Deep Learning Homework Course - Homework Solutions

Table of Contents

1. Homework 1: Dimensionality Reduction and Autoencoder on Fashion-MNIST

2. Homework 2: CNNs and Transfer Learning

   • Part 1: ResNet on CIFAR-10
   • Part 2: CNN on Fashion-MNIST

3. Homework 3: Time Series and Text Generation

   • Part 1: Time Series Prediction
   • Part 2: Character-level Text Generation

4. Homework 4: Masked Language Modeling and Graph Embedding

   • Part 1: Masked Language Model
   • Part 2: POS Tagging with Fine-tuned RoBERTa
   • Part 3: Graph Embedding using DeepWalk

5. Homework 5: Generative Models and Variational Autoencoder

   • Part 1: Autoregressive Image Modeling
   • Part 2: GAN and CGAN on MNIST
   • Part 3: Variational Autoencoder on Fashion-MNIST

6. Homework 6: Deep Q-Learning for Lunar Lander

---

Homework 1: Dimensionality Reduction and Autoencoder on Fashion-MNIST

Tasks Overview

In this practical assignment, I have worked with the Fashion-MNIST dataset to implement several neural network-based models. The tasks involve reducing data dimensionality, training an autoencoder, and visualizing latent representations using PCA and t-SNE.

• I implemented the PCA algorithm and applied it to the test set.
• I applied PCA and t-SNE to visualize the data in 2D.
• I trained a stacked autoencoder and used its encoder for classification.
• I used PCA and t-SNE to visualize encoded test set data.
• I added a classification layer on top of the encoder to predict image labels.

Comparison of input image and reconstructed image

---

Homework 2: CNNs and Transfer Learning

Part 1: ResNet on CIFAR-10

I have fine-tuned a pre-trained ResNet50 model for the CIFAR-10 image classification task. This involved loading the dataset, configuring the model for transfer learning, and evaluating its performance.

Evaluate fine-tuned Resnet over some samples of Validation Set

Part 2: CNN on Fashion-MNIST

In this part, I implemented a convolutional neural network (CNN) from scratch to classify images from the Fashion-MNIST dataset. The process included:

• Loading and splitting the dataset.
• Implementing the CNN model using PyTorch.
• Training, validating, and testing the model, with metrics tracking.

---

Homework 3: Time Series and Text Generation

Time Series Prediction

I have built and trained models for time series prediction, exploring and comparing three architectures: Simple RNN, GRU, and LSTM. I evaluated their performance and analyzed which model performed best and why.

Performance comparison of vanilla RNN, LSTM and GRU on Electric Production dataset

Character-level Text Generation

I implemented text generation at the character level using recurrent neural networks (RNNs). This involved training RNNs and LSTMs, fine-tuning the models, and generating coherent text sequences from the learned models.

---

Homework 4: Masked Language Modeling and Graph Embedding

Part 1: Masked Language Model

I have implemented a Masked Language Model (MLM) in PyTorch from scratch and compared it with a pre-trained RoBERTa model using HuggingFace's Trainer API.

Part 2: POS Tagging with Fine-tuned RoBERTa

I fine-tuned the RoBERTa model for a part-of-speech tagging task on the Parsig dataset. I leveraged RoBERTa's attention mechanisms to accurately predict and label the grammatical roles of words in a sentence.

Part 3: Graph Embedding using DeepWalk

In this task, I employed the DeepWalk technique to perform graph embedding on the small version of the Movielens dataset, transforming the movie data into vector space for downstream tasks.

---

Homework 5: Generative Models and Variational Autoencoder

Part 1: Autoregressive Image Modeling

I have implemented an autoregressive likelihood model for image generation. The model learned the conditional probability distribution of each pixel given its preceding ones. I also explored the benefits and limitations of such models.

Comparison of weighted receptive field and binary receptive field in PixelCNN

Part 2: GAN and CGAN on MNIST

I built and trained a basic Generative Adversarial Network (GAN) to generate handwritten digits from the MNIST dataset. Additionally, I implemented a Conditional GAN (CGAN) where image generation is conditioned on the digit labels.

Visualization of basic Generative Adversarial Network (GAN) results

Visualization of conditional Generative Adversarial Network (GAN) results

Part 3: Variational Autoencoder on Fashion-MNIST

I implemented and trained a Variational Autoencoder (VAE) on the Fashion-MNIST dataset. This task involved exploring how the VAE's latent space captures data variance, allowing for both reconstruction and sampling of new fashion items.

Interpolation of two images in the embedding space of a Variational Autoencoder (VAE)

---

Homework 6: Deep Q-Learning for Lunar Lander

I implemented the Deep Q-Learning (DQN) algorithm to control a lunar lander in OpenAI's Gym environment. The key aspects of this project included:

• Training an agent to safely land the lunar module.
• Using experience replay and target networks for stable training.
• Tuning hyperparameters and evaluating the model's performance.