Mihaela Rosca

About me

I am a Staff Research Scientist at DeepMind.

Before joining DeepMind as a Research Engineer in 2016, I was a Software Engineer at Google Zurich working on NLP using neural networks.

Education

I obtained my PhD at UCL, supervised by Prof. Marc Deisenroth. I passed my viva without corrections in May 2023. My examination committee was made by Dr. Ferenc Huszár and Dr. Patrick Rebeschini. You can find my PhD thesis here.

I finished my 4 year Computing MEng degree from Imperial College London in 2014, with first honours and a prize of excellence for an outstanding overall performance.

Professional Interests

Generative models, reinforcement learning, natural language processing, scalable and safe machine learning.

I am currently focused on understanding optimisation in deep learning.

Talks

Talk at ETH on the importance of discretisation drift in deep learning, August 2023. Slides.
Invited talk at the Math Machine Learning seminar at the Max Planck Institute for Mathematics in the Sciences and UCLA, February 2023.
Mediterranean Machine Learning Summer School: On GANs: the divergence and optimisation views (2022). Slides.
Talk on smoothness constraints at the UCL AI journal club (2022).Slides
Invited talk at the Leipzig Symposium on Intelligent Systems on discretization drift in deep learning (2022).
Invited talk at the Neurips Workshop ICBINB 2021. Talk video.
Invited lecture on generative models at the University of Liepzig (2021).
Invited talk at Imperial College London on discretization drift in Deep Learning (2021). Video
ProbAI summer school 2021 lecture on constructing GANs from divergences and distances. Slides.
Invited lecture on GANs for Computer Vision at Politehnica University of Bucharest. Slides
DeepMind UCL lecture series 2020: Generative adversarial networks. Video, Slides
DeepMind UCL lecture series 2020: Unsupervised representation learning. Video, Slides
Eastern European summer school 2020 unsupervised learning talk
Training language GANs from Scratch at Ganocracy 2019
VAE-GAN Hybrids at the GAN tutorial at CVPR 2018
The power and the promise of generative models (University of Bath talk)
Unsupervised learning for building intelligent systems
Autoencoder GAN talk at GAN workshop at ICCV 2017
TensorFlow talk at the Deep Learning, Tools and Methods workshop
TensorFlow talk at DevFest 2016
TensorFlow meetups in London and Madrid (2016)

Open source projects

Optax: JAX optimization library. Github
Monte Carlo Gradient estimation in machine learning source code. Github
Deep Compressed Sensing source code. Github
High performance GPU implementation of deep belief networks to assess their performance on facial emotion recognition from images. Github
Hopfield networks and RBM implementation in Haskell. Github

Other ML community contributions

Organiser for the ICML 2022 workshop on "Continous time perspectives for ML". Website.
Top ICML 2021 reviewer and expert reviewer. Top 200 reviewer NeurIPS 2018, Top reviewer NeurIPS 2019. Conference reviewer since 2017. JMLR and TMLR reviewer.
Lab instructor for all labs and creator of the Bayesian learning lab at EEML.
Lab instructor at TMLSS, teaching vision, autoregressive text models and latent variable image models.
Website for the GAN tutorial at ICCV 2017 that I co-organized with Ian Goodfellow.