Postdoc: Adaptive Optimization in Nonconvex Learning

1000scholars; Saint-Martin-d'Hères; On-site; EUR 36,000 - 54,000; Full time

Job summary

1000scholars seeks a postdoctoral researcher to deepen the understanding of adaptive optimization algorithms that adjust hyperparameters based on accumulated data. You will work with Inria GHOST and LJK teams on AdaGrad variants in stochastic, non-convex, and non-smooth contexts.

The role focuses on defining problem classes near learning models and analyzing nonsmooth algorithms, with dissemination of results through articles and conferences.

Qualifications

• Foundation knowledge in numerical optimization, in particular first-order methods.
• Scientific programming skills (Julia, Python) for empirical validations of algorithms.

Responsibilities

• Identify necessary/sufficient conditions so adaptive algorithms converge.
• Investigate robustness under noise, initialization, and initial step-size.
• Compare properties with classical methods like stochastic gradient descent.
• Map a zoology of structured optimization problems where adaptive methods outperform.

Skills

Adaptive optimization

AdaGrad/Adam/RMSProp

Scientific programming

Python

Julia

Education

PhD in optimization / ML

Tools

Python

Julia

Job description

1000scholars seeks a postdoctoral researcher to deepen the understanding of adaptive optimization algorithms that adjust hyperparameters based on accumulated data. You will work with Inria GHOST and LJK teams on AdaGrad variants in stochastic, non-convex, and non-smooth contexts.

The role focuses on defining problem classes near learning models and analyzing nonsmooth algorithms, with dissemination of results through articles and conferences.