Kevin Murphy

Profile

If you’ve spent any serious time learning machine learning the rigorous way, you already own a Kevin Murphy book — or you’ve photocopied someone else’s. Kevin Patrick Murphy is a Principal Scientist at Google DeepMind and the author of what is arguably the most complete single-author reference in the field: a 2012 doorstop, Machine Learning: A Probabilistic Perspective, followed a decade later by a two-volume rewrite. Born in Ireland, raised in England (BA, Cambridge), trained in the US (MEng at Penn, PhD at UC Berkeley under Stuart Russell, postdoc at MIT), he spent 2004–2011 as a professor at the University of British Columbia before a 2011 Google sabbatical turned permanent. He’s been at Google — first Brain, now DeepMind — ever since.

What sets Murphy apart from the celebrity-researcher tier of AI is that his lasting influence is pedagogical. His books treat machine learning as applied probability first and engineering second: everything is a distribution, inference is the unifying verb, and deep learning is one chapter in a much longer story that also includes graphical models, Bayesian inference, causality, and state-space models. For a developer who wants to understand why a transformer or a diffusion model works rather than just calling .fit(), this framing is gold. It’s also demanding — these are graduate-level texts, not weekend tutorials — but Murphy has kept the full drafts free online for years, which is why nearly every ML PhD student of the last decade has read him.

His research career predates the deep learning boom and grounds it. His 2002 Berkeley thesis on Dynamic Bayesian Networks (a unifying view of HMMs, Kalman filters, and their generalizations) remains heavily cited, and his work on graphical models, approximate inference, and probabilistic programming shaped how a generation thinks about structured uncertainty. At Google he’s ranged widely: knowledge-graph embeddings, trajectory prediction, reliability and uncertainty (the Plex line of work), multimodal generation with frozen LLMs (SPAE), and lately diffusion models for control and video.

Today he manages a team of roughly 28 researchers and engineers at DeepMind working across generative models (diffusion and LLMs), reinforcement learning, robotics, and Bayesian inference. He’s also a prolific open-source maintainer — the JAX-based dynamax library for state-space models grew directly out of the second book. If your mental model of AI is “scale is all you need,” Murphy is a useful corrective: he’s the person insisting the probabilistic foundations still matter, and he’s written the reference that proves it.

Books

📖

Probabilistic Machine Learning: An Introduction

The 2022 successor to his 2012 classic — a from-scratch introduction covering probability, linear models, deep learning, and inference, with free online drafts and code.

📖

Probabilistic Machine Learning: Advanced Topics

The 2023 sequel, going deep on graphical models, variational inference, MCMC, generative models, reinforcement learning, and causality.

📖

Machine Learning: A Probabilistic Perspective

The 2012 original — for over a decade the definitive comprehensive reference for ML through a probabilistic lens.

Key Articles & Papers

Dynamic Bayesian Networks: Representation, Inference and Learning 2002 — His UC Berkeley PhD thesis and one of his most-cited works — the unifying treatment of HMMs, Kalman filters, and their generalizations. A Review of Relational Machine Learning for Knowledge Graphs 2015 — A widely-read survey that helped define how the field embeds and reasons over structured knowledge. SPAE: Semantic Pyramid AutoEncoder for Multimodal Generation with Frozen LLMs 2023 — Lets frozen large language models understand and generate images — an elegant bridge between LLMs and pixels. Plex: Towards Reliability using Pretrained Large Model Extensions 2022 — A systematic push on uncertainty, robustness, and adaptation — reliability engineering for large models. Dynamax: A Python package for probabilistic state space modeling with JAX 2024 — The JAX-based library that turned the second book's state-space chapters into runnable, differentiable code. Diffusion Model Predictive Control 2024 — Uses diffusion models to learn both action proposals and dynamics for online model predictive control — probabilistic ML meets robotics.

Videos

YouTube