines the decision boundary using real and generated latent examples. This approach improves patient-level evaluation across multi-institutional datasets while preserving candidate generation and expert review as separate parts of the workflow. Omni-iEEG provides a harmonized benchmark spanning 302 patients, 178 hours, and more than 36,000 expert-validated annotations, enabling representation choices to be compared under common targets, splits, metadata, and clinical evidence. The second setting asks whether supervision available during one naturalistic experience can support inference during later internally generated cognition. Participant-specific multi-region attention models trained only on densely annotated movie viewing produce concept activations that preferentially precede matching verbal reports without recall labels for training. Models trained using only medial temporal lobe (MTL) spikes significantly decode concepts before but not after sleep, while models trained using only frontal cortex (FC) spikes decode concepts after but not before sleep. These combined issues create a single-exposure transfer-and-evaluation problem, where representations learned from sparse neuronal recordings during initial experience poorly generalize for evaluation in states with substantially different distributions. The final study addresses this problem using a dual-head objective that combines multi-label classification with contrastive concept alignment, followed by Recall-Adapted Sleep Alignment, which couples unlabeled sleep-domain adaptation with recall-based semantic regularization. Across 62 held-out concept-selective neurons, the adapted representation shows stronger concept-specific temporal coupling than un adapted and comparison models. Together, these studies show that progress in neural representation learning depends not only on model architecture, but also on how candidate spaces, supervision, distribution shifts, and independent evaluation evidence are jointly designed.