SoftBank Corp. and Aizip, Inc.'s
Real-Time Fish Counting Application with
On-Device AI Wins Innovation Award at CES® 2025

On-device Fish Counting AI Application

SoftBank had been developing a cloud-based AI system designed for fish counting to enhance the operational efficiency of aquaculture operators. However, the system's reliance on a server-based architecture presented challenges, particularly in securing power and communication for real-time underwater analysis. To overcome these limitations, a power-efficient, low-memory machine learning model capable of running on edge devices, such as smartphones, was needed.

To solve these problems, an AI model that can do real-time analysis on edge devices was developed, using SoftBank's CG fish schooling simulation to automatically create a training dataset for Aizip's compact TinyML technology. TinyML is known for being low power and small scale, able to run on edge devices, and yet it has similar accuracy to large, server level neural networks. For this research, a TinyAI model was optimized for counting fish, and successfully counted fish underwater, in real-time, with a standard smartphone camera, with 95% accuracy. Using this technology, the possibilities of using AI applications in a marine environment and for improvements in efficiency for the aquaculture industry are clear.

QA Application with a Built-In SLM

To investigate the possibilities of TinyML, SoftBank collaborated with Aizip to develop an employee QA system. While Large Language Models (LLMs) have made it possible to answer general questions, they typically require powerful servers to operate. However, for internal systems that only need to address specialized domains, edge device operation is preferred for lower power consumption and enhanced information security when handling sensitive internal data. Shrinking language models to fit resource-constrained edge devices while maintaining cloud-level generation quality presents a significant challenge. Aizip has addressed this by developing SLMs that not only run efficiently on mobile devices but also deliver generation quality comparable to cloud-based models through fine-tuning on task-specific datasets. Furthermore, the SLM is integrated into a RAG system that includes a database of employee service documents. This setup enables rapid adaptation to evolving data without requiring the SLM to be retrained. The newly developed QA system can clarify user questions by allowing the model to interact with ambiguous queries, achieving an impressive 96% accuracy rate in our in-house QA system. Future research is expected to deliver further accuracy improvements.

Tiny Machine Learning (TinyML) overview diagram

SoftBank and Aizip succeeded in reducing power consumption and size, allowing real-time execution on a smartphone, without reducing accuracy.

[Note]

1. *
   Aizipline Toolbox: An AI design tool developed by Aizip

Results of the Demonstration Experiment

• Counting fish underwater

  Enables estimation of the number of fish obscured by other fish

• SLM using RAG

  QA system running on a smartphone (in Japanese)

[Note]

1. *
   Some answers have been processed for security reasons.

About Aizip

Founded in 2020, Silicon Valley-based Aizip is a leader in AI for IoT (AIoT) model design. With its groundbreaking neural network architecture and proprietary automated design tools, it has demonstrated a wide range of deep neural networks (DNNs) with superior performance.
https://aizip.ai/

TinyML Glossary

Deep learning is problematic in that it requires computer power, and challenges such as power, privacy and real-time response have been highlighted in recent years. TinyML is a method that aims to solve these issues by minimizing the size of AI models. It has advantages in terms of power saving, low cost, privacy protection and real-time response, and is attracting attention in AI implementation.