 |
Assistant HES, PhD student CS UNIGE frantzeska.lavda@hesge.ch |
Bio
Frantzeska joined the DMML Group as research assistant in February 2017. Before she studied Applied Mathematics in the National Technical University of Athens, Greece and she continued her studies at the University of Geneva where she completed her Master’s in Statistics.
Research
Her main research interests include deep generative modelling, deep learning, and Bayesian inference, with applications to image processing.
Publications
2020 |
Lavda, Frantzeska; Gregorova, Magda; Kalousis, Alexandros Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data Journal Article Entropy2020, 22(8) (888), 2020. @article{Lavda2020entropy, title = {Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data}, author = {Frantzeska Lavda and Magda Gregorova and Alexandros Kalousis}, url = {https://www.mdpi.com/1099-4300/22/8/888 https://bitbucket.org/dmmlgeneva/cp-vae/src/master/}, doi = {https://doi.org/10.3390/e22080888}, year = {2020}, date = {2020-08-13}, journal = {Entropy2020}, volume = {22(8)}, number = {888}, abstract = {One of the major shortcomings of variational autoencoders is the inability to produce generations from the individual modalities of data originating from mixture distributions. This is primarily due to the use of a simple isotropic Gaussian as the prior for the latent code in the ancestral sampling procedure for data generations. In this paper, we propose a novel formulation of variational autoencoders, conditional prior VAE (CP-VAE), with a two-level generative process for the observed data where continuous z and a discrete c variables are introduced in addition to the observed variables x. By learning data-dependent conditional priors, the new variational objective naturally encourages a better match between the posterior and prior conditionals, and the learning of the latent categories encoding the major source of variation of the original data in an unsupervised manner. Through sampling continuous latent code from the data-dependent conditional priors, we are able to generate new samples from the individual mixture components corresponding, to the multimodal structure over the original data. Moreover, we unify and analyse our objective under different independence assumptions for the joint distribution of the continuous and discrete latent variables. We provide an empirical evaluation on one synthetic dataset and three image datasets, FashionMNIST, MNIST, and Omniglot, illustrating the generative performance of our new model comparing to multiple baselines.}, keywords = {}, pubstate = {published}, tppubtype = {article} } One of the major shortcomings of variational autoencoders is the inability to produce generations from the individual modalities of data originating from mixture distributions. This is primarily due to the use of a simple isotropic Gaussian as the prior for the latent code in the ancestral sampling procedure for data generations. In this paper, we propose a novel formulation of variational autoencoders, conditional prior VAE (CP-VAE), with a two-level generative process for the observed data where continuous z and a discrete c variables are introduced in addition to the observed variables x. By learning data-dependent conditional priors, the new variational objective naturally encourages a better match between the posterior and prior conditionals, and the learning of the latent categories encoding the major source of variation of the original data in an unsupervised manner. Through sampling continuous latent code from the data-dependent conditional priors, we are able to generate new samples from the individual mixture components corresponding, to the multimodal structure over the original data. Moreover, we unify and analyse our objective under different independence assumptions for the joint distribution of the continuous and discrete latent variables. We provide an empirical evaluation on one synthetic dataset and three image datasets, FashionMNIST, MNIST, and Omniglot, illustrating the generative performance of our new model comparing to multiple baselines. |
Lavda, Frantzeska; Gregorová, Magda; Kalousis, Alexandros Improving VAE generations of multimodal data through data-dependent conditional priors Conference 24th European Conference on Artificial Intelligence, 325 , IOS Press, 2020. @conference{Lavda2020ecai, title = {Improving VAE generations of multimodal data through data-dependent conditional priors}, author = {Frantzeska Lavda and Magda Gregorová and Alexandros Kalousis}, url = {http://ebooks.iospress.nl/volumearticle/55021 https://bitbucket.org/dmmlgeneva/cp-vae/src/master/}, doi = {10.3233/FAIA200226}, year = {2020}, date = {2020-08-01}, booktitle = {24th European Conference on Artificial Intelligence}, journal = {IOS press}, volume = {325}, pages = {1254-1261}, publisher = {IOS Press}, abstract = {One of the major shortcomings of variational autoencoders is the inability to produce generations from the individual modalities of data originating from mixture distributions. This is primarily due to the use of a simple isotropic Gaussian as the prior for the latent code in the ancestral sampling procedure for the data generations. We propose a novel formulation of variational autoencoders, conditional prior VAE (CP-VAE), which learns to differentiate between the individual mixture components and therefore allows for generations from the distributional data clusters. We assume a two-level generative process with a continuous (Gaussian) latent variable sampled conditionally on a discrete (categorical) latent component. The new variational objective naturally couples the learning of the posterior and prior conditionals, and the learning of the latent categories encoding the multimodality of the original data in an unsupervised manner. The data-dependent conditional priors are then used to sample the continuous latent code when generating new samples from the individual mixture components corresponding to the multimodal structure of the original data. Our experimental results illustrate the generative performance of our new model comparing to multiple baselines.}, keywords = {}, pubstate = {published}, tppubtype = {conference} } One of the major shortcomings of variational autoencoders is the inability to produce generations from the individual modalities of data originating from mixture distributions. This is primarily due to the use of a simple isotropic Gaussian as the prior for the latent code in the ancestral sampling procedure for the data generations. We propose a novel formulation of variational autoencoders, conditional prior VAE (CP-VAE), which learns to differentiate between the individual mixture components and therefore allows for generations from the distributional data clusters. We assume a two-level generative process with a continuous (Gaussian) latent variable sampled conditionally on a discrete (categorical) latent component. The new variational objective naturally couples the learning of the posterior and prior conditionals, and the learning of the latent categories encoding the multimodality of the original data in an unsupervised manner. The data-dependent conditional priors are then used to sample the continuous latent code when generating new samples from the individual mixture components corresponding to the multimodal structure of the original data. Our experimental results illustrate the generative performance of our new model comparing to multiple baselines. |
2018 |
Ramapuram, Jason; Lavda, Frantzeska; Webb, Russ; Kalousis, Alexandros; Diephuis, Maurits Variational Saccading: Efficient Inference for Large Resolution Images Conference BMVC 2019 & Bayesian Deep Learning Workshop Neurips, 2018, (Code: https://github.com/jramapuram/variational_saccading). @conference{Ramapuram2018, title = {Variational Saccading: Efficient Inference for Large Resolution Images}, author = {Jason Ramapuram and Frantzeska Lavda and Russ Webb and Alexandros Kalousis and Maurits Diephuis}, url = {https://arxiv.org/abs/1812.03170 https://github.com/jramapuram/variational_saccading}, year = {2018}, date = {2018-12-03}, booktitle = {BMVC 2019 & Bayesian Deep Learning Workshop Neurips}, journal = {Bayesian Deep Learning Workshop NeurIPS 2018}, abstract = {Image classification with deep neural networks is typically restricted to images of small dimensionality such as 224x244 in Resnet models. This limitation excludes the 4000x3000 dimensional images that are taken by modern smartphone cameras and smart devices. In this work, we aim to mitigate the prohibitive inferential and memory costs of operating in such large dimensional spaces. To sample from the high-resolution original input distribution, we propose using a smaller proxy distribution to learn the co-ordinates that correspond to regions of interest in the high-dimensional space. We introduce a new principled variational lower bound that captures the relationship of the proxy distribution's posterior and the original image's co-ordinate space in a way that maximizes the conditional classification likelihood. We empirically demonstrate on one synthetic benchmark and one real world large resolution DSLR camera image dataset that our method produces comparable results with 10x faster inference and lower memory consumption than a model that utilizes the entire original input distribution.}, note = {Code: https://github.com/jramapuram/variational_saccading}, keywords = {}, pubstate = {published}, tppubtype = {conference} } Image classification with deep neural networks is typically restricted to images of small dimensionality such as 224x244 in Resnet models. This limitation excludes the 4000x3000 dimensional images that are taken by modern smartphone cameras and smart devices. In this work, we aim to mitigate the prohibitive inferential and memory costs of operating in such large dimensional spaces. To sample from the high-resolution original input distribution, we propose using a smaller proxy distribution to learn the co-ordinates that correspond to regions of interest in the high-dimensional space. We introduce a new principled variational lower bound that captures the relationship of the proxy distribution's posterior and the original image's co-ordinate space in a way that maximizes the conditional classification likelihood. We empirically demonstrate on one synthetic benchmark and one real world large resolution DSLR camera image dataset that our method produces comparable results with 10x faster inference and lower memory consumption than a model that utilizes the entire original input distribution. |
Lavda, Frantzeska; Ramapuram, Jason; Gregorova, Magda; Kalousis, Alexandros Continual Classification Learning Using Generative Models Workshop Continual learning Workshop NeurIPS 2018, 2018. @workshop{DBLP:journals/corr/abs-1810-10612, title = {Continual Classification Learning Using Generative Models}, author = {Frantzeska Lavda and Jason Ramapuram and Magda Gregorova and Alexandros Kalousis}, url = {http://arxiv.org/abs/1810.10612}, year = {2018}, date = {2018-10-24}, booktitle = {Continual learning Workshop NeurIPS 2018}, journal = {CoRR}, keywords = {}, pubstate = {published}, tppubtype = {workshop} } |