Student projects in the EPFL NeuroAI Lab should:
- contribute to the mission of the lab, i.e. to create accurate models of behavior and underlying neural mechanisms — these models let us make sense of the mind and brain, and enable new applications in the diagnosis and treatment of neural disorders
- contribute to ongoing projects by permanent members (PhDs and Postdocs), so that the results of the project will be useful long-term, and so that the student can make use of existing expertise.
- be aligned to the interests of the student with a realistic timeline so that meaningful progress with some closure can be made over the duration of the project.
The list of projects here is non-comprehensive. It is always a good idea to reach out to lab members whose research you find interesting and inquire if you can contribute to their projects. In either case, we ask that you apply via the form on the lab website. During interviews, we will share a more comprehensive list of projects with you that we are thinking about.
Modeling Brain Response to Dynamic Visual Stimuli Using Enhanced Neural Networks for Motion Processing
Project Overview:
Our lab focuses on building dynamic vision models of the brain. We have collected 8 fMRI datasets using natural stimuli and tested over 70 models for brain alignment. By correlating model brain alignment with various task performances, we identified motion processing—recognizing high-level actions from pure motion stimuli (like moving dots)—as a key missing task, alongside object recognition.
To improve our brain model, we aim to enhance the motion processing capabilities of existing neural networks. Key research questions include:
* How much can enhancing motion processing improve on brain alignment?
* Does a dual-stream separation (analogous to the brain’s ventral-dorsal streams) improve topological brain similarity? [1]
* How can we train a model on natural video data to perform pure motion recognition, and how does this affect brain alignment? [2, 3]
Project Requirements:
* Experience training deep neural networks on image/video data.
* Familiarity with state-of-the-art (SOTA) neural network architectures for image/video tasks.
* Strong knowledge of data processing in computer vision, such as optical flow computation and frequency band-pass filtering.
Expected Outcome: A top-tier conference paper.
References:
Bakhtiari, Shahab, et al. “The functional specialization of visual cortex emerges from training parallel pathways with self-supervised predictive learning.” Advances in Neural Information Processing Systems 34 (2021): 25164-25178.
Han, Shuangpeng, Ziyu Wang, and Mengmi Zhang. “Flow Snapshot Neurons in Action: Deep Neural Networks Generalize to Biological Motion Perception.” arXiv preprint arXiv:2405.16493 (2024).
Sun, Zitang, et al. “Modeling human visual motion processing with trainable motion energy sensing and a self-attention network.” Advances in Neural Information Processing Systems 36 (2024).