Find Criminal Faces 10 Times Faster in CCTV Videos

KAIST Research Team Designs AI Chip and Completes Unit Production

Development of a New AI Acceleration Chip for Precise Visual Reasoning
- Research team from the Department of Electrical Engineering: Professors Yoon Chan-hyun and Kim Ju-young, (from left) Master’s students Kim Jun-su, Kim Ju-young (Professor), Ko Geon-woo, PhD candidates Jeon Min-su, Kim Sung-hwan, Professor Yoon Chan-hyun, PhD candidates Lee Chang-ha, Kim Tae-woo. At the bottom left, (left) KAIST prototype board for explainable AI processing and (right) newly developed EPU chip.

[Asia Economy Reporter Kim Bong-su] An artificial intelligence algorithm has been developed that can identify specific objects such as enemy tanks and airplanes in satellite images or closed-circuit (CC) TV footage, or suspects and wanted vehicles, with six times greater precision and ten times faster than before, while also explaining the reasons behind the identification.

The Korea Advanced Institute of Science and Technology (KAIST) announced on the 24th that a research team led by Professors Yoon Chan-hyun and Kim Ju-young from the Department of Electrical Engineering has succeeded in designing an AI chip equipped with a noise-resistant multi-pyramid activation map-based attention structure to process explainable artificial intelligence (eXplainable AI, XAI) techniques. With support from Samsung Electronics’ DS Division, the team has also completed the development of an Explainable Neuro-Processing Unit (EPU).

Explainable AI refers to AI techniques that provide explanations understandable and trustworthy to humans. Traditional AI, trained through mathematical algorithms, suffers from biases in training examples or the problem that humans cannot comprehend billions of parameters. To address this, explainable AI was developed to provide reasoning for why AI inferred specific results. Explainable AI guarantees higher accuracy, fairness, and reliability than conventional AI by explaining the significant factors influencing AI decision-making.

The research team implemented an AI model capable of complex interpretation of the activation levels of neural layers inside the AI that affect inference results within a pyramid-shaped neural network structure designed for multi-scale and multi-object feature extraction. They also realized an EPU chip specialized for accelerated processing using channel-direction convolution operations while maintaining accuracy. To visualize explanations in the pyramid-shaped AI model specialized for multi-scale and multi-object feature extraction, a structure is required that can extract parameter gradients from activation maps of all convolutional layers in the reverse direction of the inference process.

However, unlike conventional AI chip designs for accelerating inference processing, the backpropagation calculation process requires remembering previous parameters and states, which poses physical limitations for implementation using limited on-chip memory size and custom ASIC (Application Specific Integrated Circuit) tailored for specific purposes. Additionally, the pyramid-shaped explainable AI model requires gradient-based class activation mapping visualization processing for activation maps of N layers to guarantee explainability, increasing complexity. Furthermore, even very small input noise could completely alter class activation mapping visualization explanations, significantly reducing the reliability of explainable AI models.

To solve these problems, the research team developed a network structure that fuses the unique characteristic information of multiple activation maps of explainable AI to generate a global attention map, and a co-learning method for creating models robust to input image noise. This improved explainability metrics by up to six times compared to single activation map-based attention map generation technology. By precisely reconstructing the complementary characteristics of various attention maps at multiple scales into a unified attention map, it became possible to provide explanations with precision interpretable by humans. The team explained that this research achievement is expected to greatly enhance the explainability of AI models in image analysis with large object size variations, such as satellite imagery.

To accelerate the proposed explainable AI model, the research team developed an XAI core capable of processing not only inference and backpropagation but also activation map generation. They proposed a multi-data flow method that flexibly partitions various computational tasks for simultaneous processing. Utilizing the characteristics of activation maps containing many zero values, they proposed a new data compression format that can skip consecutive zeros and developed an acceleration unit supporting this format, enabling processing of over ten times more activation maps within the chip.

The EPU chip developed by the research team can be applied to special-purpose and high-precision AI image processing systems such as optical satellites and all-weather observation Synthetic Aperture Radar satellites. It is expected to dramatically improve the explainability of AI system decision bases with low latency and low power consumption. The research team plans to continue follow-up research on the EPU chip development.

This content was produced with the assistance of AI translation services.