Professor Choi Jaesek's KAIST Team Develops Error Correction Technology for Deep Learning Generative Models
[Asia Economy Reporter Kim Bong-su] The Korea Advanced Institute of Science and Technology (KAIST) announced on the 25th that Professor Choi Jae-sik's research team at the AI Graduate School has developed an error correction technology for deep learning generative models.
Recently, deep generative models are widely used to create new content such as images, speech, and sentences. Despite the advancements in these generative models, even the latest models often produce flawed results, making it difficult to utilize generative models in critical tasks and learning areas such as defense, healthcare, and manufacturing.
The research team devised an algorithm that identifies and removes units (neurons) causing problems during the image generation process inside the generative model by utilizing explainable AI techniques that interpret the inner workings of deep learning. This error correction technology for generative models does not require retraining of the neural network model and has low dependency on the model structure. It can be broadly applied to various adversarial generative neural networks and improves the reliability of deep learning generative models.
The research team stated, "We demonstrated that visual errors in the outputs generated by deep learning generative models can be repaired by sequentially removing units inside the generative model that show corresponding activations," adding, "This result shows that even sufficiently trained models contain untrained or incorrectly trained internal components."
The research findings were presented on the 23rd at the International Conference on Computer Vision and Pattern Recognition (CVPR).
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
