Discussion: Advancing Brain Decoding and Cognitive Analysis: Leveraging Diffusion Models for Spatiotemporal Pattern Recognition in fMRI Data

[Francode007]

Hi everyone,

Starting this discussion thread by sharing some basic research into this new project. Pattern recognition in functional MRIs is a very exciting task at hand. It could help us understand a lot regarding how our brain changes state throughout a sequence of action/s which I am very curious about. It could potentially help us model “how we think”.

[Francode007]

Introduction: I am Franchis N Saikia, currently working as a Data Analyst II in Walmart International Tech after graduating from Indian Institute of Technology, Guwahati (IITG) ‘23. I am a highly motivated individual seeking every opportunity to work on state-of-the-art projects. Previously, I have worked on the application of deep learning models to medical imaging data (MLP-UNet & an ongoing submission to MICCAI'25) and see this opportunity as an outlet for substantial research directed to understanding the functionality of the brain.

LinkedIN

Below I am sharing a few resources and summaries related to certain aspects of the project.

Functional MRI (fMRI):

A model neuroimaging method, mapping functional areas of the brain activated during a cognitive, motor, or other tasks. Mapping is performed either on the basis of a change in blood flow to a given area (perfusion) or on the basis of a change in blood oxygenation (the so-called BOLD effect). BOLD fMRI (named after the use of the BOLD effect) is the most common way today and has almost become synonymous with the more general name fMRI.
fMRI utilises an MR tomograph for functional mapping comparatively better than existing methods like PET or EEGs in terms of both spatial and temporal quality. [1]

DATASET:

[OpenNeuro: Dataset bank]

BOLD5000 dataset: Slow event related public fMRI dataset with 5000 images and stimuli mitigating the overlap of stimuli while showing images by utilising images from SUN, COCO and ImageNet respectively covering real-world indoor, outdoor scenes and objects in complex, real-world scenes. ~168 GB dataset

Natural Scenes dataset: Large-scale fMRI dataset of 8 healthy adult subjects while viewed thousands of color natural scenes over 30-40 scan sessions. The corresponding research paper discussion a latent diffusion model can be found here.

Generic object decoding: The dataset consists of preprocessed fMRI image cued against a total of 1200 ImageNet images. Two tests were performed: Image Presentation and Imagery Experiment and their fMRI patterns were noted. Preprocessing can be done according to the article Horikawa & Kamitani (2017) Generic decoding of seen and imagined objects using hierarchical visual features. Nat Commun and preprocessed data can be found here.

Human Connectome Project (HCP): Private data, need access, details
Individual Brain Charting (IBC): Large amount of diverse tasks (12), huge dataset ~ 1TB

METHODOLOGY:

1. NeuralDiffuser:

• NeuralDiffuser is built upon Stable Diffusion, utilising a guided image generation. Instead of generating new images, their method tries to reconstruct the original image.
• During training, fMRI voxels are first mapped to subject-shared space and then aligned with ground truth embeddings, CLIP text space, VAE’s latent space and CLIP’s image encoder. During inference, model runs forward a forward and reverse diffusion process with or without guidance for reconstruction.
• Benchmarked on Natural Scenes Dataset

2. MindEye2:

• It improves over MindEye1 requiring only 1 hour of data to reconstruct fMRI to seen images giving competitive results with just 2.5% of a subject’s full dataset.
• Fine-tuned on the 8th subject after pre-training on 7 subjects is the shared-subject functional alignment followed by a MLP backbone fed to a diffusion prior and two MLP projectors (low-level submodule and retrieval submodule).
• The diffusion prior outputs a reconstructed image using OpenCLIP embeddings which is then used to also get a textual description. This is then fused with the output from the MLP projectors to reconstruct the image
• Benchmarked on Natural Scenes Dataset

More methods: Semantic Brain Decoding, Sparse Masked Modeling

[314project]

First-Level AGI Reasoning Module: Kyburg Formula Rearrangement

Overview

This project shares a novel algebraic rearrangement of Kyburg’s epistemological formula, which—according to Perplexity AI—may serve as a reasoning module for “first-level AGI.”
Other AI labs are already exploring this idea, and I want to make it freely available to the developer and research community for open-source development and distribution.

The Perplexity Conversation

Read the full conversation and analysis here:
https://www.perplexity.ai/search/let-s-design-a-reasoning-modul-QwOdVhqQT42qLwIcSfJBxw

Why This Matters

• Novelty: This approach presents a new way to arrange Kyburg’s formula, potentially enabling machine reasoning that qualifies as “first-level AGI.”
• Endorsement: Perplexity AI evaluated the logic and believes it can be deployed as a model in as little as 3 months with current AI technology.
• Open Invitation: AI labs are already investigating this, but I want it to be available for anyone to build, use, or improve in the spirit of open science.

Call to Action

If you are a developer, researcher, or enthusiast interested in AGI:

• Use, experiment with, or extend this idea
• Build an open-source reasoning model for free distribution
• Discuss and improve the concept
• Share it widely

License

I intend this idea to be freely available for any use, including commercial and non-commercial. Developers are encouraged to use a permissive license (such as MIT or Apache 2.0) for any code or models derived from this idea.

---

Shared in the spirit of open science and collaboration. Please cite or reference the Perplexity conversation if you build upon this work.

[prantik-pdeb]

Hi, I am Prantik Deb an MS by Research student in CSE at IIIT Hyderabad working in the Cognitive Science Lab. My research spans medical imaging, large language models, and visual language models. As part of my coursework in Cognitive Science and AI course, I have been exploring brain encoding and decoding work—closely aligning with this project. I am familiar with datasets like BOLD5000 and NSD and am eager to contribute to research at the intersection of AI and neuroscience.

E-mail: prantik.d@research.iiit.ac.in
Thank you.

[skrd-18]

Hi everyone,

I'm Shiva (e0727167@u.nus.edu), a graduate of the National University of Singapore with a degree in Electrical Engineering, where I specialized in signal processing and machine learning. Currently, I'm publishing a paper on novel compression methods for EEG signals, which has given me hands-on experience with handling and analyzing complex neural data.

I am interested to work on the "Advancing Brain Decoding and Cognitive Analysis" project.

I'm excited about the potential to contribute to advancing brain decoding and cognitive analysis, and I look forward to discussing how I can help bring fresh perspectives to this initiative.

I am sharing a few resources too,

Foundational fMRI Connectivity and Parcellation Studies:

• The organization of the human cerebral cortex revealed by intrinsic functional connectivity. Yeo BTT*, Krienen FM*, Sepulcre J, Sabuncu MR, Lashkari L, Hollinshead M, Roffman JL, Smoller JW, Zöllei L, Polimeni JM, Fischl B, Liu H, Buckner RL. Journal of Neurophysiology, 106(3):1125–1165, 2011 [pdf]
  This paper lays the groundwork for understanding how brain regions interact—a crucial step when designing models that need to capture spatiotemporal patterns.
• Local-Global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI. Schaefer AL, Kong Ru, Gordon EM, Laumann TO, Zuo XN, Holmes AL, Eickhoff SB, Yeo BTT. Cerebral Cortex, 29:3095-3114, 2018 [pdf]
  It provides a modern framework for segmenting the cortex into meaningful regions, which is important when converting raw fMRI data into structured forms (e.g., connectivity matrices) suitable for training diffusion models.

Data Preprocessing and Harmonization:

• Goal-specific brain MRI harmonization (An et al., 2022)
  This paper directly addresses challenges in preparing MRI data for analysis by standardizing it across subjects or sessions. A good understanding of harmonization techniques will be critical when preprocessing fMRI scans for diffusion model training.

Temporal Dynamics and Connectivity Fluctuations:

• Resting brain dynamics at different timescales capture distinct aspects of human behavior (Liégeois et al., 2019)
• Interpreting temporal fluctuations in resting-state functional connectivity MRI (Liegeois et al., 2017)

Best regards,
Shiva

[ShubhamAXS19]

Hi everyone,

I'm Shubham Vishwakarma, a recent ECE graduate from DJSCE, currently working as a Data Scientist at a private company that provides data-driven solutions to clients. I have a strong interest in the intersection of AI and Neurology and am eager to contribute to the project "Advancing Brain Decoding and Cognitive Analysis: Leveraging Diffusion Models for Spatiotemporal Pattern Recognition in fMRI Data"

Previously, I worked as a research intern at IIT Patna, where I focused on finding the optimal rank for fine-tuning LLMs in a federated setting and implemented a research paper using PyTorch. I have published a research paper related to Stable Diffusion and am currently working on another paper focused on improving noise variance using RL in federated setting. I am one of the co-authors of this paper, along with my university professor.

Thank you!
Shubham
Email
LinkedIn

[karan-nanda]

Hi everyone,

I’m Karandeep Nanda, currently diving deep into the world of data science as a student in the MSc program at the University of Colorado Boulder. With a background in psychology and biological sciences, I've always been fascinated by the intersection of the brain, data, and technology. My previous work has involved applying machine learning techniques to healthcare and computational biology, and now I'm eager to explore the exciting potential of diffusion models in brain decoding.

I’m particularly drawn to how these models can reveal spatiotemporal patterns in fMRI data and offer insights into the brain’s inner workings. I’m thrilled to be part of this conversation and can't wait to collaborate with all of you as we explore this cutting-edge field together.

Here’s my GitHub and LinkedIn.

Best,
Karandeep

[Francode007]

Hey everyone,

Recently came across this interesting paper on speech processing in the human brain.
This recent study by Google shows that neural activity in the human brain aligns linearly with the internal contextual embeddings of speech and language within large language models (LLMs) as they process everyday conversations. This is quite intriguing on how we are modeling textual and speech cognition in the brain and potentially also feed as extra reference while reconstructing images from fMRI. Recent models utilise multiple embeddings as extra context for brain decoding and we could utilise this to model brain functionality even better.

Paper
Post

[niranjankumarnk]

Hello Everyone,

I'm Niranjan Kumar Kishore Kumar, a Biomedical Engineering graduate currently pursuing my Master’s in Artificial Intelligence at Yeshiva University, New York. I have a strong passion for NeuroAI, computational neuroscience, and AI-driven biomedical research, making this project particularly exciting for me.

My previous work includes:

• Deep learning for medical imaging, including projects on computer vision for predicting cardiomegaly detection and Image segmentation on Bird Sound Denoising.
• Latent Diffusion for Music Generation, where I fine-tuned and developed knowledge distillation on AudioLDM2 (SOTA).
• Machine Learning for Healthcare, including a project on breast cancer prediction using AWS cloud-based ML pipelines.

Project Interests:

For this project, I’m excited to apply diffusion models to fMRI data for spatiotemporal pattern recognition, leveraging U-Net, transformer-based architectures, and denoising probabilistic approaches. Although I have experience in PyTorch, signal processing, and multimodal AI, I am always eager to learn and refine my skills further. I am excited to collaborate, gain deeper insights into fMRI data modeling, and work closely with mentors and the community.

Looking forward to this amazing learning experience!

Best,
Niranjan Kumar Kishore Kumar
Email: nkishore@mail.yu.edu
LinkedIn | GitHub

[marver17]

Hello everyone,
My name is Mario, and I am a PhD student. My research focuses on the application of AI in medical imaging for precision medicine, with a particular emphasis on neuroimaging and neurodegenerative diseases such as Alzheimer's. My work integrates traditional medical imaging tools with deep learning approaches, and I have recently been delving into generative models.
I am particularly captivated by the Advancing Brain Decoding & Cognitive Analysis project , especially the idea of applying diffusion models to fMRI data for brain decoding and cognitive analysis. I am eager to explore opportunities to contribute to this project.

Look forward to collaborate to this amazing project

[Sarah0ravari]

Hi Dr. Mahmoudi and Dr. Kara,

My name is Sadaf (Sarah) Draper, and I’m a graduate student in Computer Science with a background in data engineering, machine learning, and cloud computing. I recently completed a Master's thesis on optimizing solar energy efficiency using GRU, LSTM, and CNN models, and I’m deeply interested in applying deep learning to real-world scientific domains—especially in neuroscience and cognitive analysis.

I came across your GSoC 2025 project, “Advancing Brain Decoding and Cognitive Analysis using Diffusion Models,” and I’m very excited by the opportunity to contribute. I’ve worked extensively with PyTorch, and I’m comfortable designing and training deep learning models, particularly for time-series data. I'm also intrigued by the use of diffusion models for spatiotemporal pattern recognition in high-dimensional data like fMRI scans.

I’d love to get involved, learn more about your vision for this project, and begin discussing ideas for how I could contribute. I’ll begin exploring the forum and any relevant datasets or literature, and I’d be happy to start with smaller tasks or drafts if available.

Looking forward to your guidance!

Best regards,
Sadaf (Sarah) Draper
GitHub Profile: github.com/Sarah0ravari
LinkedIn: linkedin.com/in/sadaf-draper

[NiyatiBisht08]

Hey everyone!

I am Niyati Bisht, a 3rd year B.Tech student in Electronics and Telecommunication at Veermata Jijabai Technological Institute(VJTI), Mumbai. I am deeply passionate about Medical Image Processing in Machine Learning, particularly in the intersection of deep learning and neuro-imaging.

I had the privilege of working as a Research Intern at the Medical Deep Learning and Artificial Intelligence Lab (MeDAL), IIT Bombay, where I focused on MRI-based segmentation using 2D U-Net. My work involved advanced feature extraction techniques, analyzing outputs from intermediate layers to enhance segmentation quality, improving both spatial resolution and feature representation.

Beyond MRI segmentation, I have gained hands-on experience in Diffusion Models, Gaussian Noise, Vision Transformers (ViT), Swin Transformer, RNNs, and CNNs. I am proficient in Python, PyTorch and deep learning libraries relevant to medical imaging.

Currently, I am exploring "Advancing Brain Decoding and Cognitive Analysis: Leveraging Diffusion Models for Spatiotemporal Pattern Recognition in fMRI Data." This project excites me because it aligns with my goal of pushing the boundaries of spatiotemporal analysis in neuro-imaging. My prior experience with MRI and diffusion-based models provides a strong foundation to contribute meaningfully to this project.

Through this opportunity, I aim to refine my expertise in diffusion models for fMRI, explore innovative conditioning techniques for brain decoding, and work on real-world neuro-imaging applications that could aid cognitive analysis. I look forward to contributing my skills while learning from experts in the field.

Warm Regards,
Niyati
niyati.bisht458@gmail.com

[saketlad75]

Hey everyone!

I am Saket, a final-year B.Tech student in Electronics and Telecommunication Engineering at Sardar Patel Institute of Technology (SPIT), Mumbai. My passion lies in Machine Learning for Medical Imaging, with a strong focus on deep learning, signal processing, and neuroimaging applications.

I previously worked as a Research Intern at IIT Bombay, where I developed a deep learning-based contactless palmprint recognition system, improving biometric authentication accuracy. Additionally, I have worked on medical image analysis, including Lung Cancer Detection and Age-Related Macular Degeneration classification, using CNNs, ensemble learning, and GANs for dataset balancing.

My experience extends to Diffusion Models, U-Net, Transformers, and probabilistic modeling, and I am proficient in Python, PyTorch, and TensorFlow. I am particularly excited about "Advancing Brain Decoding and Cognitive Analysis: Leveraging Diffusion Models for Spatiotemporal Pattern Recognition in fMRI Data", as it aligns with my goal of leveraging generative models for spatiotemporal neuroimaging analysis.

Through this project, I aim to explore diffusion-based approaches for fMRI, refine conditioning techniques for cognitive state classification, and contribute to advancing brain decoding research. Looking forward to collaborating and learning from the community!

Best Regards,
Saket Lad
saket.lad@spit.ac.in
LinkedIn

[ignasiialemany]

Hello everyone,

I'm a current postdoctoral researcher at Imperial College London, where I continue my work in numerical simulations for diffusion MRI in biological tissues following my PhD. My academic background spans Aerospace Engineering and Computational Biophysics, which has provided me with a strong foundation in mathematics, numerical methods, and the underlying physics of MRI.

Currently, I am involved in projects that leverage transformers and deep generative models. In this project, I aim to explore diffusion-based models in brain fMRI, refine innovative conditioning techniques for cognitive state classification, and ultimately contribute to advancing brain decoding research. I want to deepen my expertise in diffusion models while enhancing our understanding of cognitive processes.

I look forward to collaborating with and learning from experts in the community as we push the boundaries of neuroimaging research.

ignasi.alemany18@ic.ac.uk
LinkedIn

[prachi-kedar]

Hello all,

My name is Prachi Kedar, and I'm nearing the completion of my Master's in AI at Politecnico di Milano. My journey includes three years of practical experience in AI/ML technologies. Currently, interning for Spanish National Research Council at Department of Functional and Systems Neurobiology , focusing on analyzing neural activity from calcium imaging recordings. Utilizing machine learning techniques, I'm working to correlate this activity with observed behavior. My core interest is in computer vision and image processing, and working on this project will provide me a valuable opportunity to delve deeper into the fascinating intersection of AI and biomedical neuroscience.

I would like to propose below topic based on the given idea:

Project Title: Brain Neural Activity Decoder by using Time-Varying Dependency Structures from fMRI using Graph Neural Network

Have a summarize following steps which can be performed achieve this:

1. Data Preprocessing: Clean fMRI data, parcellate the brain into regions, and segment the time series into dynamic windows.
2. Dynamic Graphs: Calculate time-varying functional connectivity matrices, forming a sequence of brain network graphs.
3. Temporal GNNs: Use GNN models designed for time-series data (e.g., GCRNs, T-GCNs) to learn patterns from these dynamic graphs.
4. Spatiotemporal Denoising/Reconstruction: Use diffusion models to denoise or reconstruct noisy or incomplete sequences of fMRI data, improving the quality of the dynamic graphs.
5. Latent Spatiotemporal Feature Extraction: Extract latent features from the diffusion model that capture both spatial and temporal aspects of brain network dynamics.
6. Training & Evaluation: Train the GNN to perform tasks like dynamic state prediction using time-series appropriate metrics

Dataset :

1. Human Connectome Project (HCP): Good fit for high-resolution fMRI and rich behavioral data.
2. ABIDE (Autism Brain Imaging Data Exchange): Useful for studying autism-related brain connectivity.
3. ADNI (Alzheimer's Disease Neuroimaging Initiative): For research on Alzheimer's disease and cognitive decline.
4. UK Biobank: Large population dataset with diverse fMRI data.
5. OpenNeuro: A broad repository of various fMRI datasets.

I have also attached the link of some research methods which has been already implemented :

1. Spatio-Temporal Graph Convolution for Resting-State fMRI Analysis https://arxiv.org/pdf/2003.10613v3
2. DynDepNet: Learning Time-Varying Dependency Structures from fMRI Data via Dynamic Graph Structure Learning
https://arxiv.org/pdf/2209.13513v3
3. Learning Task-Aware Effective Brain Connectivity for fMRI Analysis with Graph Neural Networks https://arxiv.org/pdf/2211.00261v1

Looking forward to collaborate on this challenging yet interesting project idea !

Best Regards,
Prachi Kedar
prachibalu.kedar@mail.polimi.it
Linkedin | Github Profile

[NiyatiBisht08]

Hey everyone!

I have a few queries regarding the project:

1. GPU Support: Will additional GPU resources be provided for higher computational requirements? This would significantly aid in progressing with the project.

2. Dataset Availability:
- Will the dataset be provided, or do we need to source open-source datasets ourselves?
- If provided, in which format will the dataset be available?

@zeydabadi Looking forward to your response.

Best regards,
Niyati Bisht

[weeebhu]

Hi everyone,

I’m excited to start this discussion thread by sharing some initial research on my GSoC project. Pattern recognition in functional MRI (fMRI) data is a fascinating challenge that could provide deeper insights into how our brain transitions between cognitive states during various tasks.

Why This Matters
Understanding these spatiotemporal patterns could help us model "how we think", leading to advancements in:
-> Brain-Computer Interfaces (BCIs) for assistive technology
-> Early detection of neurological disorders
-> AI-driven cognitive state modeling

My Approach
I’m exploring diffusion models, a class of generative models, to capture complex dependencies in brain signals. Unlike traditional methods, these models can better handle noise, variability, and spatial-temporal relationships in fMRI data.

I’d love to hear thoughts from the community!
-> Have you come across any relevant datasets or papers?
->What challenges do you foresee in applying generative models to fMRI data?
-> Any suggestions on evaluation metrics for this task?

Looking forward to discussing and refining this approach with all of you!

Best,
Mrityunjay Kukreti
GitHub: https://github.com/weeebhu | LinkedIn: www.linkedin.com/in/mrityunjaykukreti

[NiyatiBisht08]

Has anybody tried downloading the dataset? It's too huge. I tried downloading BOLD5000 but it's unzipped file is nearly 512gb. The zipped file is 125gb. Do let me know if anybody has an alternative dataset with lesser size

[Clyde0513]

Hello Everyone!

My name is Clyde Villacrusis (github: clyde0513; email: clyde0513@g.ucla.edu) and I am a 3rd year UCLA Computer Science student, passionate in deep learning, AI, and leveraging skills with Python! I am currently interning at UCLA Health and have been working with AI such as ChatGPT models and Handwritten OCR APIs to convert unorganized blood pressure data into a cleaned, format that doctors can easily read. Thus, I am very excited to work on this new project on advancing brain decoding and cognitive analysis, as it is interesting to learn how spatiotemporal pattern recognition works in our brain and how to better analyze it!

I have been reading a couple of research papers related to this project topic and here is what I have found to get started on this project!

NeuralFlix: Reconstructing Vivid Videos from Human Brain Activity:
-This research paper aims to reconstruct dynamic visual experiences from brain activities. So, they explored to make NeuralFlix, a novel dual-phase framework designed to address the challenges of decoding fMRI data, especially spatial redundancy, noise, and temporal logs

-They do temporal interpolation and spatial masking for contrastive learning of fMRI representations and a diffusion model enhanced with dependent prior noise for generating videos.

-Their methods consist of fMRI feature learning and video decoding two-phase framework for reconstructing videos from fMRI-recorded brain activities.

• Phase 1: Tuning pre-trained fMRI encoder w/ temporal and spatial augmented contrastive learning to align fMRI data with CLIP’s image and text features, thus enhancing the extraction of semantic info from fMRI signals
• Phase 2: Uses trained fMRI encoder to guide a video diffusion model, incorporating prior noise to compensate for fMRI’s low signal to noise ratio

Decoding visual brain representations from electroencephalography through knowledge distillation and latent diffusion models

• They trained an EEG classifier to reconstruct images from ImageNet and THINGS-EEG 2 datasets. So, they’ve analyzed EEG recordings from 6 participants for the ImageNet dataset and 10 for the THINGS-EEG 2 dataset, both of which are exposed to images spanning unique semantic categories.

• The EEG readings are then converted into spectrograms to train CNNs that were integrated with a knowledge distillation procedure based on a pre-trained CLIP based image classification teacher network.

• Their strategy allowed their model to attain a top-5 accuracy of 87%, outperforming normal CNN and various RNN-based benchmarks. In addition, they also allowed an image reconstruction mechanism based on pre-trained latent diffusion models, which allowed them to generate an estimate of the images that had elicited EEG activity.
  Note that their computational experiences and model training were conducted on a server with 4 NVIDIA A100 GPU cards, each with 80 GB RAM connected via NVLINK and a 2 TB of system RAM.

Their codebase is accessible right here:
The way they structure their implementation is shown below:

1. Obtain Data
2. Transform eeg into spectrogram
3. Extract CLIP features from images
4. Train a classifier from CLIP features
5. Build CNN classifier from images
6. Do Knowledge distillation.

These research papers are closely aligned with the challenging, but interesting project idea. I believe that with all the community has been proposing thus far, we can manage to collaborate on this project and complete it in an efficient, timely manner!

Additionally, I also took a data science and deep learning in computer vision course, so I am familiar with data processing, augmenting data, training data, and fine-tuning a model.

Here is my website to learn more about me and my experiences! https://clyde.at/.

Lastly, I plan on collaborating on this project full-time in the summer! I hope to look forward on collaborating with you guys and feel free to contact me and/or reply to my message!

[riyaarah]

Hi everyone! I’m Riya Rahim, a student at IIT Madras with a deep interest in machine learning, data science, and open-source development. I’m particularly excited about the "Advancing Brain Decoding and Cognitive Analysis" project at Emory BMI for GSoC 2025 and eager to contribute. My experience includes working with Python, Flask, SQL, and deep learning, with a strong focus on data analysis and building ML-driven applications. Currently, I’m exploring fMRI data analysis using Nilearn and MNE-Python, along with diffusion models, to better understand spatiotemporal pattern recognition in brain imaging. I’m looking forward to engaging with the community, making meaningful contributions, and learning from experienced mentors. Any guidance on getting started, beginner-friendly issues, or relevant resources would be greatly appreciated. Excited to collaborate and be a part of this journey!

Best regards,
Riya Rahim

[Francode007]

Hi @pradeeban @zeydabadi @monjoybme @anbhimi @abdelrahman725 ,

I have submitted an initial draft of the proposal for this project. Kindly give your valuable feedback before the deadline so that I am able to refine my proposal.

Regards,
Franchis

[prachi-kedar]

Hello mentors, @pradeeban @zeydabadi

I wanted to know while submitting proposal what project size should I need to select ? Also , I have submitted my proposal it would be great if you could provide your feedback for the same .

Thanks,
Prachi Kedar

[unique-777]

Hello,

**Project Title:

Advancing Brain Decoding and Cognitive Analysis: Leveraging Diffusion Models for Spatiotemporal Pattern Recognition in fMRI Data**
2) Abstract / Project Summary:
This project aims to harness the power of diffusion models to enhance brain decoding from fMRI data. By learning meaningful spatiotemporal patterns in neural activity, we can better interpret cognitive states. As a beginner passionate about neuroscience and AI, I plan to explore the integration of generative diffusion models with neural decoding tasks. My goal is to contribute a clear, open-source, and reproducible pipeline that helps researchers better understand brain dynamics using machine learning. I will focus on learning preprocessing, fMRI data interpretation, diffusion networks, and evaluation metrics, building from the ground up with guidance from mentors and the community.
3) Contributor Name: N.VENKATA JAHNAVI
4) Contributor Email and GitHub ID:
Email: jahnavi.venkata777@gmail.com
5) Personal Background (Brief CV):
I am an undergraduate student with a strong interest in Artificial Intelligence, cognitive science, and medical data analysis. I have foundational experience in Python, neural networks, and working with datasets. I recently started exploring neuroimaging and generative models, which inspired me to take up this project. My background includes personal projects in AutoML, emotion recognition, and model evaluation. I am enthusiastic about learning through contribution and eager to build something impactful with guidance from the open-source community.I am researching before GSOC 2025 on the field of hybrid models of ML,blockchain technologies and RAG models.
6) Project Goals / Major Contributions:
Understand the fundamentals of fMRI data and cognitive pattern decoding
Preprocess raw fMRI data using libraries like NiBabel and Nilearn
Learn the theory and architecture of diffusion models
Design a baseline model to capture spatiotemporal features
Apply and fine-tune a diffusion model for brain decoding tasks
Create visualization tools to interpret and analyze model outputs
Document the full pipeline and ensure reproducibility
Engage with mentors and community to iterate and improve the syste
8) Project Schedule
8.1) Community Bonding Period (May 20 – June 16)
Interact with the mentor and understand the goals and scope
Study fMRI datasets, diffusion models, and recent papers
Set up the environment (Python, PyTorch, NiBabel, etc.)
Choose suitable open-source fMRI datasets for development
8.2) Development Phase
Week 1 : Load and explore fMRI datasets, apply preprocessing
Week 2 : Build a basic 3D CNN/LSTM model for spatial & temporal features
Week 3 : Study diffusion models and apply toy examples
Week 4 : Design the architecture combining diffusion and spatiotemporal encoder
Week 5 : Train and debug the hybrid model
Week 6 : Evaluate outputs, tweak hyperparameters
Week 7 : Analyze cognitive state prediction accuracy
Week 8 : Add visualizations and meaningful result interpretations
Week 9 : Polish code, optimize performance
8.3) Final Phase
Write user guides, documentation, and reproducibility instructions
Final testing and evaluation on unseen data
Submit final report and release code on GitHub
Share results with the organization and broader community
9) Planned GSoC Work Hours:
Type: Full-time (35 hours/week)
Timezone: IST (UTC+5:30)
Preferred Hours: 10:00 AM – 5:00 PM, flexible with mentor preference
10) Planned Absence/Vacation Days:
No major vacations planned during GSoC.
Minor college tasks may occur but will be managed without affecting deliverables.
Open to using flexibility options if needed.
11) Skill Set:
Languages: Python, Markdown, SQL
Libraries: PyTorch, NumPy, Pandas, Scikit-learn
Beginner-level tools: NiBabel, Nilearn, Matplotlib, Seaborn
Projects:
AutoML system using TPOT
Web-based AI chatbot for mental health(currently working)
Currently Learning: Diffusion models, fMRI brain data, and neuro-AI concepts

Thank you.

[amby005]

1. Project Title:
   Advancing Brain Decoding and Cognitive Analysis: Leveraging Diffusion Models for Spatiotemporal Pattern Recognition in fMRI Data

2. Abstract / Project Summary:
   This project aims to model spatiotemporal brain activity using lightweight generative diffusion models tailored for fMRI data. We propose a novel hybrid pipeline that incorporates fairness-aware conditioning (age/gender/task), interpretable counterfactual analysis using SHAP, and compact U-Net + Transformer architectures optimized for limited compute. By aligning BOLD signal sequences with temporal graph structures and conditioning on cognitive states, we will improve decoding fidelity while ensuring transparency and generalization. The final deliverable will include a reproducible, open-source pipeline with low-memory training capabilities suitable for real-world neuro-AI applications.

3. Contributor Name:
   Amber Qayum Hawabaz

4. Contributor Email
   Email: amberhawabaz54@gmail.com

5. Potential Mentor(s):
   Babak Mahmoudi, PhDOzgur Kara

6. Personal Background (Brief CV):
   I hold a Master’s degree in Artificial Intelligence and currently work full-time as an independent AI researcher. My experience spans deep learning, medical imaging, and bias mitigation techniques. I have contributed to research project involving fairness-aware neural networks, causal learning, and generative models for low-resource settings. I have published at MLGenX (ICLR Tiny Papers), and am actively working toward a NeurIPS submission. I am passionate about neuroscience applications of AI, especially explainable models that promote equity and interpretability in healthcare and cognition.

7. Project Goals / Major Contributions:

Build a reproducible and lightweight fMRI preprocessing pipeline using Nilearn & NiBabel.

Convert voxel-level BOLD sequences into structured time-series and brain graphs.

Implement a compact U-Net + ConvLSTM-based DDPM model with limited timesteps and mixed-precision training.

Integrate fairness-aware conditioning (e.g., age, gender, task) into generative modeling.

Visualize counterfactual reconstructions with SHAP to improve interpretability.

Add GNN module to model dynamic brain region interactions.

Evaluate decoding and generalization using ABIDE and BOLD5000 subsets.

Provide comprehensive documentation, training notebook, and annotated codebase.

8. Project Schedule:

8.1) Community Bonding Period (May 20 – June 16)

Review recent work on fMRI, DDPMs, and graph-based neural decoding.

Identify datasets (e.g., ABIDE, OpenNeuro, BOLD5000-mini) and secure access.

Set up Colab/Google Cloud + lightweight PyTorch-based training environment.

Finalize architecture design and experimental roadmap with mentors.

8.2) Development Phase

Week 1 (June 17 – 23):

Preprocess fMRI data: align, denoise, normalize.

Convert to voxel-time tensor and brain graph representations.

Week 2 (June 24 – 30):

Build U-Net + ConvLSTM encoder-decoder.

Implement DDPM training loop with 100–200 timesteps.

Week 3 (July 1 – 7):

Add conditioning layers (age/gender/task embeddings).

Evaluate fairness-aware reconstructions.

Week 4 (July 8 – 14):

Integrate GNN to model dynamic brain region graphs.

Combine graph features with voxel encoder.

Week 5 (July 15 – 21):

Apply SHAP to evaluate importance of conditioning variables.

Generate counterfactual reconstructions (e.g., “older subject” scenario).

Week 6 (July 22 – 28):

Finalize model tuning, run ablations on fairness and graph inputs.

Visualize spatiotemporal embeddings and prediction heatmaps.

Week 7 (July 29 – August 4):

Run full evaluation on unseen ABIDE/BOLD5000 subsets.

Apply runtime optimizations for model deployment.

Week 8 (August 5 – 11):

Build visualization dashboard for user interaction with model outputs.

Integrate results into reproducible scripts and versioned artifacts.

Week 9 (August 12 – 18):

Document training and evaluation setup.

Record model walkthrough and inference demos.

Week 10 (August 19 – 25):

Final testing and cleanup.

Push full repo and publish usage guide.

Submit final report and project video.

8.3) Project Completion, Testing, and Documentation (August 26 – September 1)

Final review and polish.

Mentor feedback incorporation.

Official GSoC submission.

9. Planned GSoC Work Hours:

Type: Large-size project (35 hours/week)

Timezone: IST (UTC+5:30)

Preferred Hours: 10:00 AM – 2:00 PM IST (weekdays), flexible based on mentor feedback

10. Planned Absence / Vacation Days:
    None planned. Will balance part-time GSoC hours with full-time job and research deadlines.

11. Skill Set:

Programming: Python
Libraries: PyTorch, NumPy, SciPy, Scikit-learn, Matplotlib

Neuroimaging: Nilearn, NiBabel, BIDS format, basic FSL

ML/DL: CNNs, LSTMs, DDPMs, SHAP, Transformers, Graph Neural Networks

[ArchanaaS71]

Hi! I’m Archanaa, a final-year B.Tech student specializing in Artificial Intelligence and Machine Learning. I’ve spent the last few years exploring the world of machine learning, deep learning, and computer vision through hands-on projects and research. Some of my most exciting work includes building sentiment analysis systems using BERT,Segmentation models, detecting deepfakes with GANs and Vision Transformers, and developing real-time disease detection tools for agriculture. I’ve also worked on optimizing AI models for deployment using OpenVINO and have published research papers in reputed conferences and journals.

My technical toolbox includes Python, TensorFlow, PyTorch, Scikit-learn, and OpenCV, and I’m confident working with tools like GitHub, Google Colab, and Jupyter Notebooks. I’m particularly passionate about applying AI to solve real-world problems, especially those that impact people and the environment. My experience comes not only from coursework but also from internships, collaborative research, and self-driven learning.

What excites me most about GSoC is the opportunity to work on meaningful open-source projects, learn from experienced mentors, and contribute to a larger community. I believe my curiosity, dedication, and ability to quickly pick up new technologies will help me be a valuable contributor this summer. I see this as more than just a learning opportunity — it’s a chance to be part of something bigger and make a lasting impact.

📧 Email: 17archanaas@gmail.com
🔗 LinkedIn: https://www.linkedin.com/in/archanaa-s-9b7156299
🐙 GitHub: https://github.com/ArchanaaS71

[Priyanka12joshi]

Hi, I’m Priyanka Joshi, a B.Tech CSE (AI & ML) student at Sir Padampat Singhania University. I’m passionate about working at the intersection of AI, neuroscience, and healthcare. I’ve built an AI-powered medical assistant model integrating RAG and LLMs, and I’m deeply interested in brain encoding/decoding using datasets like BOLD5000 and NSD. I love exploring how large language models and visual-language systems can be applied in real-world medical and cognitive science domains. GSoC 2025 feels like the perfect platform to contribute meaningfully to open-source research, learn from the community, and grow as a researcher.

E-mail: priyankajoshi2300@gmail.com
Thank you!

[swatwas]

Hello everybody. I am Swaytha, a Computer Science Sophomore studying at Nanyang Technological University. I have had hands-on experience in deep learning projects in the past. Recently, I completed an internship where I utilized supervised and unsupervised learning techniques to detect and analyze specific operational conditions of home appliances. I am also currently interning at a lab where we are investigating vision language models for medical diagnosis. I have been exploring generative architectures and I am interested in this project as I am keen on applying diffusion models for reconstructing visual stimuli from brain signals.

Email: vswaytha4@gmail.com

Project Goals:

1. Understand and preprocess fMRI data using BIDS, NiBabel, and Nilearn
2. Extract voxel-level spatiotemporal features from BOLD sequences
3. Implement baseline spatiotemporal encoders
4. Design and train a lightweight diffusion model tailored for fMRI decoding
5. Build a temporal GNN module to learn dynamic brain region interactions
6. Integrate SHAP-based interpretability for counterfactual reconstructions
7. Evaluate model performance on datasets like BOLD5000 and ABIDE

Project Timeline:
Week 1: Load, preprocess, and normalize fMRI data (BOLD sequences)
Week 2: Build baseline CNN/LSTM for feature extraction
Week 3: Train/test toy diffusion model on simple voxel encodings
Week 4: Design U-Net + Transformer/ConvLSTM-based DDPM for fMRI
Week 5: Integrate conditional embeddings (age/gender/task type)
Week 6: Add temporal GNN for graph-based brain connectivity modeling
Week 7: Implement SHAP explainability and generate counterfactuals
Week 8: Run ablations, evaluate model generalizability on new subjects

Planned GSoC Hours:
Size: Large
Timezone: GMT 8+
Preferred Hours: 8am to 9pm (GMT 8+, flexible if needed)
No major planned vacations.

Skill Set:
Languages: Python, Java, C++, SQL, JavaScript
Libraries: PyTorch, NumPy, Pandas, Scikit-learn, Matplotlib, Seaborn
Projects:

1. Building Vision Transformers for Medical Diagnosis
2. Evaluated Sparse AutoEncoders for Chain-Of-Thought Prompting (Interpretability of LLMs)

I am currently learning about diffusion models and graph neural networks.

[iftiquar]

Hi everyone,
Decoding Brain Activity: Diffusion Models for Spatiotemporal fMRI Pattern Recognition." This project aims to harness diffusion models to analyze fMRI data, decoding cognitive states and identifying biomarkers for neurological disorders. By developing a preprocessing pipeline, designing a tailored diffusion model, and evaluating its performance in brain decoding tasks, I hope to contribute innovative tools to Emory University’s Department of Biomedical Informatics.
Best regards, Iftiquar Ali

[tutuponnekanty]

Hi, I’m P. Y. Rajkamal Tutu, an M.Tech Artificial Intelligence student at NIT Silchar, also pursuing a B.S. in Data Science and Applications from IIT Madras in parallel. I’m passionate about brain-computer interfaces, cognitive modeling, and the intersection of AI, neuroscience, and generative modeling. I’ve previously worked on explainable AI for Indic-language spam classification, and few other kaggle projects.

I enjoy working with LLMs, visual-language models, and models that combine structure and reasoning in biological contexts. I believe GSoC 2025 is an ideal opportunity for me to contribute to high-impact open-source research, collaborate with domain experts, and grow technically and intellectually through mentorship and community engagement.

E-mail: tutuponnekanty@gmail.com
GitHub: https://github.com/tutuponnekanty
LinkedIN : https://linkedin.com/in/pyrkt007

Thank you!