Gerardo Duran-Martin

Introduction Getting in the door was the hardest part of the whole ordeal.

Introduction A classical method to tackle the problem of time-series regime detection and segmentation is the hidden Markov model (HMM). 1 At the core of the HMM are two sub-models: a model for latent regimes and a predictive model for observations (condi…

Abstract We develop a predictive-first optimisation framework for streaming hidden Markov models.

Introduction In this note, we introduce the BONE framework1 for learning and adaptation in non-stationary environments.

Abstract This paper develops a framework to predict toxic trades that a broker receives from her clients.

Introduction The algorithms we studied in the previous note assumed a stationary environment: the data-generating-process (DGP) does not depend on time.

Introduction In this note, we provide an introduction to the problem of sequential learning and decision making from a Bayesian perspective.

Introduction Deep neural networks (DNNs) are great at fitting data, but terrible at predicting when they don’t know the right answer.

Introduction I was first introduced to Gaussian Processes (GPs) as the quintessential example of Bayesian reasoning in machine learning.

Abstract We introduce scalable algorithms for online learning of neural network parameters and Bayesian sequential decision making.

Online learning and sequential decision making Observe sequence of features $\vx_t\in\reals^M$ and observations $\vy_t \in \reals^o$: $$ {\cal D}_{1:t} = \{(\vx_1, \vy_1), \ldots, (\vx_t, \vy_t)\}. $$ We model a mapping from $\vx_t$ to $\vy_t$ through a f…

Abstract In this thesis, we introduce Bayesian filtering as a principled framework for tackling diverse sequential machine learning problems, including online (continual) learning, prequential (one-step-ahead) forecasting, and contextual bandits.

Part III of the filtering notes series: Part I: signal plus noise models Part II: state-space-models Part III: the Kalman filter Introduction In this post, we derive the Kalman filter (KF) algorithm and explore how it can be applied to time series forecas…

Part II of the filtering notes series: Part I: signal plus noise models Part II: state-space-models Part III: state-space-models Introduction Whether modelling the trajectory of a rocket, forecasting economic trends, or powering large language models, dis…

Abstract We propose a unifying framework for methods that perform probabilistic online learning in non-stationary environments.

Part I of the filtering notes series: Part I: signal plus noise models Part II: state-space-models Part III: state-space-models Introduction From tracking epidemics and predicting financial markets to controlling complex systems like rockets and personali…

(B)ayesian (O)nline learning in (N)on-stationary (E)nvironments Paper: arxiv/2411.10153 Slides: bone/talk Repo: gerdm/BONE How do we create agents that can continually learn and adapt to their environment?

Introduction In this post, we introduce the Bayesian Online Changepoint Detection (BOCD) model and its application for estimating the probability of heads in a sequence of coin tosses where the underlying probability changes over time.

Introduction Our weighted-observation likelihood filter (WolF) got accepted at ICML 2024.

Abstract We derive a novel, provably robust, efficient, and closed-form Bayesian update rule for online filtering in state-space models in the presence of outliers and misspecified measurement models.

Abstract We propose an efficient online approximate Bayesian inference algorithm for estimating the parameters of a nonlinear function from a potentially non-stationary data stream.

The Kalman filter Suppose $\forall t=1,\ldots,T.

Abstract In this paper we present a new algorithm for online (sequential) inference in Bayesian neural networks, and show its suitability for tackling contextual bandit problems.

Do you know what inm,kij,jnm->knm is all about?

Introduction In this post I report my work during the Google Summer of Code (GSoC) 2021 program.

Introduction On July 10, 2018 I committed myself to read and do all the exercises of the book Pattern Recognition and Machine Learning from Christopher Bishop.

Statistical machine learning About me g.duran@me.com My background is in Bayesian inference, Kalman filtering, time-series modeling, state-space models, and online learning.