Code

[Reproduce] Language as an Abstraction for Hierarchical Deep Reinforcement Learning

Submitted to NeurIPS 2019 Reproducibility Challenge

We tackle the issue of long-horizon planning and temporally-extended tasks in our replication, using language as abstraction for hierarchical reinforcement learning. The proposed approach selects language as the choice of abstraction because of its compositional structure, ensuring an ability to break down tasks into smaller sub-tasks. The authors train a low-level policy and high-level policy using an interactive environment built using the MuJoCo physics engine and the CLEVR engine. The authors show that using language as the framework between low-level policy and high-level policy allows the agent to learn complex tasks requiring long term planning, including object sorting and multi-object rearrangement. We focused on implementing and training the low-level policy from scratch, as that is where HIR is first introduced. For the low-level policy, we show that encoding the instruction with a GRU and using HIR performs better than a one-hot encoded representation of the instruction. However, our results for one-hot encoded representation as the number of total instructions grew contradicted what the conclusions from the original paper.
PDF  Code  Original Paper

[Reproduce] Convolutional Neural Networks for Speech Recognition
The speech recognition models previously included Gaussian Mixture Models with a hidden Markov model (HMM).The Gaussian mixture models are replaced with the deep neuralnetwork structures. The convolutional neural networks (CNNs) have improved significantly the performance of the speech recognition systems, as they model the complex correlations ofthe speech features. In this project, we trained a CNN - HMM hybrid model, and experimented on the TIMIT phone recognition dataset. In the report, we describe the structure and the use of the CNN model applied to the speech features. Then, we describe the decoding and training process of the HMM.
PDF  Original Paper

[Reproduce] Conditional Generative Adversarial Networks
An implementation of the Conditional Generative Adversarial Networks in PyTorch, used the CIFAR-10 dataset.
Code  Original Paper

[Reproduce] Image Classification with Deep Convolutional Neural Networks
An implementation of the original ImageNet paper in C++ and LibTorch, used the CIFAR dataset.
Code  Original Paper

[Reproduce] Siamese Neural Networks for One-shot Image Recognition
An implementation of the Siamese Neural Networks in PyTorch, trained and tested on the MNIST dataset.
Code  Original Paper

[Reproduce] 3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction
An implementation of the 3D-R2N2 paper in Tensorflow. Used the ShapeNet dataset.
Code  Original Paper

[Reproduce] Generative Adversarial Networks
An implementation of the Generative Adversarial Networks in Julia using KNet. Used MNIST, CIFAR and Labeled Faces in The Wild datasets.
Code  Original Paper