The goal here is to evolve a target bitmap image that is not known. The learning signal is the similarity between any image and the target. The fitness function was defined to return this similarity measure. The population evolves over time to ultimately generate the target image.
| Target image | Evolved image |
|---|---|
 |
 |
The goal here is to optimize CTRNNs to have intrinsic oscillatory dynamics. The fitness function was defined to measure the average absolute difference in neural network activity in consecutive time-steps during the course of a simulation. Results from 5 different runs are shown here along with sample activity from the 5 neurons at the end of one of the runs.
| Fitness over time | Evolved oscillator |
|---|---|
 |
 |
The goal here is to classify the MNIST digits using a single-layer feed-forward neural network. The fitness function was defined as a network's performance on a random batch of 20000 samples from the dataset. Results from 5 different runs showing fitness over time and final performance on a validation set. [Requires hyperparameter tuning]
| Fitness over time | Avg. Validation Perf |
|---|---|
 |
89.43% |