Research Scientist - Deep Learning Models

The Mindspan Institute is a new non-profit research institute in Cambridge, MA, aimed at building novel tools to analyze the human brain, and scaling them to study its molecules and wiring. We interpret how they contribute to brain functions and dysfunctions, and share our tools and discoveries as freely as possible, to advance human health.

The role

This is a hands-on modeling role. You will design, train, and evaluate the deep learning models that turn large microscopy volumes into usable data: segmentation, the stitching and alignment of image tiles and sections, and the detection and classification models that follow. The job will heavily depend on expertise and fluency in DL frameworks like PyTorch: reading the literature, choosing an architecture that fits the data, training it at scale, and finding out where and why it fails.

You report to the Head of Computation, and work, as needed, with other Mindspan teams (Microscopy and Optics, Expansion Microscopy, Scalable Chemistry, and Automation and Industrialization), using logic and evidence to advance your work so as to boost the performance of all.

The position is intended for a generalist: someone who is as comfortable adapting a pretrained vision model as writing a new loss function, and who picks the model to fit the problem rather than the reverse. Given the multidisciplinary nature of the job, it is ideal for a person who enjoys problem solving across intellectual boundaries, at the level of concepts as well as of hands-on debugging. It also requires systematic documentation, analysis, and iterative problem solving, to arrive at working models, apply them, and share them.

What you'll own

• Design, train, and evaluate deep learning models for volumetric microscopy data, including dense segmentation, boundary and affinity prediction, and object detection.
• Build models and methods that stitch and align image tiles and serial sections into consistent volumes.
• Adapt pretrained and foundation vision models to our imaging modalities, and decide when a new architecture is worth the cost.
• Run controlled experiments: ablations, held-out evaluation, and error analysis that shows which failure modes a change fixed and which it introduced.
• Work with the pipeline and annotation teams to move models into production, and to request the ground truth that the next model needs.
• Documentation and training: keep experiment records that others can reproduce, with the reasoning behind each design choice

Culture we're building

The disorders of the brain affect over 1,000,000,000 people around the world - Alzheimer’s, Parkinson’s, stroke - and other conditions - none of which can be cured. We operate with a sense of urgency, even as we look to understand the brain at its most fundamental levels.

We are hypercollaborative, and focused on impact over short-term markers of productivity. Creativity and failure are the stepping stones to scale, and working together with “strong opinions, weakly held” is our style.

We solve real engineering problems the best way we can: inventing fast when we have to, and using good existing solutions where they exist. We believe hard problems only yield to many disciplines working together, so collaboration is required, not optional.

We expect people to explore the unknown, which means making mistakes daily, owning them, and learning from them to improve and excel.

We are recruiting a team of scientists and engineers who exemplify all these values.

What we're looking for

• Advanced degree or equivalent industry experience in computer science, electrical engineering, applied mathematics, physics, computational biology or neuroscience, or a related field.
• Direct hands-on experience designing and training deep learning models. 
• Demonstrated experience with image segmentation or dense prediction on 3D or large 2D image data.
• Strong software engineering: Python, version control, testing, and code that other people can run and extend.
• Experience training on multi-GPU systems or a compute cluster.
• Excellent engineering judgment and scientific rigor, including the discipline to tell a real improvement from noise. Experience with image registration, stitching, or alignment of large tiled or serial-section datasets.
• Experience with connectomics or other biological microscopy data (electron, light, or expansion microscopy).
• Experience with self-supervised learning, or with adapting large pretrained vision models to a new domain.
• Experience with chunked array formats (Zarr, N5, precomputed) and data too large for one machine.
• Publications, open-source models, or results on public benchmarks

What we offer

You will get to contribute to addressing one of the most important unmet medical needs of our time, as part of the initial team. As Mindspan grows, we expect this role to grow with it. We are excited to invest in someone who wants to take on increasing responsibility over time.

Comprehensive health, dental, vision, and family benefits, and a matching retirement contribution.

Competitive salary ($180,000-260,000), depending on experience.

The Mindspan Institute is an Equal Opportunity Employer. We are committed to a collaborative and scientifically rigorous environment for everyone who works here.