Machine Learning Mini-projects

• english phonetics - seq2seq model for translating english words to their phonetic transcription
  • (LSTM) Long-Short-Term-Memory units
  • (ResNetLSTM) custom Long-Short-Term-Memory with residual connections, allowing for building deep networks
  • (CNN-ED) convolutional neural network for building embeddings based on Edit Distance between phonetic transcriptions
• acrobot - reinforcement learning agent that learns to swing a double pendulum upwards
  • (DQN) Deep Q Learning
• MNIST autoencoder - convolutional autoencoder that compresses MNIST digits
  • (AE) Autoencoder solution
  • (SAE) Sparse Autoencoder solution
  • (SAE_STE) Sparse Autoencoder with Straight-Through Estimator (for binarized bottleneck layer) solution
  • (SC_STE) Sparse Classifier with Straight-Through Estimator (for binarized bottleneck layer) solution
  • (SCAE) Sparse Convolutional Autoencoder solution
  • (VAE) Variational autoencoder
  • (VAE) interactive explorer for 2D latent space
  • (GAN) Generative adversarial neural network
  • (DCGAN) Deep convolutional generative adversarial neural network (version with stride 1)
  • (DCGAN) Deep convolutional generative adversarial neural network (version with stride 2)
• Moving MNIST autoencoder - State-space model for encoding time series over latent space
  • (ConvLSTM) Convolutional LSTM
  • (Deterministic LSTM) recurrent network over latent space (with epsilon teacher forcing)
  • (RNN) deterministic recurrent network (without LSTM or any other form of long term memory) over latent space (with epsilon teacher forcing)
  • (SSM) stochastic state model - variational recurrent network over latent space (with epsilon teacher forcing)
  • (RSSM) recurrent state space model - recurrent network over latent space with both variational and deterministic connections (with epsilon teacher forcing)
• mountain car - reinforcement learning agent that learns to drive a car uphill
  • (DDPG) Deep deterministic policy gradient
  • (PPO) Proximal policy optimisation (copied from https://github.com/nikhilbarhate99/PPO-PyTorch)
• bipedal walker - reinforcement learning agent that learns to walk on two feet (on flat terrain)
  • (DDPG) Deep deterministic policy gradient (unfortunately the results are somewhat underwhelming)
  • (Dreamer) reinforcement learning with latent imagination
  • (custom Dreamer) my own adaptation of Dreamer
  • (GA) Genetic algorithm (with elite and no crossover) (the results are poor)
  • (GA) Genetic algorithm (with crossover and more relaxed survivorship policy)
• information extraction from NDA documents - named entity recognition and information extraction tasks
  • (BERT) solution based on huggingface pretrained BERT transformers
• publication date predication from text - using large language models for temporal prediction
  • (BERT) solution based on polish BERT with custom fine-tuining
• contrastive learning on CIFAR-10 - Building generative models of visual data. Next step after MNIST autoencoders and GANs
  • (SimCLR) A Simple Framework for Contrastive Learning of Visual Representations
• captcha solver
  • (Resnet+RNN) Resnet with RNN head
  • (Resnet) Resnet with linear head
  • (Resnet+SimCLR) Resnet pretraiend with contrastive learning
• English homework - My English teacher gave me a tedious homework. I didn't want to do it, so I let BERT do it for me
  • (BERT)
• Punctuation restoral for ASR - PoleEval challenge for punctuation restoral to aid ASR systems
  • (PolBERT)
• News topic classifier - classifies news fragment into one of specified categories
  • (Naive Bayes) multiclass naive bayes with logarithmic probabilities for numerical stability
• Japanese Amazon products web crawler
  • (Facets) Web scraper for facet extraction (with HTTP-header spoofing and CAPTCHA solver)
• Minecraft exploratory reinforcement learning - reinforcement learning agent that learns to navigate minecraft world
  • (SimCLR) Contrastive learning for higher-level world representations from visual inputs (based on ResNet50)
• PageRank
  • Animation showcasing Monte-carlo appraoch to page-rank
  • Animation showcasing effects of iterative linear transformation for pare-rank
  • A small website that implements page-rank integrated with Solr backend for browsing amazon products
• hierarchical clustering
  • Animation showcasing various hierachical clustering algorithms
  • unsupervised MNIST classification using Ward's method