Our guest speaker is Tengda Han from the Visual Geometry Group (VGG), University of Oxford

You are all cordially invited to the CVG's Seminar on the 17th of March at 3:00 pm CET

• via Zoom (passcode is 690015).

Abstract

Videos are an appealing data source to train computer vision models. There exist almost infinite supplies of videos online, but exhaustive manual annotation is infeasible. In this talk, I will briefly introduce a few methods to learn strong video representations with minimal human annotations, with an emphasis on long videos which go beyond a few seconds.

Bio

Tengda Han is a post-doctoral research fellow at the Visual Geometry Group at the University of Oxford. He obtained his PhD from the same group in 2022 supervised by Andrew Zisserman. His current research focuses on self-supervised learning, efficient learning and video understanding.

Our guest speaker is Karsten Kreis from NVIDIA’s Toronto AI Lab.

You are all cordially invited to the CVG Seminar on the 22nd of December at 5:30 pm CET

• via Zoom (passcode is 052316).

Abstract

Denoising diffusion-based generative models have led to multiple breakthroughs in deep generative learning. In this talk, I will discuss recent works by the NVIDIA Toronto AI Lab on diffusion models. In the first part, I will present GENIE: Higher-Order Denoising Diffusion Solvers, a novel method for accelerated sampling from diffusion models, leveraging higher-order methods together with an efficient model distillation technique to solve the generative differential equations of diffusion models. Next, I will discuss our work on Critically-Damped Langevin Diffusion. Taking inspirations from statistical mechanics and Markov chain Monte Carlo, we introduce an auxiliary velocity variable into the diffusion process, which allows the diffusion to converge to the Gaussian prior more smoothly and quickly. This makes critically-damped Langevin diffusion ideally suited for diffusion-based generative modeling. Finally, I will briefly recapitulate Latent Score-based Generative Models and then present LION: Latent Point Diffusion Models for 3D Shape Generation, which achieves state-of-the-art 3D shape synthesis and enables various artistic applications, such as voxel-guided shape generation.

Bio

Karsten Kreis is a senior research scientist at NVIDIA’s Toronto AI Lab. Prior to joining NVIDIA, he worked on deep generative modeling at D-Wave Systems and co-founded Variational AI, a startup utilizing generative models for drug discovery. Before switching to deep learning, Karsten did his M.Sc. in quantum information theory at the Max Planck Institute for the Science of Light and his Ph.D. in computational and statistical physics at the Max Planck Institute for Polymer Research. Currently, Karsten's research focuses on developing novel generative learning methods and on applying deep generative models on problems in areas such as computer vision, graphics and digital artistry, as well as in the natural sciences.

Our guest speaker is Yuki Asano from the University of Amsterdam.

You are all cordially invited to the CVG Seminar on the 9th of December at 2:30 p.m. CET

• via Zoom (passcode is 303207).

Abstract

It is a talk about pushing the limits of what can be learnt without using any human annotations. After a first overview of what self-supervised learning is, we will first dive into how clustering can be combined with representation learning using optimal transport and how this can be leveraged to unsupervisedly segment objects in images [1]. Finally, as augmentations are crucial for all of the self-supervised learning, we will analyze these in more detail in a recent preprint [2]. Here, we show that it is possible to extrapolate to semantic classes such as those of ImageNet using just a single datum as visual input when combined with strong augmentations.

[1] Self-Supervised Learning of Object Parts for Semantic Segmentation [arxiv]

[2] Extrapolating from a Single Image to a Thousand Classes using Distillation [arxiv]

Bio

Yuki Asano is an assistant professor for computer vision and machine learning at the Qualcomm-UvA lab at the University of Amsterdam, where he works with Cees Snoek, Max Welling and Efstratios Gavves. His current research interests are multi-modal and self-supervised learning and ethics in computer vision. Prior to his current appointment, he finished his PhD at the Visual Geometry Group (VGG) at the University of Oxford working with Andrea Vedaldi and Christian Rupprecht. During his time as a PhD student, he also interned at Facebook AI Research and worked at TransferWise. Prior to the PhD, he studied physics at the University of Munich (LMU) and Economics in Hagen as well as a MSc in Mathematical Modelling and Scientific Computing at the Mathematical Institute in Oxford.

You are all cordially invited to the Master Thesis defence on the 27th of October at 11 a.m. CEST

• in person at the Institute of Computer Science: room 302, Neubrückstrasse 10, 3012 Bern
• via Zoom (passcode is 327699).

Abstract

Depression is the second major cause for years spent in disability and has a growing prevalence in adolescents. The recent Covid-19 pandemic has intensified the situation and limited in-person patient monitoring due to distancing measures. Recent advances in wearable devices have made it possible to record the rest/activity cycle remotely with high precision and in real-world contexts. We aim to use machine learning methods to predict an individual's mental health based on wearable-measured sleep and physical activity. Predicting an impending mental health crisis of an adolescent allows for prompt intervention,  detection of depression onset or its recursion, and remote monitoring. To achieve this goal, we train three primary forecasting models; linear regression, random forest, and light gradient boosted machine (LightGBM); and two deep learning models; block recurrent neural network (block RNN) and temporal convolutional network (TCN); on Actigraph measurements to forecast mental health in terms of depression, anxiety, sleepiness, stress, sleep quality, and behavioural problems. Our models achieve a high forecasting performance, the random forest being the winner to reach an accuracy of 98% for forecasting the trait anxiety. We perform extensive experiments to evaluate the models' performance in accuracy, generalization, and feature utilization, using a naive forecaster as the baseline. Our analysis shows minimal mental health changes over two months, making the prediction task easily achievable. Due to these minimal changes in mental health, the models tend to primarily use the historical values of mental health evaluation instead of Actigraph features. At the time of this master thesis, the data acquisition step is still in progress. In future work, we plan to train the models on the complete dataset using a longer forecasting horizon to increase the level of mental health changes and perform transfer learning to compensate for the small dataset size. This interdisciplinary project demonstrates the opportunities and challenges in machine learning-based prediction of mental health, paving the way toward using the same techniques to forecast other mental disorders such as internalizing disorder, Parkinson's disease, Alzheimer's disease, etc. and improving the quality of life for individuals who have some mental disorder.

You are all cordially invited to the Bachelor Thesis defence on the 23rd of September at 2:30 p.m. CEST

• in person at the Institute of Computer Science: room 302, Neubrückstrasse 10, 3012 Bern
• via Zoom (passcode is 871438).

Abstract

The calculation of variables of brain morphology is computationally very expensive and time-consuming. Previous work showed the feasibility of extracting the variables directly from T1-weighted brain MRI images using a convolutional neural network. We used significantly more data and extended their model to a new set of neuromorphological variables, which could become interesting biomarkers in the future for the diagnosis of brain diseases. The model shows for nearly all subjects a less than 5% mean relative absolute error. This high relative accuracy can be attributed to the low morphological variance between subjects and the ability of the model to predict the cortical atrophy age trend. The model however fails to capture all the variance in the data and shows large regional differences. We attribute these limitations in part to the moderate to poor reliability of the ground truth generated by FreeSurfer. We further investigated the effects of training data size and model complexity on this regression task and found that the size of the dataset had a significant impact on performance, while deeper models did not perform better. Lack of interpretability and dependence on a silver ground truth are the main drawbacks of this direct regression approach.