Adversarial attacks on machine learning models has been a hot research topic for the last year. While many teams are working on understanding the implications of adversarial approach, it is still a new area.
There are two main approaches: White-box optimization requires access to trained model architecture and weights, and uses it’s differentiability property to generate adversarial sample. Black-box optimization treats model as an object with unknown hidden state that only has some I\O interface. While it’s generally more computationally involved, black-box scenario is much closer to real-life event.
In this article I will explain how to generate adversarial examples using genetic programming. …
Conway’s Game of Life is a cellural automaton that has couple of simple rules:
This seemingly simple guidelines can result in the overall complex emergent behavior.
A couple of years ago Kaggle has launched a competition whose goal was to “reverse the time arrow” in Game of Life. You are given 20x20 end board and are asked to predict the starting board that, when evolved, will resemble the end board state as much as possible. Metric used for evaluation is Mean Absolute Error: for each board number of cell states correctly guessed is divided by the total number of cells. Although some people have examined this problem, it is unknown just how difficult this will be as well as the successful approaches to this problem. …