Neil Houlsby

Also known as Neil M. T. Houlsby

Neil Houlsby is an AI researcher whose work on neural architectures spans adapters, the Vision Transformer and the PaLI vision-language models. A Google doctoral fellowship brought Houlsby to Cambridge's Machine Learning Group in 2010, and by 2014 Houlsby was a Research Scientist on Google's Pragmatics team in Zurich. Houlsby was first author of the 2019 paper that proposed adapter modules for parameter-efficient transfer learning in NLP, then coauthored the Vision Transformer paper in 2020, MLP-Mixer in 2021 and PaLI in 2022. Anthropic appointed Houlsby Head of Zürich in February 2025, after Houlsby led a team at Google DeepMind and Google Brain. At Anthropic's Zurich site, which works on safe frontier language models with a focus on multimodal capabilities, Houlsby is the head.

Now at: Anthropic
Badges: Bridge

Timeline

2010

A Google fellowship at Cambridge

A Google doctoral fellowship brings Houlsby to Cambridge's Machine Learning Group in 2010. The PhD in machine learning, co-supervised by Zoubin Ghahramani and Máté Lengyel in the Computational and Biological Learning lab, covers active learning and matrix modelling.
2011

Bayesian active learning

In December 2011, Houlsby submits Bayesian Active Learning for Classification and Preference Learning with Ferenc Huszár, Zoubin Ghahramani and Máté Lengyel. The method expresses information gain through predictive entropies and applies it to the Gaussian Process classifier.
By 2014

Researcher at Google Zurich

By November 2014, Houlsby is a Research Scientist on Google's Pragmatics team in Zurich, working on pragmatics, discourse and dialogue in natural language processing. An internship with Google's natural language processing research team comes first.
2019

Adapters for transfer learning

In 2019, as first author, Houlsby proposes adapter modules in Parameter-Efficient Transfer Learning for NLP. The adapters add a few trainable parameters per task while the original BERT network stays fixed, and on GLUE they come within 0.4% of full fine-tuning with 3.6% added parameters per task.
2020

The Vision Transformer

In October 2020, Houlsby coauthors An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. The paper applies a pure transformer directly to sequences of image patches, and the resulting Vision Transformer (ViT) attains excellent results compared to state-of-the-art convolutional networks while requiring substantially fewer computational resources to train.
2021

MLP-Mixer for vision

In May 2021, Houlsby coauthors MLP-Mixer: An all-MLP Architecture for Vision with Ilya Tolstikhin, Alexey Dosovitskiy and colleagues. The architecture is based exclusively on multi-layer perceptrons applied to image patches and across patches, and it attains competitive scores on image classification benchmarks.
2022

The PaLI language-image model

In September 2022, Houlsby is among the authors of PaLI (Pathways Language and Image model), which generates text from visual and textual inputs in many languages. PaLI pairs large pre-trained language models with Vision Transformers, including a 4-billion parameter ViT-e, and achieves state-of-the-art results on captioning, visual question answering and scene-text understanding.
Since 2025

Head of Anthropic Zurich

Anthropic appoints Houlsby, previously a Senior Staff Research Scientist and manager who led a team on neural architectures, vision and language models at Google DeepMind and Google Brain, as Head of Zürich in February 2025. The new office, Anthropic's third in Europe, works on safe frontier language models with a focus on multimodal capabilities. In July 2026, Houlsby leads the Zurich site and builds a team focused on pre-training and multimodal models.