Artificial intelligence (AI) is a field of research that artificially implements human intelligence. Hyundai KEFICO is conducting on-device AI research based on an optimal AI model for AI technology application.

Hyundai KEFICO is interested in the technology of optimizing the AI model in line with the specifications of the target device, and aims to use minimum device resources while maintaining AI inference performance. AI models customized and optimized to devices can be generated by using this technology. Hyundai KEFICO plans to apply AI technology to various application systems such as machine vision, mobility, etc. going forward.

AI model compression is a technology that minimizes the model's computation and size while maintaining the model's inference performance. Research on various compression techniques, such as pruning, quantization, distillation, etc. is being conducted. Pruning is a technique that eliminates the parts with relatively low importance from a model, while quantization is a technique of converting the scope of parameter changes within a model into a simple form. Distillation is a technique that generalizes knowledge of a large model to transmit it to small models. Hyundai KEFICO is implementing an optimal AI model on target devices based on such compression techniques.

The on-device AI platform infers AI models on the device itself. In conjunction with the generation of an optimal AI model, on-device AI is completed through development on the target device itself. Based on Hyundai KEFICO's powertrain controller development technology, tailored response to each customer can be provided by selecting the processor based on the performance required for AI computation, configuring communication and various input/output circuits.

3D flow estimation is a technique for estimating the future location of three-dimensional data such as point cloud data (PCD). Hyundai KEFICO is conducting research on AI models and compression techniques that estimate future locations by receiving three-dimensional data measured by a lidar or depth camera as input, which are the main sensors for autonomous driving. This technique can be extended to motion detection and tracking of specific objects, and can be applied to the safety field and autonomous driving mobility.