This project was done to participate in the YouTube-8M Video Understanding Competition on Kaggle.

Work

Implemented ML-kNN that supports tuning k (passing a list of k), File ml-knn.py
Implemented k-means that supports removing empty or small clusters during each iteration and supports Cosine and Euclidean distance, File kmeans.py
Implemented linear regression that does not require data to be shifted and supports tuning l2 regularization (passing a list of l2 reg. rates), Class LinearClassifier in linear_model.py
Implemented one-vs-rest logistic regression that supports class weights and instance weights. More importantly, it supports any feature transformation implemented in Tensorflow, Class LogisticRegression in linear_model.py
Implemented stable standard scale transformation that supports any feature transformation, function compute_data_mean_var in utils.py and example in rbf_network.py
Implemented multi-label RBF network that supports three-phase learning and experimented with the hyper-parameters, such as the number of centers and scaling factors, rbf_network.py
Implemented multi-layer neural networks and experimented with architecture, batch normalization and dropout and bagging, mlp_fuse_rgb_audio.py
Implemented inference that supports bagging many models, Class BootstrapInference in bootstrap_inference.py
Finally, the implementations can be easily adapted to other large datasets.