## Overview Furuno is a Japanese marine electronics company founded in 1948 when they developed the world's first fish finder. The company has since expanded into radar, GPS, and other marine technologies, becoming a global leader in marine electronic equipment. However, as acknowledged in the presentation, service-based offerings and AI capabilities have historically been a weak spot for the company. Over the past three years, Furuno has ventured into AI development, applying it to various marine use cases including real-time ship detection from camera images, farmed fish detection and size measurement for aquaculture, and most notably, sustainable fishing assistance. The presentation was given by Akino Kazian from Furuno's AI division, with additional Q&A support from a colleague named Yoshi who has been working closely on the project. This case study represents an interesting intersection of traditional maritime industry expertise and modern AI/ML techniques, with LLMs playing a specific role in capturing tacit fisherman knowledge. ## The Problem: Sustainable Fishing and Data Scarcity The core business problem addresses a global sustainability challenge: marine fish stocks are declining yearly, with overfishing being a prominent contributor. Specific issues include premature catching of juvenile fish and incidental catches despite fishing restrictions. The industry needs to move away from indiscriminatory catches toward accurate assessment of fish species and size before casting nets. The technical challenge is severe data scarcity. Building effective AI systems for image recognition typically requires hundreds of thousands of examples, but for many fish species, Furuno has only hundreds of data points. With over 300 species of fish caught in Japan alone, it's estimated that traditional data collection approaches would take more than a century to gather sufficient training data. The presentation highlighted the technological state of fishing vessels, which are described as "about 20 years behind land-based industries such as agriculture." Equipment on typical Japanese fishing vessels is not connected to the internet, and none are equipped with AI. Fishermen rely on interpreting fish finder images (showing colored blobs representing fish schools), their years of experience, and intuition to make catch decisions. Notably, the presenter admitted that not a single Furuno employee can decipher fish finder images to predict fish species—only the experienced fishermen who captured the images possess this ability. ## The Solution: Ensemble Models with Knowledge Capture Furuno's approach involves an ensemble of several models working together: **Image Recognition Model (Sonar/Fish Finder Images)**: This component processes the visual data from fish finders to help differentiate between the appearance of various fish schools. The "big red blob" on a fish finder display contains information about fish schools that experienced fishermen can interpret, and this model attempts to learn these visual patterns. **Classifier Model**: This uses structured data such as swimming depths and ultrasonic signals that indicate unique characteristics of different fish species. This represents more traditional sensor-based machine learning. **Knowledge Model**: This is the most innovative component and where LLMs come into play. The knowledge model captures the tacit expertise of individual fishermen—knowledge that has been accumulated over decades but was previously extremely difficult to translate into usable data format. Examples of this knowledge include seasonal patterns ("in October there are lots of sardines"), behavioral observations ("Pacific mackerel are the ones swimming near the surface at 10 meters"), and age indicators ("young fish swim at X depth"). The ensemble approach also employs intelligent segmentation strategies. Rather than trying to classify all 300+ species simultaneously, the system sets conditions for segmentation based on the fact that each fisherman has specific target species. By limiting models to only a few relevant species per fisherman, they can significantly reduce noise in training data and make the problem tractable with limited data. ## Role of LLMs in Knowledge Capture The presentation explicitly highlighted the emergence of LLMs as a breakthrough for their use case. Previously, while fishermen's knowledge could potentially be obtained through interviews or observation, translating this intuitive, experience-based knowledge into usable structured data presented a significant challenge. LLMs provide fishermen with "a more straightforward means to input their knowledge." This represents an interesting application of LLMs as an interface layer for knowledge extraction and structuring, rather than the LLM being the primary reasoning component. The tacit knowledge of fishermen—expressed in natural language about seasonal patterns, fish behaviors, local conditions, and experience-based heuristics—can now be more easily captured and incorporated into the AI system. The presentation mentioned plans to significantly broaden information sources from the current 2 to over 20 types of data, including diverse forms such as signal data, text, and voice. The integration of these multimodal inputs into a coherent model is a key technical challenge. The presenter referenced the SSM (Structured State Space Model) approach proposed by Mamba as a potential architecture for constructing models that enable each fisherman to become a domain expert using structured data. ## Deployment and Edge Computing Considerations The Q&A session revealed important details about edge deployment challenges. With fishing vessels operating in the middle of the ocean, far from land-based infrastructure, edge computing is essential. The presenter acknowledged the constraints of deploying models onto edge devices in this environment. The solution has evolved with the times—while the presentation showed older display equipment and fish finder models on vessels, modern fishing boats more commonly bring laptop PCs aboard that serve as edge devices. These more powerful edge devices can handle the AI applications that Furuno is developing. This represents a practical approach to edge ML deployment, leveraging existing consumer hardware rather than requiring specialized marine-grade AI hardware. Recent advancements in satellite communications have also enabled data collection from all onboard equipment, which was previously impractical. This infrastructure improvement is crucial for both training the models with real-world data and potentially for model updates or cloud connectivity when available. ## Business Model and Ownership Philosophy The monetization approach discussed involves an add-on subscription service to fishermen at approximately $300/month. This represents Furuno's transition from pure product sales to service-based revenue, an acknowledged weak spot in their historical business model. A particularly thoughtful aspect of the design philosophy is the ownership model. Rather than creating a centralized, black-box AI system, Furuno's approach makes each fisherman the owner of their personalized model. This design choice has several benefits: - Fishermen are more inclined to use AI when they feel ownership over it - They become more motivated to contribute knowledge and improve the system - Each fisherman effectively becomes an expert in their own right - The approach cultivates what the presenter called "the pleasure of nurturing AI"—fishermen become masters of their own models This decentralized, personalized approach also addresses the cost-efficiency challenge. Training a single model on all fish varieties would be enormously expensive, with a large portion of that cost consumed by learning data irrelevant to any individual fisherman's local fish populations. ## Results and Future Directions Early results are described as "very promising." For fixed-net fishing operations, the AI system has improved productivity by helping fishermen know exactly when fish arrive in the nets, allowing them to optimize their timing for collection. Furuno is thinking about 10+ additional applications for this AI technology internally, though these were not disclosed publicly at the time of the presentation. The company frames their work as a commitment to sustainable fishing practices that will help "continue to provide delicious fish to you, your children, and future generations." ## Critical Assessment While the case study presents an innovative approach to a genuine problem, several aspects warrant balanced consideration: - The system is still in relatively early stages, with the "journey towards safe and sustainable fishing" only recently embarked upon - Concrete metrics on sustainability improvements, accuracy rates, or adoption numbers were not provided - The $300/month subscription model's viability for fishermen, who may already operate on thin margins, was not validated with market data - The claims about LLMs enabling knowledge capture are conceptually sound but the specific implementation details and effectiveness remain unclear - The 10+ additional internal applications mentioned suggest the technology platform may have broader potential, but these are still in internal discussion Overall, this case study represents an interesting application of ensemble ML approaches combined with LLM-based knowledge capture in a traditionally underserved industry, with particular attention to edge deployment constraints and user-centric ownership models.