조한주 Email & Phone Number

대한민국 서울

· Multi-task learning for application-level optimization (2020. 06 – present) - Optimize scene understanding application, consisting of object detection, semantic segmentation, and depth estimation, by building a multi-task model sharing feature extractor  Developed a novel multi-task learning process that applies single-task models to balance the training progress of various tasks of different nature and to produce pseudo labels for label-absent tasks of partially annotated images  Achieved x3.47 speedup (from 73.3 ms to 21.1 ms) with 102.65%, 94.47%, and 98.97% accuracy in object detection, semantic segmentation, and depth estimation, respectively, when compared to single-task models· Optimizing vision neural networks for the specific NPU (2019. 11. – present) - Investigate inefficient operations of neural networks for specific NPUs and optimize them· Quantization-aware training to improve accuracy of quantized neural networks (2019. 11 – present) - Apply quantization-aware training (QAT) to quantized neural networks if the accuracy was seriously degraded by post-training quantization - Implement a specialized QAT platform for the target NPU by adopting FakeQuantize which imitates the target NPU’s quantization during the training· Accuracy validation of vision neural networks for various compression methods (2018. 11 – 2019. 10) - Validated the accuracy and compression ratio of ternary quantization for RefineDet compared to uniform quantization· Research on the trend of vision neural networks for designing NPU ISA (2017. 12 – 2018. 10) - Researched the trend of vision neural networks to decide which types of layers and activations the NPU should support in the early stage of NPU architecture design - Analyzed required networks' structure and MAC counts to support the hardware team

박사, Electrical And Electronics Engineering

학사, 전자전기공학