(Results and future prospects) In 2021, SimCLR [Chnetal., 2020], which achieved SOTA in self-supervised contrastive learning as of 2020, was implemented on the GPU machine of the Marine AI Development Evaluation Center and Google Colaboratory, and the performance was improved. verified.In a similar direction, in recent years, CLIP [Radfordetal., 2021] and CoCa [Yuetal., 2022] have been combined with language models to learn with super-large datasets of billions or more. Methods have been developed to do so with impressive results, but these studies are not possible without very large budgets and facilities.On the other hand, it has been reported that these methods do not perform well for images that do not contain similarities in the training data, such as microscopic images [Radford et al., 2021].In addition, autoencoders have been reviewed, such as being used in StableDiffusion [Rombachetal., 2021] to learn feature expressions, so we stopped developing a self-supervised contrastive learning method and used light and dark images in microscope images. In the task of conversion to fluorescent images and the detection of abnormalities in healthy fish diseases, relatively good results were obtained by learning with autoencoders such as U-net and VQ-VAE.As soon as the research results are finalized, we plan to publish them one by one.theme d.Construction of highly accurate nitrate profile data in the ocean (Fuminori Hashihama, Takehiro Nagai, Kohei Mizobata, Naho Miyazaki) Theme e.System development of self-supervised contrastive learning method as a basis for AI analysis of ocean big data (Tomoyuki Takenawa) A detailed understanding of the salinity distribution is the key.The purpose of this theme is to construct high-resolution and high-precision nitrate profile data by combining continuous vertical observation using a nitrate sensor and analysis values ??of samples collected using a clean water sampler. (Plan/Method) Collect nitrate concentration data in addition to regular water temperature, salinity, and pressure observations during training cruises on the Shioji Maru.By combining continuous vertical data from an underwater UV nitrate analyzer and analysis data of water samples at arbitrary depths using a Teflon-coated clean water sampler, the same vertical data as water temperature and salinity, which are the main variables of the ocean, can be obtained. Build high-precision nitrate profile data with resolution. (Results and Future Prospects) In August 2022, during the voyage of Shiojimaru (open ocean observation training), we succeeded in acquiring highly accurate nitrate profile data.At 8 stations in the western North Pacific, 6m-by-0m data from 2000-1m was obtained by an underwater ultraviolet nitrate analyzer, and at the same time nitrate samples were collected from 1 layers per station by clean water sampling and analyzed on board.The data from the underwater UV nitrate analyzer showed a high correlation (r17=2) with the nitrate concentration obtained from the analysis, and from that relational expression, we were able to estimate the nitrate concentration data for every 0.994m with the same vertical resolution as the water temperature and salinity.In the future, the Shioji Maru plans to conduct open ocean observation training every summer and winter, so it will be possible to obtain long-term, highly accurate nitrate profile data, including seasonal changes.Due to the warming and stratification of the surface layer of the ocean due to global warming, it is predicted that oligotrophic conditions and the weakening of biological production will progress. expected to be possible.The problem is that it is costly to assign teacher labels to many big data in the ocean field.Self-supervised controlled learning is a method of learning without using supervised data instead of using a large amount of data, and is achieving performance comparable to or exceeding that of supervised learning.We develop a system for self-supervised contrast learning with real data. (Plan/Method) Develop a system for self-supervised contrast learning with actual data. 1.Acquisition of marine big data and promotion of AI analysis research (XNUMX) Development of a fishery integrated support system that realizes resilient and sustainable fisheries using marine organism big data and training of marine AI human resources
Original page ../index.html#29