In the context of M/EEG source reconstruction, most distributed source models tend to strongly overestimate the spatial extent of brain activity and underestimate its depth and amplitude (Black contour line: boundary of the “aud-rh” region given by MNE-Python).

The top row shows the results of the MNE and eLORETA algorithms (implemented in MNE-Python with the default parameters), which suffer significantly from these limitations (although eLORETA manages to estimate depth). Sparse Bayesian Learning (SBL, bottom left) underestimates the spatial extent while accurately locating the activity. Combining SBL with spectral graph wavelets, as shown in the bottom right panel, correctly locates the activity (red and green lines are level curves corresponding respectively to 1% and 10% source amplitude levels), estimates its spatial extent and depth, and yields a quantitatively relevant amplitude estimate.

We compared three machine-learning approaches for segmenting discourse units. The performance of each method (“f_score”) is plotted against the number of training tokens (“size”, in log-scale). “weaksup” means data-programming weak supervision, in which the training set is annotated by an automatic noisy labeller based on a manually written set of labeling rules; “finetune” means fine-tuning an LLM (RoBERTa) with varying amounts of hand-labelled data;  “kshot” means prompting GPT-2 with different numbers of examples. Standard fine-tuning of an LLM emerges as the most effective method. It reaches the same performance as the “weaksup” approach while relying on a more straightforward training procedure. The prompting “kshot” approach was lagging behind. Although the models used with the prompting approach are improving rapidly, their intrinsic opacity makes systematic error analysis almost impossible.

Laurent Prévot and Philippe Muller. 2025.
Dialogue & Discourse 16 (2): 35–73. —  @HAL