Neural networks are vital for modern artificial intelligences to perform at their best. These networks help in pattern recognition and allow AI’s to respond quickly to external inputs such as someone talking or something moving. But how do they work exactly? Neural networks are computing systems made of many connections between different elements called ‘nodes’, like the human nervous system but much smaller. The various connections ensure that the networks can process large amounts of data in a relatively small time. The individual nodes have also been found to diverge from the group and find their own function by themselves in order to make the whole net run efficiently. But that does not mean that each node doesn’t have an individual function.
While these computing systems are great for creating artificial intelligences by maintaining a humanesque structure, once they have been finalized, their engineers cannot peer into the way they actually work or what data they are processing, similar to a black box on an airplane. The creators only know the math and coding behind the scenes which is why neural networks have been a mystery since they were first created.
The engineers at CSAIL (Computer Science and Artificial Intelligence Laboratory at MIT) have found a way to not only look into how the neural networks function after they have been completed and installed but they have also learned how to allow the network to perform upwards of 20 tasks at once rather than just one per network. The new tasks include: coloring blank or grey images, advanced pattern recognition, and solving puzzles. These tasks may seem trivial to us but they are a huge victory for computer scientists as they are a few feet closer to a fully automated AI and a sentient robot species!
Neural nets are not stopping there though. These networks are commonly made of layers with each layer completing one task for the whole group which is performing a larger task. For example the system would try to recognize a scene so one layer would look at colors, another would work on pixels, and the a third layer would try to discern different textures and so on. This hierarchy may be no more with the new and improved multitasking neural network that is up and coming. Instead, one layer could recognize images or movement, another could recognize speech, etc.
Researchers cannot only further computer science and AI systems with these findings, but they can also shed light on the organization of the human brain, an amazing feat. Computer scientists also found several new algorithms that neural networks create for themselves to help them with computational photography and computer-vision.
This new technology could improve already existing AI tech. For instance, self-driving cars could become more safe by being able to better recognize obstructions on the road and speech recognition systems could further discern accents for best possible performance. The possibilities are limitless! Just try not to think of the possibility of a whole new species of sentient robots emerging.