NEC to Build World-Leading Underwater Acoustic Foundation Model

Japan's NEC Corporation has been awarded a contract to research underwater acoustic foundation models using self-supervised learning, with dual-use applications in mind. Conducted for Japan's Agency for Defense Equipment, the project aims to train a model on vast amounts of underwater acoustic data to create a general-purpose generative AI for the field. NEC targets completion by fiscal year 2027, with applications spanning defence, marine life exploration, resource monitoring and earthquake prediction.

 

Details of the Contract

 

NEC has secured a significant research contract. It covers research on underwater acoustic foundation models. The work uses self-supervised learning techniques. It focuses on dual-use applications for the technology. The contract falls under a defence-related programme.

The project operates within a specific institutional framework. It is part of a programme by the Defense Innovation Science and Technology Institute. This sits within Japan's Agency for Defense Equipment. NEC has already commenced research on the project. It aims to complete the foundation model by fiscal year 2027.

 

The Project's Ambition

 

The project pursues a substantial technical goal. It involves training a model on vast underwater acoustic data. The aim is to build a world-leading foundational model. This would serve the field of underwater acoustics. It represents an ambitious undertaking in the field.

The concept draws on developments in artificial intelligence. The goal is a general-purpose generative AI for underwater acoustics. This would be similar to large language models. Such models have transformed language processing. The project seeks comparable capability for underwater sound.

 

Why Underwater Acoustics Matter

 

Sound is uniquely important beneath the ocean surface. Light and radio waves barely penetrate water. This makes sound waves an effective observation method. They enable a wide range of recognition tasks. This underpins the value of underwater acoustic analysis.

Current analysis methods face significant limitations. Sound is captured using sonar and hydrophones. Skilled experts analyse this data using signal processing. However, analysing vast data requires much time and concentration. There are also limits to the accuracy that can be achieved.

 

Read More: ZeroUSV Names First Long-Range Oceanus17 Hybrid USV in UK Port

 

The Technical Challenge

 

Building such a model presents distinct difficulties. Language foundation models are advancing worldwide. Yet few general-purpose underwater acoustic models exist. Development has largely focused on specialised AI models. These are limited to specific applications or tasks.

Data availability is a central obstacle. Conventional AI models require large amounts of labelled data. This makes them hard to apply where little labelled data exists. Collecting labelled data is especially challenging underwater. This has hindered building large-scale AI models in the field.

 

NEC's Approach

 

NEC brings substantial relevant experience to the project. It has over 90 years of experience in sonar technology. This provides a strong foundation for the work. It will develop a general-purpose underwater acoustics model. This builds on the company's established expertise.

The company will draw on its AI capabilities. It will leverage expertise from its proprietary AI technology. Self-supervised learning is central to the approach. This enables using large data volumes without labelled data. Training data will be collected through public-private partnerships.

 

Applications and Significance

 

The technology promises broad applications. It would enable more accurate understanding of underwater phenomena. It could also facilitate creating a digital twin of the ocean. These capabilities extend beyond defence purposes. They open possibilities across multiple fields.

The potential uses span science and safety. These include exploration of marine life and resources. Environmental monitoring is a further application. High-precision earthquake prediction is also anticipated. This breadth reflects the dual-use nature of the technology.