An experimental sandbox for implementing, testing, and visualizing ML models.
The figures below show the reverse diffusion process on MNIST digits:
each row corresponds to a target digit (0–5), and columns show samples
evolving from pure noise (left) to a clean digit (right).
This variant conditions the reverse diffusion process only on a sparse “dot image”: an input canvas with N bright dots (single pixels), where N encodes the target class. At inference time, I provide the dot image and the model denoises from pure noise into a clean MNIST digit consistent with that dot-conditioning.
This project trains a reinforcement-learning agent to track the lowest point (global minimum) of a complex 2D landscape that slowly evolves over time, similar to staying in the deepest valley while the terrain itself is moving. The agent learns this behavior purely through trial and error, without access to gradients or prior knowledge of the landscape. The image shows time-series comparisons between the agent’s position, the true global minimum, and the objective value, while the accompanying video provides an intuitive visual of the agent moving across the changing contour map as it follows the global minimum in real time.
Click to view the demo video (hosted on GitHub)
https://github.com/user-attachments/assets/4e25493f-4365-4ad8-9c86-58174bd5ac4a