anibalsolon
commented
4 years ago Hi @JacobCWBillings
The project seems exciting! Thank you for submitting it.
If you could please try to delineate the current issue and add one more, so for the newcomers. And put them on your project board 👌
Please, let us know if you need a hand on setting up the communication channels. @complexbrains added some labels too, take a look if they are enough to cover your project.
One more thing: let us know if there is an image for your project, so we can add to the website!
complexbrains
tashrifbillah
JacobCWBillings
Project info
Title: Vector Embeddings of Brain Dynamics
Project lead: Jacob Billings github.com/JacobCWBillings
[Timezone]: Montreal, UTC -4
Hub: The Americas
Description: Whereas we can build time-varying functional connectivity (TVFC) graphs easily and efficiently using instantaneous wavelet coherence. And whereas we can compare TVFC graphs using a host of graph-graph metrics---from simple edge-wise overlap, to more robust methods from topological data analysis. The capacity to understand how TVFC states compare among volunteers is limited by the transductive nature of state-of-the-art manifold embedding algorithms such as tSNE and UMAP. Transductive methods tend to overweight the temporal adjacencies of graph-graph distances measured within single volunteers, and thereby ignore the large-scale graph-graph similarities measured across volunteers. To address this issue, this project aims to build vector embeddings of several matrices of TVFC graph-graph distances. One class of inductive and vectorized embedding that may work well is the Siamese neural network. We should try to look at some other techniques too.
Link to project: https://github.com/users/JacobCWBillings/projects/1
Mattermost handle: channels/vectorize_dynamics @jacob.billings
Goals for the OHBM Brainhack Vectorize distances between time-varying functional connectivity graphs. Vectorize several graph-graph distances to compare the properties of each metric space.
Good first issues: 0) Project provides graph-graph distances using several metrics as numpy matrices. 1) Hack together a Siamese NN. Possibly use the code from git/UGraphEmb. 2) Test vector embedding distance against metadata scores: volunteer #, experimental condition, etc. (Some code for sample tests exist). 3) Hack and train other embeddings, transductive and/or inductive, and test statefulness. 4) Visualize brain states.
Skills: Python Neural networks Dimensionality reduction Scientific writing Data visualization Functional connectivity dynamics
Chat channel: https://mattermost.brainhack.org/brainhack/channels/vectorize_dynamics
Acknowledging contributions: Our goal is publication of a novel manuscript describing how the interpretation of spontaneous brain dynamics and brain states is affected by the type of embedding technique used. We'll generate and share our code for comparing graph-graph distances as a git repo.
Project submission
Submission checklist
Once the issue is submitted, please check items in this list as you add under 'Additional project info'
Please include the following above (all required):
You can also include information about (all optional):
We would like to think about how you will credit and onboard new members to your project. We recommend reading references from this section. If you'd like to share your thoughts with future project participants, you can include information about (recommended):