A method for learning disentangled representations of videos, the method comprising: feeding each frame of video data into an encoder to produce a sequence of visual features; passing the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; applying Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger representa-tion disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approximate a conditional distribution to facilitate minimization of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenating the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A non-transitory computer-readable storage medium comprising a computer-readable program for learning dis-entangled representations of videos, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of: feeding each frame of video data into an encoder to produce a sequence of visual features; passing the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; applying Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for 10 a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger representa-tion disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approximate a conditional distribution to facilitate minimization of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenating the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A system for learning disentangled representations of videos, the system comprising: a memory; and one or more processors in communication with the memory configured to: feed each frame of video data into an encoder to produce a sequence of visual features; pass the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; apply Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger represen-tation disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approxi-mate a conditional distribution to facilitate minimi-zation of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenate the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A method for learning disentangled representations of videos, the method comprising: feeding each frame of video data into an encoder to produce a sequence of visual features; passing the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; applying Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger representa-tion disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approximate a conditional distribution to facilitate minimization of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenating the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A non-transitory computer-readable storage medium comprising a computer-readable program for learning dis-entangled representations of videos, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of: feeding each frame of video data into an encoder to produce a sequence of visual features; passing the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; applying Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for 10 a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger representa-tion disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approximate a conditional distribution to facilitate minimization of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenating the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A system for learning disentangled representations of videos, the system comprising: a memory; and one or more processors in communication with the memory configured to: feed each frame of video data into an encoder to produce a sequence of visual features; pass the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; apply Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger represen-tation disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approxi-mate a conditional distribution to facilitate minimi-zation of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenate the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A method for learning disentangled representations of videos, the method comprising: feeding each frame of video data into an encoder to produce a sequence of visual features; passing the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; applying Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger representa-tion disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approximate a conditional distribution to facilitate minimization of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenating the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A non-transitory computer-readable storage medium comprising a computer-readable program for learning dis-entangled representations of videos, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of: feeding each frame of video data into an encoder to produce a sequence of visual features; passing the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; applying Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for 10 a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger representa-tion disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approximate a conditional distribution to facilitate minimization of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenating the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A system for learning disentangled representations of videos, the system comprising: a memory; and one or more processors in communication with the memory configured to: feed each frame of video data into an encoder to produce a sequence of visual features; pass the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; apply Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger represen-tation disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approxi-mate a conditional distribution to facilitate minimi-zation of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenate the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A method for learning disentangled representations of videos, the method comprising: feeding each frame of video data into an encoder to produce a sequence of visual features; passing the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; applying Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger representa-tion disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approximate a conditional distribution to facilitate minimization of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenating the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A non-transitory computer-readable storage medium comprising a computer-readable program for learning dis-entangled representations of videos, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of: feeding each frame of video data into an encoder to produce a sequence of visual features; passing the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; applying Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for 10 a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger representa-tion disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approximate a conditional distribution to facilitate minimization of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenating the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.

A system for learning disentangled representations of videos, the system comprising: a memory; and one or more processors in communication with the memory configured to: feed each frame of video data into an encoder to produce a sequence of visual features; pass the sequence of visual features through a deep convolutional network to obtain a posterior of a dynamic latent variable and a posterior of a static latent variable; apply Jensen-Shannon divergence for a penalty of static embedding and maximum mean discrepancy for a penalty of dynamic embeddings, and simultaneously minimizing a Wassertein distance and a Kullback-Liebler divergence regularization to trigger represen-tation disentanglement; sampling static and dynamic representations from the posterior of the static latent variable and the posterior of the dynamic latent variable, respectively, with a general sample-based mutual information upper bound employed with a neural network to approxi-mate a conditional distribution to facilitate minimi-zation of a disentangled representation learning (DRL) objective and to reduce dependency of static and dynamic embeddings; and concatenate the static and dynamic representations to be fed into a decoder to generate reconstructed sequences.