Summary: This study measured sub-second activity within thalamocortical networks and nine thalamic nuclei in the human brain during spontaneous transitions in behavioral arousal state, using ultra-high field fast fMRI. The research discovered a stereotyped sequence of activity across thalamic nuclei and the cingulate cortex that preceded behavioral arousal after a period of inactivity, followed by widespread deactivation. These thalamic dynamics were linked to whether participants subsequently fell back into unresponsiveness, with unified thalamic activation reflecting the maintenance of behavior. Setzer, B., Fultz, N.E., Gomez, D.E.P. et al. A temporal sequence of thalamic activity unfolds at transitions in behavioral arousal state. Nat Commun 13, 5442 (2022).
Summary: In this seminar, we will explore a novel fMRI-to-text decoding framework named MindLLM. It combines a neuroscience-informed, subject-agnostic fMRI encoder with an off-the-shelf large language model to translate brain activity into coherent text. It introduces Brain Instruction Tuning (BIT), which enriches the model’s capacity to extract and represent diverse semantic information from fMRI signals, enabling versatile decoding across different tasks and subjects. We will discuss how to implement these techniques in our study. Qiu, Weikang, et al. “MindLLM: A Subject-Agnostic and Versatile Model for fMRI-to-Text Decoding.” arXiv preprint arXiv:2502.15786 (2025).
Summary: In this seminar, we will explore the structure of an open-source rtfMRI-NF toolbox. They use Python and MATLAB to preprocess data, generate feedback, and provide it.
Summary: In this study, to enhance our understanding of visual processes, they developed WAVE, which reconstructs visual stimuli from fMRI data. By integrating three modalities (fMRI, image, and text) to perform contrastive learning, the features are then passed to a diffusion model for final image reconstruction.
Summary: The study performed traditional binary GWAS, continuous univariate GWAS using wearable combination scores, and multivariate GWAS to identify genetic variants associated with ADHD. The identified variants showed associations with heart function (MYH6, CMTM5) and ADHD-related genes (ELFN1), with some variants potentially having a protective effect against ADHD.
Summary: This study introduces BrainRGIN, a novel graph neural network (GNN) model designed to predict intelligence using resting-state fMRI data. By leveraging graph isomorphism networks and clustering-based embeddings, the model effectively captures brain sub-network structures. The authors validate their approach using the Adolescent Brain Cognitive Development Dataset and demonstrate superior predictive performance compared to traditional machine learning models.
Summary: A meta-analysis of existing literature on the effects of slow-paced breathing on cardiovascular indices, including HR, HRV, and BP, as well as on negative emotions.The training showed a moderate effect in reducing SBP, moderate-to-large effect in increasing time-domain HRV, and a small effect in reducing HR.Also, slow-paced breathing may reduce negative emotions such as perceived stress.Long-term effect of slow-paced breathing was found reducing SBP and DBP among prehypertensive subjects. Shao, R., Man, I.S.C. & Lee, T.M.C. The Effect of Slow-Paced Breathing on Cardiovascular and Emotion Functions: A Meta-Analysis and Systematic Review. Mindfulness 15, 1–18 (2024).
Summary: This study presented the first controlled study of a short, on-road breathing intervention with both calm and stressful driving conditions with a sample of experienced drivers familiar with the regular, daily, commuting experience in the US.Their stress inducing task confirmed that people were more stressed during the stressor inducing condition.Also, for those who engaged with intervention showed decrease in breathing rate during normal driving was about 15% or about one half of the intended decrease.
Summary: Here they presented BreatheBuddy, a passive respiratory sensing system that monitors comprehensive breathing biomarkers during breathing exercises in real-time using earbud’s accelerometer.Their evaluation with independent test users shows quite accurate performances. The interfaces they presented facilitates real-time breathing biofeedback to potentially make earbud as an effective tool for breathing exercises towards stress relaxation.
Summary: This study examined the effects of six breaths per minute breathing, soothing rhythm breathing, and nature video viewing on HRV, respiratory rate, and emotional regulation through a randomized controlled experiment. The results showed that both breathing techniques significantly increased HRV (SDNN, LF HRV, LF/HF ratio), with six breaths per minute breathing being the most effective, while nature video viewing had no impact; however, HF HRV did not show significant changes across conditions, and emotional responses to breathing exercises were not significantly different between clinically at-risk and non-at-risk participants.
Summary: This study evaluates the test-retest reliability of heart rate variability (HRV) and cardiopulmonary coupling in healthy individuals by analyzing ECG and respiration data collected one week apart. The results indicate high reliability for mean heart rate (HR), but greater variability in RMSSD, highlighting the importance of careful HRV metric selection in research.
Summary: This paper proposes a method to directly align language models using human preference data without reinforcement learning. It simplifies the optimization process by utilizing a binary classification loss based on human preferences, eliminating the need for explicit reward model training. This approach is simpler, more stable, and more efficient than RLHF.
Summary: ReST is a reinforcement learning algorithm that improves language model policies by sampling outputs from an initial model and refining them using offline RL. It enhances data reusability, reduces computational costs, and aligns outputs with human preferences.
Summary: RLCD aligns language models using contrastive distillation by generating preference data from positive and negative prompts. A reward model trained on these preferences refines the model via reinforcement learning. This approach reduces noise in preference data and achieves superior alignment compared to RLHF.
summary: This study explored how the hierarchical structure of scene grammar is reflected in our object recognition. Scenes are divided into several “phrases,” each consisting of a central “anchor” object and its surrounding “local objects.” Participants consistently judged object pairs within the same phrase to be more similar, and this tendency was observed across both images and words. This demonstrates that the hierarchical structure of the visual environment is integrated into our abstract mental representations. Consequently, this study is expected to provide insights into how stimuli can be extracted from natural scene data, such as COCO images.