Stony Brook University SUNY Korea

Two Computer Science students and Professor Youngmin Kwon in one team developed a data processing technology concerning training AI applications with image classification. This new technology is unlike the AI models in general which need completed decoded images for classification, and has been registered at the Korean Intellectual Property Office.

Saving time and memory space, this technology is even expected to save peoples’ lives if it assists the safety device as an auxiliary method.

Below is the team member Yousun’s written interview!

1. Please introduce yourself and your team members.

Hi. I am Yousun, a Ph.D. student in the CS department of the IoT lab. The two members of my team are my advisor, Prof. Youngmin Kwon, and another Ph.D. student, Jay, who graduated last Feb.

2. Could you explain briefly about this newly developed data processing that has been registered at the Korean Intellectual Property Office?

There is a study field of image classification in AI applications. The idea we developed is training an AI model with not fully decoded images. In general, pictures for AI training are completely decoded, but it is also likely to be trained as a deep learning model if we feed intermediate data taken from the decoding process.

The word ‘frequency’ is also used not just for radio waves, but pictures too. Image pixels are arranged in frequency ascending order while it’s in the encoding and decoding stage, and we take intermediate data from this stage.

Then the high-frequency parts are discarded because it is enough to train a model with the remaining parts if training targets are lumps of features. To do so, training data can be reduced up to one-fourth from the original size. It helps to save training time as well as memory space requirements.

3. How do you think this new technology will help people in general?

It requires fewer parameters of an AI model than using general pictures, therefore, the classification time is also fast with the same hardware. That small time gap is important to prevent the car accident of autonomous vehicles moving at high speed. If it assists the safety device as an auxiliary method, it will be helpful to save lots of lives.

4. Tell us about some difficulties or challenges your team faced while working together.

The first challenge was to design and implement an image decoder that can freely handle all sub-processes for decoding. We read lots of books about image processing and lecture notes.

Another difficulty was finding the best model for learning our outputs and repeated parameter tunings for each model. This is because it takes a long time to train the model, even if the input data has a small data size.

5. What did you learn from this experience?

One big lesson that I felt, and I would like to say to other students, is ‘You have to believe in what you think with confidence’. It would have never been invented if we ignored it as a trivial idea, and we would have given up without doing our best.