This paper presents great visual images and quantitative scores for an autoencoder-based generative model. All reviewers agree on this aspect, and this is primarily the reason why acceptance is warranted. Certainly an AE pipeline with this capability is a worthwhile contribution to the community. However, the proposed method is mostly some engineered enhancements to the basic VQ-VAE model that has already been published. Moreover, full architectural details and hyperparamter settings were not provided in the original submission but were promised for the final version.

In addition, the model also inherit the nice property of AE-based models that it does not suffer from the mode collapse issue. However, it seems to me that the only difference between this paper and the VQ-VAE paper is that this work introduces the hierarchical structure to learn different levels of latent representations and priors. The novelty looks a bit low. In addition, this paper didn't provide any idea about why such a design can make the generative performance better. The loss function (2) is not a reasonable objective to optimize considering the stop gradient operator. During the optimization procedure, the loss function may increase by taking a step in the gradient directions.

We explore the use of Vector Quantized Variational AutoEncoder (VQ-VAE) models for large scale image generation. To this end, we scale and enhance the autoregressive priors used in VQ-VAE to generate synthetic samples of much higher coherence and fidelity than possible before. We use simple feed-forward encoder and decoder networks, making our model an attractive candidate for applications where the encoding and/or decoding speed is critical. Additionally, VQ-VAE requires sampling an autoregressive model only in the compressed latent space, which is an order of magnitude faster than sampling in the pixel space, especially for large images. We demonstrate that a multi-scale hierarchical organization of VQ-VAE, augmented with powerful priors over the latent codes, is able to generate samples with quality that rivals that of state of the art Generative Adversarial Networks on multifaceted datasets such as ImageNet, while not suffering from GAN's known shortcomings such as mode collapse and lack of diversity.

Vector quantization (VQ) is a technique to deterministically learn features with discrete codebook representations. It is commonly performed with a variational autoencoding model, VQ-VAE, which can be further extended to hierarchical structures for making high-fidelity reconstructions. However, such hierarchical extensions of VQ-VAE often suffer from the codebook/layer collapse issue, where the codebook is not efficiently used to express the data, and hence degrades reconstruction accuracy. To mitigate this problem, we propose a novel unified framework to stochastically learn hierarchical discrete representation on the basis of the variational Bayes framework, called hierarchically quantized variational autoencoder (HQ-VAE). HQ-VAE naturally generalizes the hierarchical variants of VQ-VAE, such as VQ-VAE-2 and residual-quantized VAE (RQ-VAE), and provides them with a Bayesian training scheme. Our comprehensive experiments on image datasets show that HQ-VAE enhances codebook usage and improves reconstruction performance.

With development of deep learning, researchers have developed generative models in generating realistic images. One of such generative models, a PixelCNNs model with Vector Quantized Variational AutoEncoder 2 (VQ-VAE-2), can generate more various images than other models. However, a PixelCNNs model with VQ-VAE-2, I call it PC-VQ2, requires sufficiently much training data like other deep learning models. Its practical applications are often limited in domains where collecting sufficient data is not difficult. To solve the problem, researchers have recently proposed more data-efficient methods for training generative models with limited unlabeled data from scratch. However, no such methods in PC-VQ2s have been researched. This study provides the first step in this direction, considering generation of images using PC-VQ2s and limited unlabeled data. In this study, I propose a training strategy for training a PC-VQ2 with limited data from scratch, phased data augmentation. In the strategy, ranges of parameters of data augmentation is narrowed in phases through learning. Quantitative evaluation shows that the phased data augmentation enables the model with limited data to generate images competitive with the one with sufficient data in diversity and outperforming it in fidelity. The evaluation suggests that the proposed method should be useful for training a PC-VQ2 with limited data efficiently to generate various and natural images.

Unsupervised learning of discrete representations from continuous ones in neural networks (NNs) is the cornerstone of several applications today. Vector Quantisation (VQ) has become a popular method to achieve such representations, in particular in the context of generative models such as Variational Auto-Encoders (VAEs). For example, the exponential moving average-based VQ (EMA-VQ) algorithm is often used. Here we study an alternative VQ algorithm based on the learning rule of Kohonen Self-Organising Maps (KSOMs; 1982) of which EMA-VQ is a special case. In fact, KSOM is a classic VQ algorithm which is known to offer two potential benefits over the latter: empirically, KSOM is known to perform faster VQ, and discrete representations learned by KSOM form a topological structure on the grid whose nodes are the discrete symbols, resulting in an artificial version of the topographic map in the brain. We revisit these properties by using KSOM in VQ-VAEs for image processing. In particular, our experiments show that, while the speed-up compared to well-configured EMA-VQ is only observable at the beginning of training, KSOM is generally much more robust than EMA-VQ, e.g., w.r.t. the choice of initialisation schemes. Our code is public.

We explore the use of Vector Quantized Variational AutoEncoder (VQ-VAE) models for large scale image generation. To this end, we scale and enhance the autoregressive priors used in VQ-VAE to generate synthetic samples of much higher coherence and fidelity than possible before. We use simple feed-forward encoder and decoder networks, making our model an attractive candidate for applications where the encoding and/or decoding speed is critical. Additionally, VQ-VAE requires sampling an autoregressive model only in the compressed latent space, which is an order of magnitude faster than sampling in the pixel space, especially for large images. We demonstrate that a multi-scale hierarchical organization of VQ-VAE, augmented with powerful priors over the latent codes, is able to generate samples with quality that rivals that of state of the art Generative Adversarial Networks on multifaceted datasets such as ImageNet, while not suffering from GAN's known shortcomings such as mode collapse and lack of diversity.