GSoC 2026 - Discussion

[pollytur]

Understanding the brain requires aligning high-dimensional stimuli with large-scale neuronal recordings. Experanto is a Python framework designed to bridge this gap, providing a unified interface for interpolating and aligning time series (calcium imaging, continuous behavior such as running speed), events (spiking data, action-based behavior), and media (images, video). Optimized for deep learning, Experanto recently powered OmniMouse, a large-scale multimodal neural predictive model.

The core strength of Experanto is its flexibility: it allows users to resample frequencies, filter data by behavioral criteria (e.g., "only when running"), and integrate multiple experiments into a single pipeline.

The goal of expreanto for GSoC 2026 is to expand Experanto’s functionality, visualization tools, and robustness.
Proposed focus areas include:

• Open-Source Data Integration: Currently supporting the Allen Brain Calcium Imaging dataset, we aim to expand to the Brain Wide Map, Dandi [Cebra Example with Nandi dataset] and Allen Neuropixels datasets. Ideally, this involves creating a "brainsets-style" library to automatically download and map these datasets to Experanto. (Example with calcium dataset from Allen Brain)
• Extend experanto functionality:
  • Enhanced Filtering & Parsing: Developing a string-based configuration interface for complex data filtering—for example: (treadmill > 0 and eye is not nan) & screen == video.
  • Event-Driven Data Handling: Explicitly adding support for event-based data types to better process free-animal behavior experiments.
  • Visualization & Documentation: Building built-in visualization tools and expanding tutorials for data analysis and neural predictive modeling.

We are also happy to discuss with the community what are the other important features that we should implement.

For contributions we inspire you to use uv (link) locally (more consistent than conda, more simple than docker). We will try to add a uv.lock file to the repo soon.
While experanto has some GPU-related code in it fordata loading - don't worry about it, you don't need a GPU.

For questions please ping @reneburghardt, @schewskone, @pollytur.

Thanks a lot for your interest in experanto ❤️ and let's build it together 🚀

[pollytur]

@VaibhavGarg2003 Thanks a lot for making us to start the discussion page and we are happy to answer your questions if any.

[VaibhavGarg2003]

Hi @pollytur , thank you for setting this up and for the quick review on PR #100

Understood completely on the .toml file vs. README ,it makes total sense that dependencies should resolve automatically without an extra requirements.txt step. I also saw your note here about transitioning to uv for local development. I'll get uv set up on my end today to match your preferred workflow.

Regarding the project scopes, I am highly interested in the 'Extend experanto functionality' bullet point. Given my background in Python and building data parsing/RAG pipelines, I'd love to tackle the string based configuration interface for complex data filtering.

To get a better grasp of the current architecture before drafting a proposal, which specific modules (e.g., inside experanto/datasets or interpolators) currently handle the baseline filtering logic? I'd like to study how data is currently parsed before proposing the string-based upgrades.

[pollytur]

Thanks a lot for your interest in experanto!

I guess to understand the structure of the data the best entry point is to replicate this export of the publicly available data - https://github.com/sensorium-competition/allen-exporter/blob/main/examples/example_experiment_export.ipynb

And then load it to experanto
https://github.com/sensorium-competition/allen-exporter/blob/aff4cdcad2f4707c165082d617368fae1ffcd65a/examples/example_interpolation.ipynb#L29

You should be able to reproduce it using google colab or sth like it.

And of course look into the guts of the functions that are used in the second notebook

Current filtering logic is addressed in the PR #88 but this still needs a bit of work and we would try to finalize it within next 2 weeks.

PS We are releasing more data soon as well 🎉

[pollytur]

Dear GSoC participants, thanks a lot for your interest in experanto and your enthusiasm to contribute.

Before submitting a PR, please open an issue to discuss your proposed change. PRs that merely reorganize or copy existing code without clear motivation, or that don't align with the project's direction, will be closed without review. Maintainer time is limited — please use it respectfully

[Vrittigyl]

Hi! I've been going through the codebase and the GSoC ideas, and I'm interested in working on the Open-Source Data Integration area for my proposal.
A couple of questions before I start implementing:

• Dataset priority
  I noticed that PR Added DANDI dataset example notebook (issue #107) #111 adds a DANDI example notebook (for issue Dandi Dataset example #107). If the goal is to move toward a deeper “brainsets-style” integration where datasets can be automatically downloaded and mapped to Experanto’s format, which dataset should be prioritized first: Allen Neuropixels, Brain Wide Map, or DANDI?
• Integration architecture
  For implementing these integrations, would the preferred approach be to build a separate library (similar to how allen-exporter is structured), or should the dataset adapters/loaders live directly inside the Experanto repository?

Happy to start with whichever direction is most useful. Thanks!

[pollytur]

Hi @Vrittigyl!
Sorry for long reply - please ping us next time for faster replies :)

• As for your questions - DANDI is not a single dataset but an online-arxiv for neuroscience dataset. The datasets there are typically small. Added DANDI dataset example notebook (issue #107) #111 addressed how to match DANDI data to interpolators, one thing which is still missing is to extend that example to show how to actually make an Experiment object from DANDI data (this would require making tiers as a separate modality and puting them into TimeIntervalInterpolator) - this is actually a quite important example and addressing it is very welcome. Another thing with DANDI is a bit more vague - it would require going through its dataset and seeing if there are modalities which we don't cover (issue Make the codebase less dependent on the screen #66 might also help for it).
• For Allen Neuropixels, Brain Wide Map - its probable best to do PRs directly into allen_repository (though it might require restructuring a bit later to handle many datasets import)
  I would suggest starting with a DANDI and tiers example as its needed for Allen Neuropixels, Brain Wide Map anyways

Let us know if you have more questions

PS we will try to get to your PR this week - thanks a lot for actively contributing!

[Vrittigyl]

Hey @pollytur, Thank you so much for the detailed explanation along with references.
While working on my proposal, I had understood that DANDI is a collection of NWB-based datasets, and I had been thinking of a general nwb_to_experanto()-style pipeline for such data.
I'll start working on to first extend the current DANDI example by constructing a proper Experiment object (by handling tiers for TimeIntervalInterpolator). I’ll begin with DANDI + tiers example, and then see how it can be extended into a more general pipeline. Will also keep Allen Neuropixels and Brain Wide Map in mind while doing this.

Also, thanks for mentioning that I can ping you if I get stuck, that’s really helpful! I’ll begin by looking into issue #66 and will reach out if I have any questions.

Thanks again for the guidance!

[abdelrahman725]

Thanks @pollytur for always helping !

I will share some updates here:

• allen dataset: Create colab example with Allen demo dataset. #139 is ready for review
• A reminder to review the issue Add missing type hints and make them consistent #131

I'm sorry if I diverged from the priorities of experanto and GSoC by opening these issues. I only discovered them by chance and thought they were worth it, especially the missing type hints issue as it affects development experience.

Anyway, from now on, I will put all my effort into Brain Wide Map especially that SpikeInterpolator is now merged.

Some Questions:

• I'm still not sure if it's still worth it to ask to be assigned to a new issue?
• I don't really think it's a good idea to do PRs directly into allen-exporter (can complicate things and make dependencies conflict), I believe either doing PRs into experanto or creating a new generic repo for integrating the different datasets
• What would be a good example of a small deliverable (i.e., something you want to see) for integrating BWM ?

[abdelrahman725]

Thanks @pollytur !

I've actually started exploring and studying BWM. A script that downloads BWM seems like a really good start, I will be working on that and I will keep you updated !

[abdelrahman725]

@pollytur
I've created an experanto-exporter repo with some progress on BWM, no export logic just yet, but the linked notebook has some progress and is worth looking at to verify that I'm moving in the right direction; once I verify that I will proceed immediately and hopefully reach the minimal deliverable before May

[pollytur]

Hi @abdelrahman725 I have looked at the notebook, great that you got Open Neurophysiology Environment API working. Looks good for me for now. Maybe rephrasing the question would help me to understand what type of feedback you want

[abdelrahman725]

Hi @pollytur, The linked BWM notebook has:

1. A data flow of searching sessions, loading sessions, loading specific data, and filtering spike data.
2. Written documentation representing my own thoughts and assumptions.

I want feedback on those 2 points to make sure there are no misassumptions and that I'm moving effecitvely toward the end goal of that repo.

[pollytur]

@abdelrahman725 Some comments based on the notebook

• Yes, BMW is big. What we want to support in export if probably this a) export single session by id b ) export all sessions - same way as its done in allen export, where we loop through all ids and export them all c) and example of filtering, probably sth like export all sessions for laboratory = X would be good enough
• for Available search terms - all of this should be saved at the top-level meta-data yaml. If sth is missing - set it to None
• Yes, project=brainwide
• Things like probes.description should go to the meta of the corresponding modality
• Spikes clusters should probably be a .npy referenced in meta (this comes from spike sorting preprocessing)
• Good and mua should go to metadata as well, I would not filter based on it during the export stage

Lmk if this makes sense / any other questions left

[ojasvatsyayan]

Hello @pollytur!

I’ve been going through Experanto’s codebase and the GSoC ideas, and I’m particularly interested in the event-driven data handling direction. As of right now, I've raised one issue on first review, but I'm super excited to dive deeper into the codebase.

From exploring the interpolators, my current understanding is:

• SequenceInterpolator handles continuous signals over time
• TimeIntervalInterpolator works with valid time intervals
• SpikeInterpolator handles discrete spike events

I’m trying to better understand how spike/event-based data is intended to fit into the overall pipeline.

In particular:
• When querying specific times, should spike data be treated fundamentally differently (like event lookup rather than interpolation), or is there an intended way to align it more closely with the other modalities?

• Which parts of the codebase would you recommend focusing on to understand and extend event-based handling (would focusing on specific interpolators vs. higher-level Experiment integration be better for understanding)?

• Are there any known gaps or limitations in how event-based data is currently handled that would be useful to explore as a starting contribution?

I’d like to align with the intended design before starting implementation work.

Thanks!

[pollytur]

Hi @ojasvatsyayan

Thanks for your interest to experanto.

As for your questions

"When querying specific times, should spike data be treated fundamentally differently (like event lookup rather than interpolation), or is there an intended way to align it more closely with the other modalities?"

Yes, spikes are irregular time series, so they cannot be interpolated same way as regular time series. That's why SpikeInterpolator and SequenceInterpolator interpolation work differently.

"Which parts of the codebase would you recommend focusing on to understand and extend event-based handling (would focusing on specific interpolators vs. higher-level Experiment integration be better for understanding)?"

SpikeInterpolator and TimeIntervalInterpolator the closest to the even-based interpolator we currently have. Spikes is basically the binning mechanism for events, TimeIntervalInterpolator could be modified to do sth like take the event of "saccade start" and take the next 0.5sec after each saccade

"Are there any known gaps or limitations in how event-based data is currently handled that would be useful to explore as a starting contribution?"

I guess adopting experanto for the example above would be the easiest known gap. If you decide to work on it - please dont do abstract coding and find a dataset on DANDI with saccades and test things on it as well - attaching it as the .ipynb demo to the PR

Let me know if this helps

[ojasvatsyayan]

Thanks @pollytur , this is super helpful! The event-aligned window idea makes a lot of sense, and I’ll explore extending TimeIntervalInterpolator to support extracting fixed windows relative to events. I’ll look into relevant datasets on DANDI and try to build a concrete example with a notebook demo. I'll update here when I have a PR ready.

[ojasvatsyayan]

Thanks @pollytur, this was super helpful, especially the suggestion to ground the work in a real DANDI dataset and build a concrete notebook workflow.

I ended up structuring my GSoC proposal around that exact idea. I’ve now submitted my proposal, and really appreciate the guidance- it helped shape both the design and the focus on real-world usability.

[BitForge95]

Hi @reneburghardt, @schewskone, and @pollytur, since this thread is primarily for brainstorming ideas, I wanted to ask what the preferred protocol for reviewing early student drafts?

[reneburghardt]

Sorry @BitForge95 unfortunately we won't have the resources to review drafts

[binary69]

Hey @pollytur
interpolator.py
`idx_lower = np.floor(
(
valid_times[:, np.newaxis]
- self._phase_shifts[np.newaxis, :]
- self.start_time
)
/ self.time_delta
).astype(int)

times_lower = idx_lower * self.time_delta
times_upper = idx_upper * self.time_delta` do we have to add start_time here in SequenceInterpolator/interpolate and PhaseShiftedSequenceInterpolator/interpolate

[pollytur]

@binary69
I am not sure what was the goal of the comment, it seems like sth got broken

[binary69]

sorry for the confusion :(
I will open it as an issue I was trying to verify on why self.start_time is not added in times_lower, if strat_time !=0 will it be a little innacurate on lower_signal_ratio = ((times_upper - times_valid) / denom)[:, None] numerator (SequenceInterpolator/interpolate)