Matthias Minderer

I am a Staff Research Scientist and manager at Google DeepMind in Zürich, where I work on computer vision.

Before joining Google, I received a PhD from Harvard University, working with Christopher Harvey. My thesis focused on the representation of visual and action-related information in the mammalian cerebral cortex. Earlier, I studied neuroscience at ETH Zürich and biochemistry at the University of Cambridge.

Email  /  CV  /  Google Scholar

I work on vision-language models and multimodal large language models, with a focus on spatial understanding and localization.

Besides my research work, I collaborate with product teams across Alphabet on integrating state-of-the-art deep learning models into applications covering a wide range of industries.

For a full list, see Google Scholar.

PaliGemma is an open Vision-Language Model (VLM) that is based on the SigLIP-So400m vision encoder and the Gemma-2B language model. I contributed to the model's spatial understanding capabilities.

We propose a simple end-to-end architecture for detecting objects and their visual relationships.

We scale open-vocabulary object detection through self-training on pseudo-labeled web image-text data.

We develop a simple transformer-based architecture and training recipe that achieves strong performance in open-vocabulary object detection.

We study the uncertainty calibration and its relationship with accuracy of recent state-of-the-art image classification models.

We show that a pure transformer, applied directly to sequences of image patches, can perform very well on image classification tasks.

We tackle the problem of low-level shortcuts in self-supervised learning by training an adversarial "lens" to remove shortcut features from images.

By using a spatially structured (keypoint-based) image representation, we improve video prediction quality and the usefulness of the learned video representations.

Using large-scale neural recordings and deep models of neural encoding, we show that navigation-related information is distributed and varies gradually across large parts of the posterior cortex, even across retinotopic boundaries.

Contrary to the idea that representations of sensory stimuli or the activity patterns that accompany motor actions are stable, neuronal representations in the parietal cortex can change across days, possibly allowing for the tradeoff between stable encoding of information and flexibility for incorporating new information.

We discuss the advantages of using virtual reality to study sensorimotor representations in the brain.

We present a method for fast fluorescence imaging of map-level cortical activity using a calcium indicator protein. Sensory-evoked neuronal activity can be imaged repeatedly in the same mouse over weeks, enabling new opportunities for the longitudinal study of cortical function and dysfunction.

Website design by Jon Barron.