A college sophomore is using machine learning to detect Alzheimer’s disease from brain scans in ways that conventional statistics simply cannot — and that single fact tells you everything about where serious ML research is happening right now. It is not just at Apple’s booth at ICML 2026 in Seoul. It is happening in campus labs, funded by fellowships, run by undergraduates who have not yet graduated but are already pushing past the limits of clinical methodology.
According to Bowdoin College, 180 students received research fellowships this summer, with 456 students total securing funding for internships and fellowships across disciplines. One of those projects is applying machine learning to Alzheimer’s prognosis in a way that ought to make certain well-funded research labs look over their shoulders.
The facts:
- Bowdoin undergraduate student Enbar-Salo is using brain imaging data and ML to study how connected regions of the brain can reveal diagnostic signs of Alzheimer’s that traditional biomarkers miss.
- His project is titled Temporal Brain Connectivity and Alzheimer’s Prognosis: A Machine Learning Approach and is funded through the Surdna Foundation Undergraduate Research Fellowship.
- The work is advised by Assistant Professor of Computer Science Jeová Farias.
- Enbar-Salo previously worked in the Vannini lab at Massachusetts General Hospital, where he hit the ceiling of traditional statistical methods.
- Apple presented research at ICML 2026 in Seoul, South Korea, from July 6–11, 2026, covering topics from video tokenization to memory architecture in large language models.
Brain Networks as Social Networks
The conceptual move at the center of Enbar-Salo’s research is genuinely clever. He is not just feeding MRI data into a model and hoping for a pattern. He is borrowing from social network analysis — the same mathematical frameworks used to study how information flows through human communities — and applying them to brain connectivity. Certain brain regions act as hubs. The question is whether the strength or fragility of those hubs carries a diagnostic signal that shows up before other Alzheimer’s markers do.
Machine learning is suited to this problem precisely because the relationships between brain regions are non-linear, high-dimensional, and deeply contextual. Traditional regression models struggle with that. They were not built for it. The MGH work Enbar-Salo did earlier gave him a front-row seat to that limitation, which is exactly why he pivoted toward computational methods at Bowdoin.
This is what real ML research looks like in 2026 — not the press release version. It is messy, specific, and rooted in a concrete clinical gap rather than a pitch deck.
Big Tech Research vs. The Undergrad in the Lab
Apple showed up at ICML 2026 with a formidable portfolio. VideoFlexTok, MemoryLLM, speculative expert prefetching — the company is investing serious engineering hours in making large models more efficient and more capable. That is genuinely important work. But there is a version of the ML conversation that gets completely captured by the corporate conference circuit, where the only research that registers is the research with a press office behind it.
The Bowdoin fellowships are a useful corrective. They represent a category of ML application that does not get enough oxygen: small-scale, hypothesis-driven, domain-specific research where the goal is not to ship a product but to answer a question that medicine has not been able to answer yet. That is a different kind of ambition, and frankly, it is a more defensible one. We should be more excited about undergraduate researchers finding new Alzheimer’s biomarkers than we are about a video tokenizer being presented in Seoul.
The data center cost angle inside the Bowdoin fellowship lineup is worth sitting with too. Someone at that school is studying what AI infrastructure actually costs — not in marketing terms, but in economic and environmental terms. As we’ve reported elsewhere on questions of cloud infrastructure and its real-world tradeoffs, the price tag of AI at scale is a story the industry prefers to obscure.
The Broader Pattern Is Not Accidental
Machine learning research in 2026 is bifurcating. On one track, you have the industrial research apparatus — Apple, Google, NVIDIA — producing papers at ICML and NeurIPS that optimize systems already in deployment. On the other track, you have domain specialists using ML as a tool to crack open problems in medicine, economics, and history that have resisted conventional analysis for decades.
The two tracks are not in competition. But they are not equally covered, and they are not equally funded. A fellowship at Bowdoin is a rounding error compared to Apple’s ICML sponsorship budget. The attention economy around ML research reflects that disparity in ways that distort what people think the field is actually doing.
The internet has a long memory for platform pivots and rebranding moments — just look at how we’ve tracked the collapse and reinvention of Twitter into X — but it has a shorter memory for incremental scientific work that does not come packaged as a launch event.
That is the real tension here. Enbar-Salo’s Alzheimer’s research might not produce a deployable product for years, if ever. It might produce a finding that reshapes how neurologists think about early-stage diagnosis. That outcome is harder to quote in a headline. It is also more important. The fellowship exists precisely to protect that slower, harder kind of inquiry — and the fact that it is happening at the undergraduate level, right now, in 2026, should matter more to this industry than it currently does.
Watch the Breakdown
Sources
- Summer Fellowships Include the History of History, a Machine Learning Approach to Alzheimer’s Research, and the Cost of AI Data Centers — www.bowdoin.edu
- International Conference on Machine Learning (ICML) 2026 — machinelearning.apple.com
- How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo — developer.nvidia.com
