OHBM 2026: Neuroinformatics and Data Sharing + AI and the Analysis of brain structure and function — does AI learn science, or does it learn bias?

OHBM 2026 brings a very important theme for anyone who wants to think about the future of neuroscience without naivety: on one side, AI and the Analysis of brain structure and function; on the other, Neuroinformatics and Data Sharing. Within these tracks, topics include A Novel Foundation Model for Structural Brain Health Analysis, Federated Label-Noise Filtering of Neuroimaging Data, BIDS-Flux: High-throughput, Federated Neuroimaging Data Management Platform for Multisite Studies, and ClinicaDL v2: an open-source Python library for reproducible deep learning in neuroimaging. That combination alone already signals an important shift: the discussion is no longer only about having more data or stronger models. OHBM 2026 itself makes it clear that infrastructure, curation, data sharing, label noise, reproducibility, and machine learning sit at the heart of the scientific problem.

This has major value for a Decolonial Neuroscience reading. For a long time, part of science treated datasets as if they were neutral mirrors of reality. But data never arrive pure. Every dataset is shaped from the beginning by who was included, who was left out, how things were measured, in which language, in what context, with which technology, under what definition of normality, and with what prior idea of a healthy brain. When OHBM 2026 places Foundation Model, federated data, label-noise filtering, and open deep-learning libraries side by side, it signals an important maturity: method is also politics.

In Brain Bee language, the question can become:

If data come from biased populations, will AI discover the human brain, or will it only repeat the already accepted standard brain?

This is a strong question because it goes straight to the point. Teenagers understand this quickly. If you train a machine on limited examples, it learns one piece of the world and begins to treat that piece as if it were the whole. In neuroscience, this becomes even more delicate. We are not talking only about recognizing images; we are talking about health, development, attention, emotion, aging, and human difference.

Here, the avatars that help most are Math/Hep and Olmeca.

Math/Hep matters because it watches over method. It reminds us that a strong model is not the same thing as strong truth. A Foundation Model may look impressive, but the central question remains: what was it trained on, what noise did it inherit, which population does it represent, and what kind of generalization does it truly support? The same applies to noise filters, reproducible pipelines, and federated systems: all of them can improve science, but only if the methodological question remains alive.

Olmeca matters because data are also culture, history, and territory. When a system learns from samples concentrated in certain countries, certain languages, certain social profiles, and certain health patterns, it does not learn “the human brain” in general. It learns more strongly the brain that already had the best chance of becoming data. And this matters deeply for Latin America. If we do not ask this question, we risk importing models that look universal but were already born limited by earlier exclusions.

The decolonial critique here does not need to be aggressive. It can be simple: AI does not automatically overcome bias already present in the data. In many cases, it organizes that bias better, accelerates that bias, and returns that bias with an appearance of neutrality. That is why the OHBM 2026 program is so interesting at this point. When themes such as Federated Label-Noise Filtering of Neuroimaging Data, BIDS-Flux, and ClinicaDL v2 appear, the congress is recognizing that data quality, responsible sharing, and reproducibility are not minor technical details; they are part of scientific validity itself.

A better question, then, would be this:

What changes when context enters the model instead of remaining hidden behind the appearance of pure data?

That is a good question for OHBM 2026, a good question for Brain Bee, and a very important question for Latin America. Because here we need strong science, but also science that knows how to recognize who has historically been left out of datasets, protocols, and dominant definitions of normality.

A Brain Bee proposal for an EEG + NIRS experiment

The proposal can be simple: classify task-performance profiles using EEG + NIRS in two model versions. In the first version, only neural and behavioral variables are included. In the second, sociocultural and contextual variables are also added, such as language, type of school, sleep routine, or stress context.

The point would not be to prove that “context explains everything,” but to show how the model changes when context is no longer invisible. The central hypothesis is direct: part of what AI calls a brain pattern may actually be mixed with differences in environment, language, care, or opportunity.

Where OHBM 2026 is already pointing in this direction

This blog grows directly out of the official program. The track AI and the Analysis of brain structure and function includes topics such as A Novel Foundation Model for Structural Brain Health Analysis and Federated Label-Noise Filtering of Neuroimaging Data. The track Neuroinformatics and Data Sharing includes BIDS-Flux and ClinicaDL v2, along with other tools focused on integration, reproducibility, and federated data management. This helps shift the question.

Instead of asking only “which AI works better?”, the discussion becomes more mature: what kind of data support this model, what noise does it carry, which population does it represent, and what kind of science does it actually help build?

Why this matters for Latin America

In our region, thinking about neuroinformatics and AI also means thinking about scientific sovereignty. It is not enough to use ready-made models and expect them to represent us. We need to ask who entered the data, who was left out, in which language the protocols were designed, what kind of infrastructure is required to participate, and how open science can become truly open for countries with fewer resources.

This is especially important for young people between 14 and 17 years old. They already live with algorithms all the time. That is why they can learn early a decisive lesson: a machine does not become neutral just because it performs calculations. It learns from what it receives. And if it receives an unequal world, it may return that world with the appearance of technical truth.

The beauty of this OHBM 2026 theme is exactly this: it already opens space to move beyond shallow fascination with AI and enter a more serious discussion about data, method, and scientific responsibility.

Instead of asking only whether AI is more accurate, we can ask:

Which brain did it learn to recognize?
Which differences does it erase when it calls everything a pattern?
How can reproducibility and data sharing improve science without hiding bias?

When neuroscience begins to measure that clearly, it stops being only a science of automation and starts becoming also a science of methodological responsibility.

References used in this blog

• OHBM 2026 — track “AI and the Analysis of brain structure and function”, including A Novel Foundation Model for Structural Brain Health Analysis and Federated Label-Noise Filtering of Neuroimaging Data.

• OHBM 2026 — track “Neuroinformatics and Data Sharing”, including BIDS-Flux: High-throughput, Federated Neuroimaging Data Management Platform for Multisite Studies and ClinicaDL v2: an open-source Python library for reproducible deep learning in neuroimaging.

• OHBM 2026 Schedule at a Glance — confirmation that sessions related to AI, neuroinformatics, and data sharing are part of the official congress program.