What would an AI have us try next?

The lab actively researches affective neuroscience, including awe. Interpretable EEG foundation models can bridge the gap between complex model predictions and physiologically meaningful insights into emotions.

Train an EEG foundation model on an affective EEG dataset (e.g., emotions induced by music/videos). Apply EEG-PRISM to convert time-channel attribution scores into frequency-domain power changes, then correlate these changes with self-reported affective state ratings (valence/arousal) using statistical measures.

The lab already works with multimodal genetics and connectomics. Agentic AI can now integrate vast, complex datasets to identify subtle, non-obvious patterns, accelerating hypothesis generation for personalized medicine.

Train an agentic AI with access to structural and functional connectome data (e.g., from ENIGMA-OCD or ABCD), polygenic scores, and clinical phenotypes. Evaluate its generated hypotheses against expert-curated ones using a panel of neuroscientists for novelty and plausibility metrics.

EEG-JEPA offers structured latent prediction rather than plain masking. Carrying it over to dynamic connectomics could give the lab's fMRI foundation models and polygenic risk score (PRS) association work a new source of insight.

Extract dynamic functional connectome time series from ABCD fMRI, train a JEPA-style model with an N-MET-like masking strategy informed by inter-regional anatomical connectivity (tractography), and test how useful the learned latent space is for PRS and psychopathology prediction.

EEG-PRISM opens a way to read a model's internal behaviour in physiologically meaningful terms, which would accelerate clinical application of the lab's EEG foundation models such as DIVER-0.

Transform DIVER-0 attention maps into frequency and source space with the EEG-PRISM method, then check how far the transformed interpretation agrees with established EEG analysis knowledge, using expert review and statistical similarity measures.

EEG-PRIME attacks the generalization problem, and the lab's QML expertise could push prototype expressivity further. It is a chance to show a practical benefit of QML on high-dimensional EEG data.

Build a 'Q-ProtoEEG' model that applies quantum feature maps within the EEG-PRIME framework on a public EEG dataset (e.g. BCIC IV 2a), and compare cross-dataset and cross-subject classification accuracy against EEG-PRIME and non-quantum baselines.

BLPM and EEG-JEPA enable semantic and state prediction. Combining them with the lab’s memory reconstruction and latent causal discovery can turn plausible generation into testable causal control.

Use existing affect-context memory EEG and stimulus descriptions to train a CELP+N-MET encoder with a causal latent head. Compare Recall@10, affect CCC, and intervention specificity against the current generator.

ZIPBrain enables local deployment, while SW-ProxyCE exposes a new attack surface. Testing both together extends the lab’s trustworthy clinical AI work from prediction reliability to model security.

Add ZIPBrain pooling to DIVER-0 and train on TUAB EEG with SW-ProxyCE examples from a public encoder. Measure CPU latency, MACs, balanced accuracy, and attack success rate on a private head.

EEG-PRIME shows that conditioned prototypes can improve transfer. The lab can extend this idea to genotype–fMRI alignment using its developmental genetics and foundation-model expertise.

Use ABCD genotypes, resting fMRI, adversity, and depressive symptoms with NeuroMamba plus a PRS-conditioned prototype head. Test held-out-site R², calibration error, and bootstrap prototype ARI.

The lab's art-science and EEG-reconstruction threads already link brain activity to aesthetic experience; QRC's strength on complex time series makes a continuous, art-specific semantic alignment feasible now.

Record EEG during abstract art and music; extract CLIP/audio embeddings from the stimuli; train a QRC to predict those embeddings from EEG and compare cosine similarity against classical reservoirs and RNNs.

The lab has a strong commitment to the convergence of neuroscience and art, with projects like 'EEG Reconstruction & Interactive Media' and 'Revisiting Your Memory'. However, current foundation models can be computationally intensive, limiting real-time interaction. Recent external work on ZIPBrain and understanding/correcting low-frequency bias directly addresses the need for faster, more deployable, and robust EEG foundation models. This synergy can enable truly interactive and responsive neuro-art experiences that are both scientifically grounded and artistically innovative.

Integrate the ZIPBrain redundancy-aware token pooling module into an existing EEG-guided audiovisual generation pipeline (e.g., the one from 'Revisiting Your Memory'). Additionally, implement a low-frequency bias correction method (as suggested by the 'Understanding and Correcting Low-Frequency Bias' paper) within the EEG encoder. Evaluate the model's performance in terms of generation quality (e.g., FID score for generated images/audio), latency, and computational resource usage on a low-power edge device, comparing it against the original, unoptimized pipeline for reconstructing target affective states from EEG.

The lab already predicts mental-health trajectories from multimodal data; QML's strength in high-dimensional, nonlinear settings suits causal inference that integrates polygenic scores and connectomes.

Train a quantum-enhanced causal-discovery model on the lab's PRS, white-matter tractography, adversity, and longitudinal depression data, then compare causal graphs and predictive accuracy against classical latent-space methods.

The lab's Neuro-X Project aims to decode the universal language of the mind, and the Mind the Gap paper highlights the need for nonlinear and multimodal approaches to align brain and language models. The BLPM (Continuous-Latent Predictive Modeling with Semantic Alignment) offers a concrete architecture for this alignment. Integrating this with the lab's strengths in computational psychiatry and multi-modal genetics (e.g., polygenic risk scores, white matter tractography) will allow for a powerful new paradigm in predicting mental health outcomes and understanding their neural and genetic underpinnings.

Develop a pilot BLPM-inspired model using existing fMRI and EEG data from cohorts with psychiatric assessments (e.g., from the ABCD dataset used in 'Polygenic architecture...'). Use a pre-trained fMRI/EEG foundation model (like NeuroMamba or a Mamba-based EEG encoder) as the Continuous EEG Latent Predictive (CELP) encoder. Train the Multi-Query Semantic Decomposition (MQSD) module to align these neural latents with embeddings from a clinical language model (e.g., fine-tuned BERT for psychiatric symptom descriptions). Evaluate the model's ability to predict clinical scores (e.g., depression severity) and generate text summaries of associated brain states.

The lab has strong expertise in both QML for high-dimensional brain data (e.g., Q-DIVER, Hybrid Quantum Temporal Convolutional Networks) and developing EEG foundation models (e.g., NeuroMamba, EEG Reconstruction & Interactive Media). Recent external advances in EEG-PRISM and EEG-JEPA offer new avenues for interpretability and structured learning. Combining quantum advantages in high-dimensional pattern recognition with physiologically-grounded interpretability could lead to truly trustworthy clinical AI for complex neurological and psychiatric disorders, building on the lab's work on trustworthy clinical AI for OCD.

Implement a quantum-enhanced interpretability module by replacing a classical attention or latent space component within an existing EEG-JEPA or EEG-PRIME architecture with a parameterized quantum circuit (PQC). Evaluate the physiological plausibility of the resulting attribution maps (using EEG-PRISM's frequency/source domain mapping) and compare generalizability metrics (e.g., cross-subject accuracy) on a publicly available EEG dataset for a cognitive task (e.g., P300) against the classical baseline.

Recent EEG foundation models (STEAM, EEG-PRISM) demand robust spatio-temporal alignment and interpretability — a natural fit for a lab strong in both QML and state-space models.

Add a parameterized-quantum-circuit alignment module to a STEAM-like model, train on PhysioNet motor-imagery EEG, compare cross-subject accuracy/F1 against classical STEAM, and interpret features with EEG-PRISM.