Machine Learning

Artificial Intelligence, made beautifully human.

Machine learning is the process by which computers learn patterns in data. These patterns can be used to examine the past (causality), or predict the future (forecasting), or imagine new possibilities altogether (generation).

We build systems that learn from data without seeing a single data point and algorithms that forget data when asked. We create models that adapt without retraining from scratch to preserve energy and the Earth. We do this work because we believe that models have the right to forget, and people have the right to be forgotten.

The questions we ask here are not solely technical. We ask: how do we build systems that respect user privacy? What is the role of machine intelligence in a world where every person is unique? How do we create models that reflect resilient and diverse human communities?

Sometimes the answers are surprising, and sometimes they are not. But they are always interesting.

Check out the papers we're reading.

  • Bandit-based Monte Carlo Planning
    Kocsis & Szepesvári

    Introduces bandit-based sampling methods for Monte Carlo planning, enabling efficient exploration of large decision trees with theoretical guarantees.

    banditsmonte-carlo+1 more
  • Bandit Algorithms
    Lattimore & Szepesvári

    A comprehensive treatment of stochastic and adversarial bandit algorithms, covering regret bounds, exploration strategies, and theoretical foundations.

    banditsregret+1 more
  • Learning Near-Optimal Policies
    Various

    Studies the conditions under which learning algorithms can achieve near-optimal decision policies under uncertainty and partial feedback.

    reinforcement-learningcontrol+1 more
  • Regret Bounds for Adaptive Control
    Abbasi-Yadkori & Szepesvári

    Develops regret guarantees for adaptive control in unknown dynamical systems, bridging control theory and online learning.

    adaptive-controlregret+1 more
  • To believe or not to believe your llm: Iterative prompting for estimating epistemic uncertainty
    Yadkori et al.

    Examines methods for quantifying epistemic uncertainty in learned models, with implications for exploration, safety, and robustness. They also distinguish between aleatoric and epistemic uncertainty.

    uncertaintybayesian+1 more
  • Almost Free: Self-concordance in Natural Exponential Families and an Application to Bandits
    Liu et al.

    Proves the self-concordant properties of exponential family models, yielding sharp optimization guarantees for second-order methods. This is a great example of how to use the theory of exponential families to get practical results.

    optimizationinformation-geometry+1 more
  • IMSTATS Special Issue on Reinforcement Learning
    Various

    A special issue on reinforcement learning, featuring papers from the field of reinforcement learning. It's a great overview of the field and features two University of Michigan faculty members.

    reinforcement-learningoverview
  • Reinforcement Learning
    Sutton & Barto

    The foundational text on reinforcement learning, formalizing learning through interaction via value functions, policies, and returns.

    reinforcement-learningfoundations
  • Meta-Learning Dynamics
    Neil C. Rabinowitz.

    Explores the optimization and generalization dynamics underlying gradient-based meta-learning algorithms.

    meta-learningoptimization
  • Ray Interference and Plateaus in Deep Learning
    Schaul et al.

    RL agents have the unique ability to adaptively adjust their learning algorithms using a control policy. The paper shows how this can lead them to control the data-generating process from which they're supposed to learn.

    deep-learningdynamics
  • Exact Solutions to Nonlinear Dynamics of Learning
    Saxe et al.

    The team analytically describes the dynamics of learning in deep linear networks. Those dynamics are surprisingly nonlinear, indicating that the learning process is not as simple as it seems.

    learning-dynamicstheory
  • On the Spectral Bias of Neural Networks
    Rahaman et al.

    Shows that neural networks learn low-frequency components first, linking optimization dynamics to function complexity.

    spectral-biasgeneralization
  • Learning to Reinforcement Learn
    Wang et al.

    Question of interest: if we train an agent to learn its own algorithm, how does it choose what to prioritize? What does it learn to ignore?

    meta-learningreinforcement-learning
  • Federated Learning for Internet of Things
    Nguyen et al.

    Presents federated optimization methods for learning across decentralized, privacy-sensitive devices.

    federated-learningdistributed-systems
  • Deep Policy Gradient Methods
    Ilyas et al.

    Authors analyze the behavior of deep policy gradients for RL agents. This indicates that benchmarks aren't the most reliable way to evaluate RL agents.

    policy-gradientsreinforcement-learning

Posts related to this pillar.