I am a scientist and entrepreneur working at the intersection of computational neuroscience, machine learning and computer vision. My goal is to understand how neural systems – both biological and artificial – perform visual perception. I am currently based in Tübingen, Germany, and will start a lab at the University of Göttingen in October 2019. I am also co-founder of Layer7 AI and DeepArt.io.
Learn more about my research and my publications.
I am happy to announce that I accepted an offer as a faculty position at the University of Göttingen, where I will start my lab at the Computer Science Department in October 2019.
My lab will continue to work on the intersection of computational neuroscience and machine learning, investigating visual perception in both biological and artificial systems.
I have a couple of open positions for PhD students and postdocs. If you want to work with me, please send me an email with your CV and a few sentences about why you would like to join my group and what you’d like to work on.
Our paper on system identification with deep neural networks in monkey V1 has finally been published in PLoS Computational Biology. In the paper we show that multi-layer convolutional neural networks outperform other existing predictive models of V1 at predicting spiking responses to natural images. As usual, data and code are available.
Finding optimal stimuli for neurons has long been central for understanding information processing in the brain. However, it is hard because the search space is high-dimensional and sensory information processing is fundamentally nonlinear. With our collaborators Fabian Sinz, Andreas Tolias, Jake Reimer and Xaq Pitkow, we developed Inception Loops: a closed-loop optimization method combining in vivo recordings and in silico nonlinear modeling to find Most Exciting Images (MEIs) we show back to the brain. MEIs drove cells better than control stimuli revealing fascinating properties of mouse V1 cells. MEIs had sharp corners, curved strokes and pointillist textures, deviating strikingly from the standard V1 Gabor model.
The preprint of our work on one-shot object detection and instance segmentation is on arXiv. In this work, we learn to detect and segment instances of previously unseen object categories based on a single visual instruction example. For example, given the image below and either a person (left) or a car (right) as the reference, the goal of the system is to detect all persons (far left) and cars (far right), respectively. Note that in this case, neither persons nor cars were annotated in the training set.
Together with Fabian Sinz, we developed a deep recurrent neural network for predicting the activity of thousands of mouse V1 neurons simultaneously recorded with two-photon microscopy, while accounting for confounding factors such as the animal’s gaze position and brain state changes related to running state and pupil dilation. We investigated how well this large-scale model generalizes to stimulus statistics it was not trained on. While our model trained on natural movies can correctly predict some neural tuning properties in responses to artificial noise stimuli, unadapted transfer is not perfect. However, it can fully generalize from movies to noise and maintain high predictive performance on both stimulus domains by fine-tuning only the final layer’s weights. Check out the preprint on bioRxiv.
I developed an approach to organize and classify neurons in V1 according to their nonlinear computation, ignoring receptive field location and preferred orientation. We use a rotation-equivariant convolutional network to perform weight sharing not only across space, but also across orientation. Our preprint describes the approach and some early results we obtained using recordings of around 6000 neurons in mouse V1.
The final version of our ECCV 2018 paper on visualizing invariances in convolutional neural networks is available. We find that early and mid-level convolutional layers in VGG-19 exhibit various forms of response invariance: near-perfect phase invariance in some units and invariance to local diffeomorphic transformations in others. At the same time, we uncover representational differences with ResNet-50 in its corresponding layers.
Work with Santiago Cadena, Marissa Weis, Leon Gatys and Matthias Bethge.
Image content is more important than Bouma’s Law for scene metamers. After many iterations it’s now finally public. Work with Tom Wallis, Christina Funke, Leon Gatys, Felix Wichmann and Matthias Bethge.
Comments welcome!
Variability in neuronal responses to identical stimuli is frequently correlated across a population. Attention is thought to reduce these correlations by suppressing noisy inputs shared by the population. However, even with precise control of the visual stimulus, the subject’s attentional state varies across trials. In 2016, we put out the hypothesis that such fluctuations in attentional state could be a cause for some of the correlated variability observed in cortical areas. To address this question empirically, we designed a novel paradigm that allows us to manipulate the strength of attentional fluctuations.
In the new paper just published in Nature Communications, we recorded from monkeys’ primary visual cortex (V1) while they were performing this task. We found both a pronounced effect of attentional fluctuations on correlated variability at long timescales and attention-dependent reductions in correlations at short timescales. These effects predominate in layers 2/3, as expected from a feedback signal such as attention.
Our paper on one-shot segmentation in clutter has been accepted to ICML. In this paper, we tackle a one-shot visual search task: based on a single instruction example (the red Φ in the image below), the goal is to find the same letter in a cluttered image that consists of many letters (left) and segment it. This task is pretty hard for computer vision systems, because the image clutter consists of other letters (i.e. very similar statistics), the letters can have arbitrary colors, are drawn by different people, transformed by affine transformations, and have not been seen during training.
