5:15 PM - 7:15 PM
[MGI27-P02] Weather field super-resolution using Restricted Boltzmann Machines
Keywords:Boltzmann Machine, Machine learning, Super Resolution, Precipitation
The super-resolution of weather fields has been approached in various ways in the field of deep learning. Among them, generative adversarial models and diffusion models are well-known approaches because of their highly accurate restoration of weather fields and their inference speeds. On the other hand, such models require a large number of learning parameters. While further improvements in the accuracy of these models are expected in the future, the improvements would accompany increase in number of parameters and necessary computational resources. Here, it is important to explore more efficient learning methods and the basis of the model structures to realize accurate restoration with less computational resources.
This study proposes using the Restricted Boltzmann Machine, one of the foundational models of deep learning models, to perform super-resolution of spatial distribution of precipitation. The Restricted Boltzmann Machine consists of only two layers: a visible layer and a hidden layer. Learnable weights fully connect both layers, and there are learnable biases on each layer, but there are no connections between nodes on the same layers. The discover of such structure allow us to train the model using contrastive divergence method, which is one of the Gibbs sampling methods. Owing to this structure of the model, the restricted Boltzmann machine has a much smaller number of parameters than current deep learning models have. By testing this model, we expect to find more efficient training method for current deep learning models.
As precipitation data, we used the Japan Meteorological Agency's Radar AMeDAS precipitation product and aimed to up-sample the spatial distribution of precipitation in the northern Kyusyu region, which is a 28x28 km squared region consisting of a plain and mountainous area. During the training process, the model learns to infer the inputted precipitation at the visible layer from the hidden layer. In the inference process, the input was the precipitation distribution at 4x4 times coarser resolution than the original one. We evaluated whether the model could restore the original distribution of precipitation. The experiments revealed that while issues such as blurring effects of the super-resolved distribution remain, there is a possibility of reproducing terrain features of precipitation from the coarsened precipitation distribution. Although we trained the model solely on precipitation data, the results suggested that the model has implicitly learned the characteristics of the terrain with only two layers.
This study proposes using the Restricted Boltzmann Machine, one of the foundational models of deep learning models, to perform super-resolution of spatial distribution of precipitation. The Restricted Boltzmann Machine consists of only two layers: a visible layer and a hidden layer. Learnable weights fully connect both layers, and there are learnable biases on each layer, but there are no connections between nodes on the same layers. The discover of such structure allow us to train the model using contrastive divergence method, which is one of the Gibbs sampling methods. Owing to this structure of the model, the restricted Boltzmann machine has a much smaller number of parameters than current deep learning models have. By testing this model, we expect to find more efficient training method for current deep learning models.
As precipitation data, we used the Japan Meteorological Agency's Radar AMeDAS precipitation product and aimed to up-sample the spatial distribution of precipitation in the northern Kyusyu region, which is a 28x28 km squared region consisting of a plain and mountainous area. During the training process, the model learns to infer the inputted precipitation at the visible layer from the hidden layer. In the inference process, the input was the precipitation distribution at 4x4 times coarser resolution than the original one. We evaluated whether the model could restore the original distribution of precipitation. The experiments revealed that while issues such as blurring effects of the super-resolved distribution remain, there is a possibility of reproducing terrain features of precipitation from the coarsened precipitation distribution. Although we trained the model solely on precipitation data, the results suggested that the model has implicitly learned the characteristics of the terrain with only two layers.