Razvan Pascanu

Also known as Răzvan Pașcanu

Razvan Pascanu is a machine learning researcher at Google DeepMind whose work centers on recurrent networks, optimization and continual learning. He grew up in Romania, studied computer science and electrical engineering in Germany, earned an MSc at Jacobs University Bremen in 2009, and completed a PhD under Yoshua Bengio at the Université de Montréal in 2014. In 2012, he proposed gradient norm clipping for training recurrent networks in a paper with Tomas Mikolov and Bengio. After joining DeepMind in 2014, he coauthored Progressive Neural Networks and a catastrophic forgetting paper in 2016, Progress & Compress in 2018, and the Linear Recurrent Unit paper in 2023. He is a Senior Research Scientist at Google DeepMind and an Affiliate Member at Mila.

Now at: Google DeepMind, Mila and Jagiellonian University
Badges: Bridge

Timeline

2009

A master's at Jacobs University

After growing up in Romania and studying computer science and electrical engineering in Germany, he earns an MSc at Jacobs University Bremen in 2009 under the supervision of Herbert Jaeger.
2012

Clipping exploding gradients

In November 2012, he posts "On the difficulty of training Recurrent Neural Networks" with Tomas Mikolov and Yoshua Bengio. The paper analyses the vanishing and exploding gradient problems and proposes a gradient norm clipping strategy for exploding gradients and a soft constraint for vanishing gradients.
2014

PhD and saddle-free Newton

He completes a PhD at the Université de Montréal in 2014 under Yoshua Bengio, working on optimization, recurrent models and deep learning. In June 2014, he coauthors a paper with Yann Dauphin, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli and Bengio that proposes the saddle-free Newton method for escaping saddle points in high-dimensional optimization.
2014

Joining Google DeepMind

He becomes a research scientist at Google DeepMind in 2014. From there he goes on to contribute to deep reinforcement learning, continual learning, meta-learning and graph neural networks.
2016

Fighting catastrophic forgetting

In June 2016, he coauthors Progressive Neural Networks with Andrei Rusu, Raia Hadsell and colleagues, an architecture that uses lateral connections to previously learned features and is immune to forgetting. In December 2016, he is second author of "Overcoming catastrophic forgetting in neural networks", led by James Kirkpatrick, whose method slows learning on the weights important for earlier tasks and learns several Atari 2600 games in sequence.
2018

Continual learning and graph networks

In 2018, he coauthors Progress & Compress, a continual learning framework accepted at ICML 2018 that distills each newly learned task from an active column into a knowledge base. He is also among the 27 authors of "Relational inductive biases, deep learning, and graph networks", which presents the graph network as a building block for relational reasoning. In October 2018, he presents Progress & Compress at Politehnica University of Timișoara as a research scientist at DeepMind London.
2023

The Linear Recurrent Unit

In March 2023, he coauthors "Resurrecting Recurrent Neural Networks for Long Sequences" with Antonio Orvieto, Soham De and four others, which introduces the Linear Recurrent Unit, an RNN block that matches deep state-space models on the Long Range Arena benchmark. In November 2023, he gives the IC Colloquium "Towards understanding the learning dynamics of neural networks" at EPFL. He is a main organizer of the Conference on Lifelong Learning Agents (CoLLAs) and the Eastern European Machine Learning Summer School (EEML).
As of October 2026

Research scientist at Google DeepMind

He is a Senior Research Scientist at Google DeepMind and an Affiliate Member at Mila. In 2025, he is a keynote speaker at ICSTCC 2025 with a talk on recurrent neural networks and state space models, and he is a faculty member at the Faculty of Mathematics and Computer Science of Jagiellonian University in Kraków.