Mgr. Petr Šimánek

Proceedings paper

10.5220/0012317000003636

Department of Applied Mathematics
Faculty of Information Technology

Optimization is an integral part of modern deep learning. Recently, the concept of learned optimizers has emerged as a way to accelerate this optimization process by replacing traditional, hand-crafted algorithms with meta-learned functions. Despite the initial promising results of these methods, issues with stability and generalization still remain, limiting their practical use. Moreover, their inner workings and behavior under different conditions are not yet fully understood, making it difficult to come up with improvements. For this reason, our work examines their optimization trajectories from the perspective of network architecture symmetries and parameter update distributions. Furthermore, by contrasting the learned optimizers with their manually designed counterparts, we identify several key insights that demonstrate how each approach can benefit from the strengths of the other.

Proceedings paper

10.1609/aaai.v38i21.30514

Department of Applied Mathematics

Modern machine learning heavily relies on optimization, and as deep learning models grow more complex and data-hungry, the search for efficient learning becomes crucial. Learned optimizers disrupt traditional handcrafted methods such as SGD and Adam by learning the optimization strategy itself, potentially speeding up training. However, the learned optimizers' dynamics are still not well understood. To remedy this, our work explores their optimization trajectories from the perspective of network architecture symmetries and proposed parameter update distributions.

Article

10.1007/s42979-024-02670-2

Department of Applied Mathematics

The intuitive approach of comparing like with like, forms the basis of the so-called nearest neighbor analysis, which is central to many machine learning algorithms. Nearest neighbor analysis is easy to interpret, analyze, and reason about. It is widely used in advanced techniques such as uncertainty estimation in regression models, as well as the renowned k-nearest neighbor-based algorithms. Nevertheless, its high inference time complexity, which is dataset size dependent even in the case of its faster approximated version, restricts its applications and can considerably inflate the application cost. In this paper, we address the problem of high inference time complexity. By using gradient-boosted regression trees as a predictor of the labels obtained from nearest neighbor analysis, we demonstrate a significant increase in inference speed, improving by several orders of magnitude. We validate the effectiveness of our approach on a real-world European Car Pricing Dataset with approximately rows for both residual cost and price uncertainty prediction. Moreover, we assess our method’s performance on the most commonly used tabular benchmark datasets to demonstrate its scalability. The link is to github repository where the code is available: https://github.com/koutefra/uncertainty_experiments.

Proceedings paper

Department of Applied Mathematics

Weather4cast again advanced modern algorithms in AI and machine learning through a highly topical interdisciplinary competition challenge: The prediction of hi-res rain radar movies from multi-band satellite sensors, requiring data fusion, multi-channel video frame prediction, and super-resolution. Accurate predictions of rain events are becoming ever more critical, with climate change increasing the frequency of unexpected rainfall. The resulting models will have a particular impact where costly weather radar is not available. We here present highlights and insights emerging from the thirty teams participating from over a dozen countries. To extract relevant patterns, models were challenged by spatio-temporal shifts. Geometric data augmentation and test-time ensemble models with a suitable smoother loss helped this transfer learning. Even though, in ablation, static information like geographical location and elevation was not linked to performance, the general success of models incorporating physics in this competition suggests that approaches combining machine learning with application domain knowledge seem a promising avenue for future research. Weather4cast will continue to explore the powerful benchmark reference data set introduced here, advancing competition tasks to quantitative predictions, and exploring the effects of metric choice on model performance and qualitative prediction properties.

Proceedings paper

10.1109/IJCNN55064.2022.9892364

Department of Applied Mathematics

Designing faster optimization algorithms is of ever-growing interest. In recent years, learning to learn methods that learn how to optimize demonstrated very encouraging results. Current approaches usually do not effectively include the dynamics of the optimization process during training. They either omit it entirely or only implicitly assume the dynamics of an isolated parameter. In this paper, we show how to utilize the dynamic mode decomposition method for extracting informative features about optimization dynamics. By employing those features, we show that our learned optimizer generalizes much better to unseen optimization problems in short. The improved generalization is illustrated on multiple tasks where training the optimizer on one neural network generalizes to different architectures and distinct datasets.

Proceedings paper

10.1109/IV51971.2022.9827089

Department of Applied Mathematics
Faculty of Information Technology

Prediction of the movement of all traffic participants is a very important task in autonomous driving. Well-predicted behavior of other cars and actors is crucial for safety. A sequence of bird’s-eye view artificially rasterized frames are used as input to neural networks which are trained to predict the future behavior of the participants. The Lyft Motion Prediction for Autonomous Vehicles dataset is explored and adapted for this task. We developed and applied a novel approach where the prediction problem is viewed as a problem of spatiotemporal prediction and we use methods based on convolutional recurrent neural networks.