AI Research Scientist Intern

Modern Intelligence is seeking exceptional students and recent graduates to join the team as an AI Research Scientist Intern, and contribute to Modern’s in-house research program on hierarchical neural networks and their application to computer vision, natural language, optimal dimensional reduction and transferable encodings, and more. This position is full-time, and runs during Summer 2022 with the option to extend through the end of the year. This opportunity is open to remote candidates.

About Modern’s Research:Modern Intelligence is an Artificial Intelligence product and R&D company. Our products are primarily geared towards the Department of Defense, and tackle some of the hardest and most interesting problems of the decade in machine learning, with a current focus on multi-modal domain awareness and scene understanding. And our research program focuses on foundational improvements to neural networks that aim to close the performance gap between artificial and biological neural networks. Whereas the current trend in machine learning is towards more data, more GPUs, and models with trillions of parameters, the fruit fly brain needs less than 200,000 neurons to achieve vision, target identification, navigation, and obstacle avoidance that is orders of magnitude more robust than any artificial neural network. 
The brittleness and inefficiency of artificial neural networks likely comes from a combination of: * artificial neurons having “less compute”,* artificial neural architectures produced by NAS and human engineers having suboptimal topologies, and* SGD and other training processes being highly inefficient,
relative to their biological equivalents. Modern’s research program uses proprietary advances in information theory to train hierarchical neural networks, or HNNs, with modular intermediate levels of structure that address the problem of multi-scale competence. (See a recent interview with Dr. Michael Levin [video][transcript] to learn more about the high-level biological insights that motivate multi-scale competence as a desirable [and possibly necessary] factor for robustness and efficient learning of high-complexity problems.)
About the position:Pay: $9,000 / month (full-time employment)Work period: Summer 2022, with option to extend

We’re looking for a curious and insightful early-stage researcher who believes that the answer isn’t just more data, and who has a deep appreciation of the benefits of sparsity, hierarchy, efficiency, and taking the right lessons from biology. With your team at Modern, you’ll get hands-on experience with some of the hardest problems in ML from both the product and the research direction. You’ll help build Modern’s research community, and will contribute directly to a research program that creates and trains HNNs in order to address the multi-scale competence problem, work towards natural robustness in neural networks, and create general re-usable neural modules that can be leveraged in future learning. You’ll have as much guidance as you need and as much project ownership as you want, including the opportunity and encouragement to propose, design, and run your own experiments, and to publish the results. 

Basic Qualifications:

• Recent PhD or active enrollment in a PhD program in Computer Science, Data Science, Mathematics, Statistics, Physics, or related field
• Strong proficiency in coding and data science
• Proficiency in at least one popular ML framework, e.g. Tensorflow or Pytorch.
• Strong project-level experience in deep learning and/or reinforcement learning, with publications preferred but not required
• Excellent communication skills, and the ability to work effectively in a team

Deeper Qualifications:

• Interest and project or research experience in non-incremental improvements to the state of AI performance, either from a foundational direction (e.g. via neural or sub-neural architectures), empirical (e.g. domain robustness in CV models), or resource-driven (e.g. using less data and requiring less training time).
• Strong mathematical foundations, including experience in probability theory and information theory
• Experience with computer vision models, autoencoders and variational autoencoders
• Preferred but not required: experience with self-supervised learning more broadly, experience in neural architecture search, experience with AutoML and meta learning

To apply, submit a resume or CV, and a cover letter, here on Lever. Letters of recommendation are not required but are strongly encouraged, especially if written from professors, researchers, or engineers who can speak to your interest and abilities.