Research Scientist — Frontier Physical AI

Sigma Nova; Paris; Sur place; EUR 90 000 - 130 000; Plein temps

Résumé du poste

Sigma Nova in Paris is building foundation models for complex scientific signals, starting with EEG, fMRI, and ultrasound, with the aim to extend to other data-scarce domains. The role emphasizes architecture-level thinking, pre-training strategies, and multimodal alignment in a collaborative team.

As a key contributor, you will own experiments from hypothesis to deployment, contribute to engineering pipelines, and advance scientific understanding of physical signals while thriving in an open,

Qualifications

• PhD (preferred) in Machine Learning, Deep Learning, or a closely related quantitative field.
• Strong proficiency in PyTorch and experience running large models on GPU infrastructure.
• Hands-on experience handling challenging data, including irregular time series and scientific signals.
• Data-first mindset prioritising raw signals and data quality over loss curves.
• Ability to thrive in ambiguous environments and explain research to non-scientific teams.
• Fluent English.

Responsabilités

• Architecture-level thinking: address why standard methods fail on irregular, noisy, or scarce data.
• Develop pre-training strategies tailored to physical signals with small samples.
• Pursue multimodal alignment to bridge gaps between signals and inductive biases for physical systems.
• Take end-to-end ownership: design experiments from hypothesis to deployment and contribute to pipelines.

Connaissances

PyTorch

Large models

GPU infrastructure

Time-series data

Multimodal alignment

Formation

PhD (preferred) in ML/DL or related field

Description du poste

The Challenge: Complex Scientific Data & Representation Learning

We’re building foundation models for scientific signals, starting with the brain (EEG, fMRI, ultrasound) and aim to extend the same approaches to other complex domains where data is scarce and difficult to collect.

We finished as silver medalists at the NeurIPS 2025 EEG challenge and have had recent papers accepted to CVPR and submitted to ICML. This feels like we’re just beginning to find our footing. For every problem we solve, we seem to stumble upon three more fundamental questions we can’t answer yet. We’re still trying to figure out the basics of how these signals behave, and we find that uncertainty is much more motivating than the ‘solved’ parts.

What the Work Looks Like

This role is about architecture-level thinking:

• Understanding Failures

  • Why do standard methods fail on irregular, noisy, or scarce data? How can we adapt?

• Pre-Training Strategy

  • Develop methods tailored to physical signals (small samples, irregular structures).

• Multimodal Alignment

  • Bridge gaps between signals and find the right inductive biases for physical systems.

• End-to-End Ownership

  • If the core is scientific exploration (experimental and evaluation design), note that as a member of a small team, you’ll own experiments from hypothesis to deployment and contribute to engineering (pipelines, infrastructure).

RequirementsEducation:

• RequirementsEducation: PhD (preferred) in Machine Learning, Deep Learning, or a closely related quantitative field.
• Technical Expertise: Strong proficiency in PyTorch and experience running large models on GPU infrastructure
• Data Experience: Hands-on experience handling challenging/complex data, including irregular time series, scientific signals, and geometric or physical structures.
• The “Data-First” Mindset: A practical approach to research where you prioritise analysing raw signals and data quality over just monitoring loss curves.
• Navigating Ambiguity: Ability to thrive in an environment with open-ended problems and the communication skills to make complex research approachable for non-scientific teams.
• Fluent English (the team speaks English in the day-to-day)
• Bonus Skills
• Specialised Deep Learning: Experience with multimodal alignment or large-scale pre-training and foundation models.
• Advanced Domains: Specific knowledge in geometric deep learning or physical signal processing.
• Academic Track Record: A history of publications in top-tier venues such as NeurIPS, ICML, or equivalent conferences.
• Unconventional Backgrounds: Experience working with “exotic” signals or unique data systems that demonstrate deep thinking and adaptability.

Beyond Technical Skills:

While technical excellence is critical, we place equal importance on how we work together. We believe the best teams are built on:

• Integrity & Respect We are striving for honesty, kindness, and fairness. We value people who treat others with dignity and foster an environment where everyone feels heard.
• Open Communication & Humility Great ideas come from collaboration. We look for teammates who listen actively, communicate clearly, and approach challenges with self-awareness and humility.
• Psychological Safety & Camaraderie We strive to create a space where people feel safe to take risks, ask questions, and grow.