기업명 비공개

[Problem-solving skills: Designing an approximate multiplier] When I was an undergraduate student, I had experience publishing a research paper on the approximate multiplier. Since this applies to convolution operations, we aimed at precise multipliers with minimal error deflection and relative error. Traditional approximate converters had a problem with error deflection of more than 3%, but we determined that the reason was to use only one type of compressor. To solve this problem, we used two types of compressors that have the same error case on the truth table but are deflected differently by + and -, respectively. To take advantage of this, we have applied the following two rules to the topology configuration:

When the probability is the same, set the error to occur on the side with the higher multiplication expectation. By applying the above rules, the error indicators of multipliers and other multipliers with lower relative errors were simulated and compared using C++, and HW area and power consumption were tested using the Verilog code written. As a result, we derived 0.02% error bias and 0.9% relative error with more than 26% lower power consumption and area than the correct multiplier, and learned that optimizing the design according to the application can achieve high performance at a low cost.

[Challenge: Achieved 99.66% recognition of special characters] I'm always positive that I can do it. Thanks to this personality, my strength is not to be afraid of challenges and to achieve my goals. With these advantages, I have gained recognition rates that exceed initial goals in a comprehensive design project. The design theme was handwriting recognition using machine learning, and the goal was to achieve a recognition rate of more than 99% based on ultra-light neural networks in line with the flow of AI small processors. This was a big challenge because the similarity between letters was greater than that of MNIST and the number of labels doubled. In particular, the main challenge was how to create tens of thousands of datasets. I thought there was a limit to creating and photographing tens of thousands of datasets, so I created Python code that automatically corrects and recognizes characters using OpenCV and stores them as datasets. In addition, we decided that the data aggregation technique was suitable for a special character dataset that can be recognized as the same character even with a slight rotation, and applied a method of rotating left and right. As a result, 149,000 datasets per character were constructed to derive 99.66% recognition rate with neural networks with only 5 layers. [Focusing on their work by establishing a unit plans] My weakness is that if I focus on one thing, I lose concentration on the other. The first semester of the fourth grade left disappointing results because I couldn't concentrate on my studies even though I applied for 21 credits because I was busy writing research papers and comprehensive design projects. When I had time, I focused on the design and experiments needed to write a thesis, so I always prepared for the major examination as if I were pressed for time. As a result, it became the only semester out of four regular semesters and five seasonal semesters that did not exceed the average A after returning to school. Recognizing these problems, I am trying to compensate for my shortcomings by making it a habit to divide the project into detailed units and allocate a period of time.