Study fast learning rates for square loss in dependent data with hypercontractivity condition.
problem Learning from dependent data with fast rates matching independent data.
method Martingale difference noise, trajectory hypercontractivity condition, least-squares estimator.
result Excess risk bound matches iid rate after burn-in time, independent of mixing-time.
New algorithm learns optimal policies with minimal memory and time.
problem Learning optimal policies in discounted MDPs with short burn-in time.
method Variance reduction and adaptive policy switching.
result First regret-optimal model-free algorithm with low burn-in time.
Optimal algorithms for non-linear ridge bandits reduce burn-in cost.
problem Non-linear models introduce a burn-in period with fixed cost.
method Two-stage algorithm: find initial action, then treat locally linear.
result Two-stage algorithm is statistically optimal.
Recent work on imitation learning has generated policies that reproduce expert behavior from multi-modal data. However, past approaches have focused only on recreating a small number of distinct, expert maneuvers, or have relied on supervised learning techniques that produce unstable policies. This work extends InfoGAI…
New algorithm reduces cold-start costs in multi-armed bandits for many products.
problem High burn-in costs in multi-armed bandits for new products.
method Two-phase bandit algorithm using subsampling and low-rank matrix estimation.
result Reduces burn-in costs and expedites experiment in large product sets.
New RL algorithms reduce costs for single-agent and federated learning.
problem Minimizing costs in RL and federated RL settings.
method Q-EarlySettled-LowCost and FedQ-EarlySettled-LowCost algorithms.
result First algorithms to achieve low burn-in and logarithmic switching costs.
New bounds on trajectory safety in training models with Langevin Dynamics.
problem Bounding the probability of a model's trajectory staying away from a designated failure region.
method Analyzes Langevin dynamics on smooth, strongly convex loss landscapes, introducing shape-free and local relaxation bounds.
result The in-set probability relaxes to the static value after a burn-in time of order d, using only the global spectral gap of the loss.
ULA estimates covariance of log-concave distributions efficiently.
problem Estimating covariance matrices of log-concave distributions efficiently.
method Unadjusted Langevin algorithm (ULA) for sampling and covariance estimation.
result Sample complexity of single-chain ULA is smaller than that of parallel ULA by a logarithmic factor.
Two algorithms learn Gaussian graphical models from Glauber dynamics trajectories.
problem Learning Gaussian graphical models from dependent data.
method Two complementary approaches: local edge-testing and burn-in/thinning reduction.
result Both approaches provide finite-sample recovery guarantees and empirical comparisons.
Parallelized bandit algorithms speed up decision-making.
problem Sequential exploration limits decision-making speed.
method Developed parallel contextual bandit algorithms.
result Regret nearly identical to sequential algorithms.
A new algorithm reduces memory and computational needs for reinforcement learning.
problem Memory and computational inefficiency in model-free reinforcement learning.
method Memory-Efficient Nash Q-Learning (ME-Nash-QL) for two-player zero-sum games.
result Proves ME-Nash-QL reduces space and sample complexity for tabular and long-horizon cases.
SGD shows distinct phases in learning single-index models, achieving optimal sample complexity and regret.
problem Learning single-index models with SGD in adaptive data settings.
method Stochastic gradient descent (SGD) with an optimal learning rate schedule.
result SGD achieves near-optimal sample complexity and regret guarantees across both burn-in and learning phases.
GD with large, adaptive stepsizes achieves optimal risk in logistic regression.
problem Optimizing logistic regression with large stepsizes.
method Gradient Descent with adaptive stepsizes.
result GD achieves minimax optimal convergence in logistic regression.
Study bounds noise level in linear regression with dependent data.
problem Analyzing noise level in linear regression with dependent data.
method Derive upper bounds for random design linear regression with β β β -mixing data, without realizability assumptions. result Correctly recovers the noise level of the problem, exhibiting graceful degradation with misspecification.
Observer learns optimal policy from learner's actions without rewards.
problem Learning optimal policy from non-rewarded actions of a non-stationary learner.
method Two-Phase Suffix Imitation framework.
result Observer achieves convergence rate of O ~ ( 1 / N ) \tilde O(1/\sqrt{N}) O ~ ( 1/ N ) . This paper is concerned with offline reinforcement learning (RL), which learns using pre-collected data without further exploration. Effective offline RL would be able to accommodate distribution shift and limited data coverage. However, prior algorithms or analyses either suffer from suboptimal sample complexities or …
We present a general framework for accelerating a large class of widely used Markov chain Monte Carlo (MCMC) algorithms. Our approach exploits fast, iterative approximations to the target density to speculatively evaluate many potential future steps of the chain in parallel. The approach can accelerate computation of t…
Particle MCMC is a class of algorithms that can be used to analyse state-space models. They use MCMC moves to update the parameters of the models, and particle filters to propose values for the path of the state-space model. Currently the default is to use random walk Metropolis to update the parameter values. We show …
Near-optimal tests and confidence sequences for non-parametric data.
problem Flexible statistical inference and decision-making with non-parametric data.
method Classic delayed-start normal-mixture sequential probability ratio tests with asymptotic guarantees.
result Asymptotically optimal type-I error and expected rejection time guarantees.
New compression methods handle biased input sequences for more accurate posterior summaries.
problem Handling biased input sequences for accurate posterior summaries.
method Stein kernel thinning, low-rank SKT, Stein recombination, Stein Cholesky.
result Achieves accurate posterior summaries with biased input sequences.
In this paper we address the following question: Can we approximately sample from a Bayesian posterior distribution if we are only allowed to touch a small mini-batch of data-items for every sample we generate?. An algorithm based on the Langevin equation with stochastic gradients (SGLD) was previously proposed to solv…
New GLB algorithm handles non-stationary data with forgetting.
problem Non-stationary GLB with non-convex projection or burn-in phases.
method Self-concordant GLB with sliding window or exponential weights for forgetting.
result Novel confidence-based algorithm for maximum likelihood estimator.
UCRL3 improves UCRL2's efficiency in reinforcement learning by reducing exploration.
problem Long burn-in phases in numerical experiments of UCRL2.
method UCRL3 uses state-of-the-art time-uniform concentration inequalities and adaptive support computation to tighten exploration.
result UCRL3 achieves a better numerical improvement over UCRL2 in standard environments.
The paper analyzes Adam and SGD in nonstationary optimization, revealing tradeoffs between noise and drift.
problem Analyzing Adam and SGD in nonstationary optimization problems.
method Theoretical analysis of Adam and SGD under non-stationary stochastic objectives, separating two regimes.
result Characterizes the tradeoff between noise and drift in Adam and SGD, revealing when adaptive step-sizing is beneficial or harmful.
The paper improves importance sampling and MCMC methods for complex distributions.
problem Improving sampling efficiency for distributions with atoms or heavy tails.
method Develops minimax optimal trial distributions and importance-tempered MCMC.
result Importance-tempered MCMC can be uniformly ergodic for certain distributions.
Communication costs, resulting from synchronization requirements during learning, can greatly slow down many parallel machine learning algorithms. In this paper, we present a parallel Markov chain Monte Carlo (MCMC) algorithm in which subsets of data are processed independently, with very little communication. First, w…
Restricted Boltzmann machines (RBMs) are powerful machine learning models, but learning and some kinds of inference in the model require sampling-based approximations, which, in classical digital computers, are implemented using expensive MCMC. Physical computation offers the opportunity to reduce the cost of sampling …
Enhanced MH algorithm reduces expensive function evaluations and improves sampling efficiency.
problem Computational expense of evaluating target distributions or likelihood functions, especially with big data.
method Accelerated MH algorithm using Bayesian optimization and Gaussian processes.
result Significant improvement in sampling efficiency and reduced function evaluations.
BART improves predictive performance but slows down with more data.
problem Understanding and improving the computational efficiency of BART with large datasets.
method Asymptotic analysis of a modified BART sampler, focusing on hitting time of high posterior density sets.
result The convergence time of the BART sampler increases with the number of training samples due to multi-modality, but can be mitigated by increasing the number of trees or raising the sampler temperature.
Particle Metropolis-Hastings (PMH) allows for Bayesian parameter inference in nonlinear state space models by combining Markov chain Monte Carlo (MCMC) and particle filtering. The latter is used to estimate the intractable likelihood. In its original formulation, PMH makes use of a marginal MCMC proposal for the parame…
Optimizes MCMC output compression by selecting a subset of states.
problem Improving the efficiency of MCMC output compression.
method Greedy minimization of kernel Stein discrepancy for selecting a subset of states.
result The method provides an optimal subset of states for empirical approximation.
New algorithm finds critical points in non-convex optimization with heavy-tailed gradients.
problem Non-convex stochastic optimization with heavy-tailed gradient estimates.
method Gradient clipping, momentum, and normalized gradient descent.
result High-probability convergence to critical points with best-known rates.
This work analyzes the convergence rate of unrolling for optimizing quadratic objectives.
problem The challenge of accurately computing Jacobians through optimization.
method Non-asymptotic convergence-rate analysis of unrolled differentiation for gradient descent and Chebyshev method.
result There is a trade-off between fast asymptotic convergence and immediate but slower convergence due to the learning rate.
Regularized Stein thinning improves MCMC output approximations.
problem Pathologies in Stein thinning leading to poor approximations.
method Theoretical analysis and regularization to improve KSD.
result Regularized Stein thinning alleviates pathologies and improves efficiency.
Bayesian realized EGARCH models improve tail risk forecasting.
problem Forecasting tail risks in financial markets.
method Developed a Bayesian framework for realized EGARCH models, incorporating multiple realized volatility measures and using robust adaptive Metropolis algorithm for estimation.
result Standardized skewed Student-t distribution and sub-sampled realized range models outperform other models in tail risk forecasting.
This study provides an explicit expansion of KL divergence's gradient flow in Fisher-Rao geometry.
problem Sampling techniques struggle to traverse between modes in non-convex potential functions.
method Explicit expansion of KL divergence's gradient flow in Fisher-Rao geometry.
result The convergence rate to π is independent of the potential function.
New analysis shows SGD with noise doesn't leak more privacy with more iterations.
problem Privacy loss in noisy SGD with more iterations.
method Privacy Amplification by Iteration and Sampled Gaussian Mechanism.
result Privacy loss remains constant after a burn-in period, not increasing with more iterations.
Curriculum learning improves model training efficiency and performance.
problem Expensive training of reasoning models using human or synthetic data.
method Autocurriculum, where models use their own performance to decide training problems.
result Autocurriculum reduces training costs and improves model performance.
Spectral gradient methods outperform Euclidean in certain deep learning scenarios.
problem When do spectral gradient updates outperform Euclidean in deep learning?
method Layerwise condition comparing squared nuclear-to-Frobenius ratio to stable rank of activations.
result Spectral updates can be more effective than Euclidean in deep networks and transformers.
New RL algorithm reduces sample complexity for optimal learning.
problem Achieving optimal learning with minimal samples in RL.
method Early-settled variance reduction method with Q-learning sequences.
result Near-optimal regret achieved with sample size S A p o l y ( H ) SA\,\mathrm{poly}(H) S A poly ( H ) . New algorithm reduces regret in infinite MDPs with optimal variance-dependent bounds.
problem Infinite horizon MDPs lack optimal algorithms with low regret.
method Developed a UCB-style algorithm for average-reward and γ-regret.
result Achieved optimal variance-dependent regret bounds for both objectives.
Analyzes intrinsic time in financial markets, linking it to physical time.
problem Understanding the intrinsic nature of time in financial data.
method Presented an analytic relationship linking intrinsic and physical time, using empirical scaling laws.
result A novel empirical scaling law relating intrinsic time variability to overshoots.
New continuous-time optimization algorithms converge in finite time to local minima.
problem Finding local minima in optimization problems.
method Discontinuous dynamical systems with finite-time convergence via Lyapunov-based differential inequality.
result Finite-time convergence to strict local minima with provable settling time.
Consider power utility maximization of terminal wealth in a 1-dimensional continuous-time exponential Levy model with finite time horizon. We discretize the model by restricting portfolio adjustments to an equidistant discrete time grid. Under minimal assumptions we prove convergence of the optimal discrete-time strate…
New distances defined between space-times, proving some definite.
problem Defining distances between space-times.
method Introducing causal-null-compactifiable space-times and using cosmological time and null distance.
result Various definite distances defined, proving convergence of space-times.
Proposes a method to allocate time budgets in mixed criticality systems.
problem Managing execution time variability in mixed criticality systems.
method Quantifies execution time variability using statistical dispersion parameters and proposes a heuristic to allocate time budgets.
result The proposed heuristic reduces the probability of exceeding allocated budgets.
Paper analyzes venture capital exit decisions under inconsistent preferences.
problem Time-inconsistent preferences in venture capital exit timing.
method Modeling four types of venture capitalists with varying levels of inconsistency.
result Time-inconsistent venture capitalists exit earlier than consistent ones.
TSMB handles time delays in multivariate time series data.
problem Varying time delays in multivariate time series data complicate predictions.
method Time Series Model Bootstrap (TSMB) framework for nonparametric time delay estimation.
result TSMB improves model performance in dynamic data environments.