Difference between revisions of "Talks"
From Research management course
| (One intermediate revision by the same user not shown) | |||
| Line 1: | Line 1: | ||
[[File:vadim_photo_1K.jpeg|class=img-responsive|right|alt|120px|noborder|Vadim]] | [[File:vadim_photo_1K.jpeg|class=img-responsive|right|alt|120px|noborder|Vadim]] | ||
| − | * '''Title:''' Model selection in neurocomputing | + | ==Model selection in neurocomputing== |
| − | * '''Speaker:''' Vadim Strijov | + | <!-- * '''Title:''' Model selection in neurocomputing |
| + | * '''Speaker:''' Vadim Strijov--> | ||
* '''Date:''' September 14th 2024 | * '''Date:''' September 14th 2024 | ||
* '''Abstract:''' The talk focuses on the problem of dimensionality reduction in neurocomputing modeling. In a neurocomputing experiment, a participant responds to a stimulus with a limb motion. Three modalities are measured: EEG for brain activity, and IMU for limb motions. The goal is to predict limb motions using brain signals. A prediction model is selected through causal inference. We show how to create a state space, process cross-correlated signals, and select a model. We introduce a model that grabs relationships within and between the source and target data. | * '''Abstract:''' The talk focuses on the problem of dimensionality reduction in neurocomputing modeling. In a neurocomputing experiment, a participant responds to a stimulus with a limb motion. Three modalities are measured: EEG for brain activity, and IMU for limb motions. The goal is to predict limb motions using brain signals. A prediction model is selected through causal inference. We show how to create a state space, process cross-correlated signals, and select a model. We introduce a model that grabs relationships within and between the source and target data. | ||
| − | * '''Title:''' Generative machine learning models for scenario simulation | + | ==Generative machine learning models for scenario simulation== |
| − | * '''Speaker:''' Vadim Strijov | + | <!-- * '''Title:''' Generative machine learning models for scenario simulation |
| + | * '''Speaker:''' Vadim Strijov--> | ||
* '''Date:''' December 2nd 2023 | * '''Date:''' December 2nd 2023 | ||
* '''Abstract:''' This talk presents the fundamental principles of generative modeling within the context of time series applications. Given the presence of high variance and substantial covariance among time series, we represent them as phase trajectories, multimodal sets, and dynamic graphs. Our talk delves into three types of generative models: variational autoencoders, normalizing flows, and diffusion probabilistic models. The objective is to reconstruct the distribution of time series or their dynamic interrelations. The pivotal method to elucidate is Principal Component Analysis. It introduces autoencoder as an essential part of generative neural networks. | * '''Abstract:''' This talk presents the fundamental principles of generative modeling within the context of time series applications. Given the presence of high variance and substantial covariance among time series, we represent them as phase trajectories, multimodal sets, and dynamic graphs. Our talk delves into three types of generative models: variational autoencoders, normalizing flows, and diffusion probabilistic models. The objective is to reconstruct the distribution of time series or their dynamic interrelations. The pivotal method to elucidate is Principal Component Analysis. It introduces autoencoder as an essential part of generative neural networks. | ||
| − | * '''Title:''' Machine learning Model selection for biomedical signals | + | ==Machine learning Model selection for biomedical signals== |
| − | * '''Speaker:''' Vadim Strijov | + | <!-- * '''Title:''' Machine learning Model selection for biomedical signals |
| + | * '''Speaker:''' Vadim Strijov--> | ||
* '''Date:''' December 2nd 2022 | * '''Date:''' December 2nd 2022 | ||
* '''Abstract:''' We model spatial-time series: audio-video streams and brain signals like EEC, ECoG, and IMU from wearable devices. The practical application is human motion analysis for health monitoring. We discuss a forecasting model to approximate phase trajectories. Since this kind of data is highly correlated, the model selection is a fruitful way to obtain a simple, stable, and accurate model. To optimize the model structure, we use a quadratic programming problem statement. To set the criterion of optimality, we use Bayesian inference. | * '''Abstract:''' We model spatial-time series: audio-video streams and brain signals like EEC, ECoG, and IMU from wearable devices. The practical application is human motion analysis for health monitoring. We discuss a forecasting model to approximate phase trajectories. Since this kind of data is highly correlated, the model selection is a fruitful way to obtain a simple, stable, and accurate model. To optimize the model structure, we use a quadratic programming problem statement. To set the criterion of optimality, we use Bayesian inference. | ||
| + | ==About== | ||
* '''Speaker's short bio:''' Dr. Vadim Strijov is a researcher at m1p.org. He served as a professor at the Computing Center of the Russian Academy of Sciences, and the head of the Intelligent Systems Department. He obtained his D.Sc. in Physics and Mathematics with theses on Mathematical Modelling and Machine Learning. In 2020 he received the Yandex Segalovich Prize award for his significant impact on the scientific community development in the CIS countries. His research fields are AI, Machine Learning, and Functional Data Analysis. | * '''Speaker's short bio:''' Dr. Vadim Strijov is a researcher at m1p.org. He served as a professor at the Computing Center of the Russian Academy of Sciences, and the head of the Intelligent Systems Department. He obtained his D.Sc. in Physics and Mathematics with theses on Mathematical Modelling and Machine Learning. In 2020 he received the Yandex Segalovich Prize award for his significant impact on the scientific community development in the CIS countries. His research fields are AI, Machine Learning, and Functional Data Analysis. | ||
<!-- He graduated with PhD (2002) and DSc (2014) degrees in Theoretical Foundations of Computer Science from the Russian Academy of Sciences. His research field is AI and Machine Learning, Bayesian Model Selection, and Functional Data Analysis with applications to Brain-Computer Interfaces and biomedical signals. He is the chief editor of the journal "Machine Learning and Data Analysis." His educational Youtube channel "Machine Learning Phystech" holds 1200 hours of monthly watch time. In 2020 he received Yandex' Segalovich award for his impact on scientific community development in CIS countries.--> | <!-- He graduated with PhD (2002) and DSc (2014) degrees in Theoretical Foundations of Computer Science from the Russian Academy of Sciences. His research field is AI and Machine Learning, Bayesian Model Selection, and Functional Data Analysis with applications to Brain-Computer Interfaces and biomedical signals. He is the chief editor of the journal "Machine Learning and Data Analysis." His educational Youtube channel "Machine Learning Phystech" holds 1200 hours of monthly watch time. In 2020 he received Yandex' Segalovich award for his impact on scientific community development in CIS countries.--> | ||
Latest revision as of 15:39, 13 August 2024
Contents
Model selection in neurocomputing
- Date: September 14th 2024
- Abstract: The talk focuses on the problem of dimensionality reduction in neurocomputing modeling. In a neurocomputing experiment, a participant responds to a stimulus with a limb motion. Three modalities are measured: EEG for brain activity, and IMU for limb motions. The goal is to predict limb motions using brain signals. A prediction model is selected through causal inference. We show how to create a state space, process cross-correlated signals, and select a model. We introduce a model that grabs relationships within and between the source and target data.
Generative machine learning models for scenario simulation
- Date: December 2nd 2023
- Abstract: This talk presents the fundamental principles of generative modeling within the context of time series applications. Given the presence of high variance and substantial covariance among time series, we represent them as phase trajectories, multimodal sets, and dynamic graphs. Our talk delves into three types of generative models: variational autoencoders, normalizing flows, and diffusion probabilistic models. The objective is to reconstruct the distribution of time series or their dynamic interrelations. The pivotal method to elucidate is Principal Component Analysis. It introduces autoencoder as an essential part of generative neural networks.
Machine learning Model selection for biomedical signals
- Date: December 2nd 2022
- Abstract: We model spatial-time series: audio-video streams and brain signals like EEC, ECoG, and IMU from wearable devices. The practical application is human motion analysis for health monitoring. We discuss a forecasting model to approximate phase trajectories. Since this kind of data is highly correlated, the model selection is a fruitful way to obtain a simple, stable, and accurate model. To optimize the model structure, we use a quadratic programming problem statement. To set the criterion of optimality, we use Bayesian inference.
About
- Speaker's short bio: Dr. Vadim Strijov is a researcher at m1p.org. He served as a professor at the Computing Center of the Russian Academy of Sciences, and the head of the Intelligent Systems Department. He obtained his D.Sc. in Physics and Mathematics with theses on Mathematical Modelling and Machine Learning. In 2020 he received the Yandex Segalovich Prize award for his significant impact on the scientific community development in the CIS countries. His research fields are AI, Machine Learning, and Functional Data Analysis.