Program in progress!

Human sleep consists of multiple stages (Wake, N1, N2, N3, REM), each characterized by distinct neurophysiological processes. Traditional sleep staging relies largely on power spectral features (delta, alpha, spindle/sigma, beta) extracted from individual EEG channels. However, sleep is also a highly network-driven process: slow waves exhibit large-scale synchrony, spindles reflect localized thalamo-cortical coupling, and REM sleep shows desynchronized, small-world-like interactions.Despite this, frequency-resolved functional connectivity [1], has been underused in sleep-stage research and rarely incorporated into automatic staging systems. Frequency-dependent covariance allows to capture signatures of collective behavior and of critical behavior, typical of dynamic phase transitions, across diverse frequency bands. Furthermore, the spectral analysis of frequency-dependent covariance may allow projecting brain activity onto a basis of spatiotemporal patterns, facilitating time-dependent dimensionality reduction. In brief, this quantity may reveal network-level signatures and traits of collective behavior - beyond the single-channel behavior -, invisible to traditional methods. This project aims to design an algorithm to efficiently evaluate frequency-resolved connectivity in EEG data, and to systematically evaluate its utility for sleep-stage characterization. Eventually, we would develop a prototype automatic sleep staging algorithm leveraging the network features unveiled by this quantity.

Goals for the BrainHack:

• Design an algorithm to efficiently estimate the Frequency-Resolved Covariance matrix, and reproduce the fundamental results of reference [1].
• Characterize connectivity patterns across sleep stages using frequency-dependent covariance and related measures. An interesting quantity in this context could be the PCA-based effective dimensionality of the global signal as a function of the frequency, in different sleep stages.
• Identify robust network signatures that consistently differentiate stages within and across subjects.
• Explore dynamic connectivity within-stage variability and transition dynamics.

• Minimal Python coding skills
• Data manipulation

Large language models like BERT and GPT have revolutionized natural language processing but often encode social biases, particularly gender stereotypes. These biases associate certain professions or traits with specific genders, reflecting and sometimes amplifying societal prejudices. This project focuses on uncovering subtle gender biases embedded within the internal representations of these models, even without explicit prompts. Using innovative methods, we measure how likely models are to link professions with male or female pronouns through crafted test sentences. We then develop a mathematical technique to adjust the model’s internal word and sentence representations to neutralize these biases. Additionally, we investigate whether reducing occupational gender bias influences other biases, like those tied to family roles or personal traits, to understand the broader structure of gender bias in AI. This work aims to contribute toward a computational model of stereotypes that can be validated via social experiments.

Goals for the BrainHack:

• Quantify gender bias in BERT's occupational embeddings using log-probability bias scores (a rigorous metric from Kurita et al., 2019)
• Learn an optimal linear transformation that remaps BERT's latent space to reduce gender bias while preserving semantic information
• Test generalization across multiple domains (occupations, family roles, personality traits, sentiment) to ensure the debiasing method is robust

• Python Coding
• Statistics
• NLP and Embeddings
• Social Psychology
• Critical Thinking
• Good Vibes ✨

Human brain is able to track hierarchical information embedded in naturalistic continuous speech. What about a phantom head with no brain but an artificial ear? The project aims to decode speech features information embedded in stories presented as audiobooks using activity recorded from an EEG phantom head equipped with a device converting sounds into electrical signals. The hierarchical speech features that we will try to decode include low-level acoustic features (e.g., envelope and spectrogram) and higher-level linguistic features (e.g., word frequency and lexical surprisal). We will assess how well this ‘brain-less’ EEG captures different information embedded in the speech signal by directly comparing it to real-brain EEG. What makes this project exciting is that we will try to test speech features decoding in a fully controlled, artificial system. To get started, one can begin by familiarizing themselves with EEG recording data and acoustic and linguistic features that can be extracted from continuous speech. Whether you are an expert or a complete beginner in decoding continuous, naturalistic information from EEG signals, it is a unique opportunity to learn and master cutting-edge neuroscience skills to work with ecological continuous stimuli. Plus, how many people can say they’ve worked with EEG activity from a phantom head?

Goals for the BrainHack:

• Day 1) Check if we are equipped with correct matrices of speech features and EEG data to work on decoding scripts using mTRF toolbox in MATLAB. Develop the method to decode acoustic features (i.e., spectrogram and envelope) from EEG phantom head and compare it with real-brain EEG data
• Day 2) Apply the developed method to decode other speech features (e.g., phoneme onset, word onset, lexical surprisal) and extract more complex linguistic information
• Day 3) Compare accuracy of each model between phantom EEG sessions and real-brain EEGs, and summarize findings in a report or presentation

• Knowledge of acoustic features in continuous speech
• Knowledge of semantic features in continuous speech
• Linguistic knowledge
• Decoding
• Matlab
• Python
• Sharing ideas

Soundscapes conceptualize acoustic environments in terms of how they are perceived, capturing higher‑order attributes such as affective quality and sound‑source composition beyond purely acoustic or psychometric measures. Although neuroscientific research increasingly adopts naturalistic paradigms, neurophysiological evidence linking soundscape attributes to brain activity remains scarce, largely due to the lack of EEG datasets recorded during real soundscape exposure. In this project, we aim to develop a proof‑of‑concept framework for inferring neurocorrelates of dynamic soundscapes by combining two complementary open datasets: (i) a corpus of annotated urban soundscape recordings, and (ii) an EEG dataset with continuous affective ratings (valence and arousal) of urban images. First, we will place the soundscape recordings in the circumplex affective space by extracting audio features and deriving perceptual descriptors from the literature and available annotations. In parallel, we will map the EEG‑derived affective ratings of urban images into the same affective space. By aligning audio-derived affective features with EEG affective responses obtained in a visual modality, we will test whether shared affective structure enables the construction of an encoding model that predicts EEG spectral properties from soundscape acoustic and source‑based features. This exploratory cross‑modal framework will allow us to assess whether affective descriptors of real sound environments can, in principle, be linked to neurophysiological markers, even in the absence of direct EEG‑during‑soundscape datasets.

• EEG
• Soundscapes
• Deep learning
• Audio & acoustics

• What are you doing, for whom, and why? Making a system to help students like us focus utilizing a commercial EEG. With our system we hope to demonstrate the possibilities of EEG-based focus apps.
• What makes your project special and exciting? While there are other products like Muse which track focus they don’t have an interface that actually helps train focus and their system is quite costly. Ours will make use of an open source software and more cost effective device.
• How to get started? First we’ll stream and build a classifier to understand when a user is in the alpha frequency band. Then, we’ll develop an interface to help guide them when they aren’t based on previous literature regarding meditative states and alpha.
• Where to find key resources? NeuroPawn and OpenCortexBCI documentation

• Interface design
• Frontend/backend
• Data Science