Embedded Files
Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring
Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring
Identify whether the sound emitted from a target machine is normal or anomalous(faulty). Experimented with ResNet, WaveNet, RotNet, and semi-supervised learning techniques to analyze toy-car valve sounds using spectrogram images. Implemented techniques like mixup, temperature scaling, label smoothing, and learning rate scheduling to optimize performance. Obtained ROC-AUC score of 0.9363.
Recommendation Systems - The Netflix challenge
Recommendation Systems - The Netflix challenge
Paticipated in a recommendation system competition where the goal was competing against some modern recommendation system architectures. Developed a custom collaborative filtering technique to get an RMSE of 0.882.
Using Web Data to Reveal 22-Year History of Sneaker Designs
Using Web Data to Reveal 22-Year History of Sneaker Designs
The study analyzed 23,492 sneaker images and metadata from StockX.com spanning 22 years (1999-2020) to investigate cultural trends in sneaker design. A sneaker design index was introduced using a contrastive learning method, revealing a shift in sneaker designs trends across the years. The data also highlighted how leading brands maintain distinct design identities, and provided insights into which sneaker models might command higher resale prices. This paper was one of the best paper candidates in WWW'22.
Molecular Activity Prediction
Molecular Activity Prediction
From a very small amount of training data (3000 compound names and labels), train a set of networks to predict the probability of a new unseen compound being active or inactive for certain bioactivity. The task includes understanding the domain, appropriate feature engineering, and building model(s). Morgan fingerprints, Avalon fingerprints, molecular descriptors were used as features to train Light Gradient Boosting and XGBoost, the results of which were combined for the final prediction. The final result was an AUC-ROC score of 93% and a log loss of 0.3 Please email me for details and code.
AI in ART competition - National Science Museum of Korea
AI in ART competition - National Science Museum of Korea
Using results from our research paper titled "Using Web Data to Reveal 22-Year History of Sneaker Designs", I tried to visualize the embedding space of 23492 sneakers from the last 20 years.
The artwork titled "the sneakers universe" won 3rd place in a competition held by the National Science Museum of Korea. KAIST School of Computing also covered the work.
Social Network Analysis
Social Network Analysis
Using a real dataset from an social networking site, utilized pyspark RDD programming to find potential friends on social media network. Designed an program that could run in parallel on different computing nodes. Please email me for details or code.
Page updated
Google Sites
Report abuse