Code for the labs of the Deep Learning course of the MSc in Artificial Intelligence at the University of Amsterdam.
One of the core problems in Computer Vision is that of Image Classification, which consists of assigning a label to an image from a fixed set of categories. In this work we explore Neural Network-based approaches towards solving this problem using the CIFAR-10 dataset. We perform extensive experimentation on hyperparameter effect for the MLP architecture. According to our experimental results, the convolutional architectures achieved significantly higher performance. [Theoretical Exercises] - [Report]
In this work we explore the capabilities of Recurrent Neural Networks (RNNs) and Long Short-Term Networks (LSTMs) for language modelling. In the first part we apply this models to a synthetic dataset of palindrome numbers; and in the second part we use a stacked LSTM architecture to learn and generate text trained on real books. Our experiments show a clear advantage of LSTMs over RNNs for learning dependencies which span over very large time steps. This advantage in turn allows for a good performance in the text generation task. [Report]
The bold characters are picked randomly and used the initialize the LSTM state.
| Train Step | Sampled Text |
|---|---|
| 0 | 72qoU)qV:UUD7Pb9wg8HAJ:606u1(h4BMN3FuZ!(cu59;dQWSOT0-E0DientlAU0b6zK |
| 3000 | d gone, and said, Wifhin, so no? I flew to me have I not have her shill came boy, the room |
| 6000 | Queen, and they went as before, and restored and spread and beasts. And say needle, and had |
| 9000 | You sit into his eat and at an each of the king s bride should go into the golden bird am I! |
| 12000 | Just take home again, I may go into his husband in the courtyard of dress; and how she saw |
| 15000 | Gretel off again in the garden. With all my heart and more fellow, and they were once tired; b |
| Train Step | Sampled Text |
|---|---|
| 0 | N0É0Aü„As2PoGxLwí ÉcJéTy 3!Á-PeLK4A:ZyúHí?Q!449e’ñ¿hc’3Tc(ñrs;36üáhÑ’G |
| 3000 | -dijo Sancho- que a decirlo de nadir como las comestamentes anday de lo que así se acompaña |
| 6000 | üera de no saber caballero de la ruega por su corta que la llama a poner las hermosas |
| 9000 | Y padres y allí se consentiría de mi amigo. Pero, viendo que había sido salir de don Quijote |
| 12000 | Éstos, así de sus almas y abajos los dueñas de la buena caballería en la cual y pusiese a quien |
| 15000 | Lo que antes se puedo vencido bien hermo de escudero, la presencia con los diablos y responde |
In this assignment we study how the problem of intractability arises in generative modelling, and how Variational Inference deals with that intractability. We train a Naive Bayes model and a Variational Autoencoder to generate images from the MNIST dataset. [Report]
Manifold learned by the VAE
- Tensorflow
- SciPy
- NumPy
- Matplotlib
Copyright © 2017 Jose Gallego.
This project is distributed under the MIT license. This was developed as part of the Deep Learning course taught by Efstratios Gavves at the University of Amsterdam. Please follow the UvA regulations governing Fraud and Plagiarism in case you are a student.