2019-2 YBIGTA CONFERENCE

YBIGTA is the Big Data Academy in the college of Engineering, Yonsei University.

They study and practice a number of fields, such as Data Science, Machine Learning, AI, NLP, Computer Vision.

Although I’m not a member of this club, I registered to participate in their regular conference because I’m interested in.

This is the post about the contents I was able to get in the conference.

Obstacle warning system for the sidewalk

(by Jaehoon Kang and 4 others, Yonsei University)

Object Detection = Localization(bounding box) + Classification(detecting the category)
If an object exists too close, the bounding box becomes to disappear, which leads to the detecting problem. In addition, it considers an object as a threat even if it does not come to the user but just stays still. This is because the process is based on images and it cannot detect the depth.

Auto Machine Learning

(by Sang Hyun Baek and 4 others, Yonsei University)

Rule-based Auto ML

Meta Learning based Auto ML

AutoML is the automatic Machine Learning method to automate repetitive works. Especially, labeling and pre-processing are too tiresome tasks.
Rule based model is based on an algorithm that determines pre-processing methods or learning models according to some specific rules.
Meta learning based model uses the meta learning as its basis. Based on data’s meta features, it analyzes the meta features of new data and finds an optimal model by comparing cosine similarities. Meta features can be considered as the specific features of data itself, such as the extent of data’s bias.
Bayesian optimizer chooses optimal hyper parameters.

Prediction of course register and its web application

(by Joo Young Lee and 4 others, Yonsei University)

Description of XGBoost

Comparison between XGBoost and LightGBM

The pyMySQL is a MYSQL library for Python allowing several commands to be executed by Python codes.
Feature Engineering is a process of converting data’s features into acceptable shapes for training.
GridSearchCV is a sequence of processing for finding optimal hyperparameters.

Visual question answering

(by Joon Sung Park and 2 others, Yonsei University)

Multimodal Compact Bilinear Pooling is a process for reducting features by combining the information of image and text with outer product of vectors.

Reinforcement learning and game

(by Seung Yoo Kim and 5 others, Yonsei University)

Markov Decision Process

MCTS - Backpropagation

Actual playing

In QWOP game, they used a method that makes the athlete’s pose an image and gives it as a state. With this, the difference between previous state and current state becomes an input.
Actor-Critic structure consists of an actor which determines which action is conducted and an critic that evaluates values.
The algorighm used in Alpha Go is the Monte Carlo tree search, MCTS.

Prediction of MBTI results with deep learning

(by Ye Hee Go and 5 others, Yonsei University)

The most commonly used morphemes by the people in each characteristic category are classified and used as features.
BERT’s weakness is that pre-processing is a little bit tricky. This is because we need token embeddings, position embeddings and segment embeddings.
They stacked an additional linear layer at the end of KoBERT and conducted fine-tuning.

Text2Moji

(by Jin Woo Lee and 4 others, Yonsei University)

It is a system that recommends an emoji based on the context when a user types texts.

Real Time Face Swapping

(by Joon Hyeong Kim and 3 others, Yonsei University)

An example of Deep Fake

Real-time demonstration

One of the possible ways is extracting landmarks from each face and conducting segmentation process into original images.
The second is a method that extract landmarks and combines them with interpolation.
the third one uses few-shot learning, which is a method to train with small number of samples.
Meta learning is a training for finding a way to get a number of labels.

A model of handwriting font generation

(by Joon Hyeong Kim and 2 others, Yonsei University)

The encoder and the decoder exist in the generator. The encoder finds the feature of given font, applys it and gives the result to the decoder.
L1 loss is good at getting the shape of texts, but it leads to slow down of training after certain extent. So they fixed some details with Constant loss.

Although this conference is held by candidates of bachelor’s degree, but it was very useful for me.

They tried a lot of creative and professional methods to improve their results and I think these are comparable to graduate-level researches.

I hope they keep working hard.