Michael Poli

Research:
I lead a research group at Liquid AI. We are building hardware-aware, custom and efficient AI systems at every scale, with a design theory for models grounded in classical first principles of dynamical systems, signal processing and numerical linear algebra. We are hiring exceptional researchers with diverse backgrounds (AI, systems, hardware, control theory etc), reach out if our mission resonates with you.

My interests are at the intersection of machine learning, systems and signal processing:

• Architectures for long context and new applications
  T1, Hyena and HyenaDNA, StripedHyena, Evo (on the cover of Science!);
• Automating and understanding the process of architecture design
  Zoology, Mechanistic Architecture Design; Evolutionary Synthesis of Tailored Architectures (STAR)
• Hybridization of methods from dynamical systems, signal processing, with learning
  Hypersolvers, Differentiable Multiple Shooting; RTF
• Efficient systems and algorithms for scaling and deployment on the edge
  LaughingHyena;

In a previous life, I have also spent time working on neural differential equations, time series and dynamical systems. These days I am mostly interested in "full-stack" design of large deep learning models, from numerics, systems, training, all the way to finetuning and deployment.