Computers nowadays have this special way of making us feel less special for being human, beating us at some of our most challenging board games like Go- even at videogames and poker! Apparently, that wasn’t enough. Fellas, they are now invading one of our most beloved games, Pokémon.
These machines now have the ability to study and train until they can look at a Magicarp and tell you it’s a water-type Pokémon (not that Magicarps are any useful, smh). With 720 different Pokémon having 18 different types (water, fire, psychic, flying, etc.), I guess we can all agree that it’s not that easy- but it is possible.
Henrique Soares, a Brazil-based independent researcher wanted to find out how its image recognition ability would do in a game played by millions of people (not just children, okay?). So he utilized a technology called neural networks are “computer programs made up of "layers" of connected nodes that run semi-random computations on input data like images, and rearrange themselves until they "learn" how to recognize the objects in them. This is called "training," and once the process is completed, the network should be able to recognize unfamiliar objects that are similar to the ones it's been trained on already,” according to Jordan Pearson of Motherboard
He started by assembling a database of Pokémon game sprites (in computer graphics, a sprite is a two-dimensional bitmap that is integrated into a larger scene) for the neural network to train on. After gathering character sprites of the first five generations (that’s 649 total!), he then visually optimized the sprites and annotating them by type, keeping 20 percent of them for training, with the other 80 percent used for testing purposes.
Well, the program was able to recognize a Pokémon’s type with 39 percent accuracy overall. The AI was even highly accurate on some individual Pokémon, with 90 percent accuracy! It was also better at recognizing some types of Pokémon than others. For instance, the algorithm could detect a dark-type Meowth 92 percent of the time, but only 17 percent for the ghost-types. This low overall accuracy could be because of a small set of a few thousand sprites since the best neural networks to date train on millions of examples.
We don’t really know how the algorithm works in tiny detail, but thanks to newer visualization tools, Henrique was able to look into the “layers” and see which sprites “activate” neurons.
"We have a kernel that is clearly firing up over red regions of the sprite, which leads us to think that the network is learning to discern colors, [and] associating it to the types. But as we go deeper in the network and further from the original image those interpretations become harder and harder to make from simple inspection” says Soares.