Hanke Chen

Hanke Chen

How romantic it is to learn things together with my AI model on weekends — [2019/02/02]

Academic / AI Research

kaggle_discussion.png My motivation for learning AI is that it can be used in many fields. “It is only a tool,” I thought but later I realized that the theory of training neural networks can be used for human learning too. “The biggest benefit that AI brought to us is not only a strong algorithm but also the theory of learning.”

My interest in wanting to know the fundamental mechanisms behind AI drove me into Udacity’s Machine Learning online course. In 1.5 years, I graduated from Machine Learning (Basics) and Deep Learning (Advanced) Nanodegrees. Although the knowledge these courses provide is too basic compared to tricks used in Kaggle competitions by PHDs and graduate students, these courses introduced me to a group of alumni and dragged me into the AI research circle. Three years ago, I started with nothing, but now I compete with other AI researchers as a high school student.

Here are some projects I did as I gradually entered the field of AI.

Google Science Fair: Extracting Cellular Location of Human Proteins Using Deep Learning

Video: Report Presentation
—
Time: 2018
Achievements: Model Prediction Accuracy 63.07%

 » I implemented a Convolution Neural Network with Residue and Squeeze-Excitation layers classifier to locate given proteins of any type in a subcellular structure. After training the model using a series of techniques, it can locate thousands of proteins in 27 different human cell types into 28 subcellular locations, way significant than historical approaches. The model can classify 4,500 images per minute with an accuracy of 63.07%, surpassing human performance in accuracy (by 35%) and speed.

Read Paper

Kaggle: SIIM-ACR Pneumothorax Segmentation

SIIM-ACR Pneumothorax Segmentation
—
Time: 2019
Achievements: 84.15% AUC Score Bronze Medal Top7%

 » I I cooperated with two graduate students from SJTU University to develope an AI model to classify (and if present, segment) pneumothorax from a set of chest radiographic images, aiding in the early recognition of pneumothoraces and save lives. With our new proposed “three-stage method”, we are able to achieve +0.02 accuracy on the test dataset.

Competition Website

Predicting the Usage of Shared Bikes Using Neuron Networks (NN)

Image: Predicted Trend vs. Real Trend
—
Time: 2017
Achievements: Model Prediction Accuracy 80.3%

 » I used a simple feed forward Neural Network to predict the hourly usage amount of shared bikes, using features like the season, weather, and date.

See Project

Detecting Dogs Using Convolutional Neuron Networks (CNN)

Image: Digits and Faces generated by my code
—
Time: 2017
Achievements: Generated Human-recognizable faces Generated handwriting digits

 » In this project, I detected breeds of dogs using a Convolutional Neutron Network. I first loaded a human face dataset from scikit-learn and detected faces as a practice. After training the dataset using ResNet50 architecture, I found out that the pre-trained model could out-compete the performance of OpenCV library by about 10% accuracy. Then, I went a step further; predicting dog breeds using CNN and trained it from scratch. Although the accuracy of standard CNN is low (9.2% accuracy), with ResNet50 architecture, I was able to successfully classify 41.5% of the dogs into 133 correct categories, far better than my own performance.

See Project

Generating Faces Using Generative Adverserial Networks (GAN)

Image: Predicted Trend vs. Real Trend
—
Time: 2017
Achievements: Generating Human-recognizable Pictures

 » I generated handwritten digit pictures and faces using GAN on MNIST and the CelebA dataset. After data pre-processing and building the Generator and Discriminator, I trained the GAN and fine-tuned the parameters. The parameters(learning rates) are hard to tune, but after only one epoch, the network could output recognizable handwritings and faces completely generated by the computer, derived from random numbers.

See Project

Generating TV Scripts Using Recurrent Neural Networks (RNN)

Image: novel/TV script generated by my code
—
Time: 2017
Achievements: Generated Human-recognizable novel

 » Using a Recursive Neuron Network I generated TV Scripts for “The Simpsons”. I first tokenized each word in The Simpsons Dataset and used Tensorflow to train the embedded words in mini-batches. After 124 epochs, the model c generate scripts that actually made some sense.

See Project and Generated Scripts

Teaching Drones to Fly Using Reinforcement Learning (RL)

Image: trajectory of fully-automatic drone in simulation
—
Time: 2017
Achievements: Using RL to balance the drones Command the drone to go to a specific coordinate

 » In this project, I was able to make a virtual drone dynamically hover and go to a specific location on its own using Reinforcement Learning. Usually, people have to control four different motors to ensure that drones do not crash, but with this algorithm, you can tell a drone “x,y,z” where it should go, and it can go there automatically.

See Project

Kaggle: Histopathologic Cancer

Image: Kaggle Competition
—
Time: 2019
Achievements: 97.82% AUC score 1st Ranked on Public Leaderboard (Top 1%) 113rd Ranked on Private Leaderboard (Top 10%)

 » In this Kaggle Competition, I use neuron-network to detect cancer samples from histopathologic images with unbelieveable 98% accuracy, surpassing the accuracy of PatchCamelyon benchmark model in the paper B. S. Veeling, J. Linmans, J. Winkens, T. Cohen, M. Welling. "Rotation Equivariant CNNs for Digital Pathology". arXiv:1806.03962 (AUC 96.3%). Although people cheated using the method I discovered in the competition, I was still ranked TOP 10% with my hardwork.

See Technical Detail

Kaggle: IMet Collection

Image: Kaggle Competition
—
Time: 2019

 » In this Kaggle Competition, I used image classification techniques to automatically tag artworks for The Metropolitan Museum of Art.

Competition Website

Kaggle: Instant Gratification

Image: Kaggle Competition
—
Time: 2019
Achievements: 97.140% AUC score

 » In this Kaggle Competition, I used the Gaussian Mixture Model to build a predictive model to generalized numerical relationship in a USB stick that washed ashore in a bottle according to Kaggle.

Competition Website

Kaggle: Understanding Clouds from Satellite Images

Image: Kaggle Competition
—
Time: 2019
Achievements: 65.239% Dice score

 » In this Kaggle Competition, our team used voting ensembles of 3 different deep learning models to segment regions of cloud images into 4 categories.

Competition Website