-
The paper presents a sequential VAE with each cell as an LSTM, with the decoder incrementally adding the outputs with a attention update mechanism for restricting the input area by the encoder and the output update area by the decoder.
-
The model is tested on MNIST ,SVHN and CIFAR0-10 dataset.
- Previously proposed ideas are based either on VAEs or on generative modelling approach which generate images in a single pass.
- The model could have been tested against the performance of GANs and other generative models available. Also, the model when tested against the CIFAR-10 dataset doesnot seem to perform upto mark. These could be some of the potential areas for future work.