Deep Neural Networks

These are projects that I have done during Deep Neural Networks course at MIMUW.

Task1

Neural network recognizing MNIST dataset digits (written only using NumPy) and brief noise experiments when randomizing labels and when randomizing values.

Task2

Neural network which given an image with plain background and MNIST digits randomly placed on it (possibly rotated) it recognizes digit value, its rotation and placement on the image.

Obtained 44% accuracy.

Task3

Simple sentiment analysis of IMDB reviews dataset.

Obtained 90% accuracy.

Task4

Neural network using DDPG (Deep Deterministic Policy Gradient) which obtained optimal scores on games:

• Pendulum-v1

  • Mean episode reward: -158

  • Mean episode length: 200.0 (it's constant)

• LunarLanderContinuous-v2:

  • Mean episode reward: 195

  • Mean episode length: 404.02

• MountainCarContinuous-v0:

  • Mean episode reward: 93

  • Mean episode length: 71.99