A new kind of neural network that is capable of adapting its own inherent behavior after the first training stage might be the secret to large improvements in circumstances where conditions may change rapidly — such as autonomous driving, controlling robots, or even diagnosing medical problems.
These so-called “liquid” neural networks were invented by MIT Computer Science and Artificial Intelligence Lab’s Ramin Hasani along with his staff at CSAIL, plus they have the capability to significantly extend the flexibility of AI technology following the training stage when they are engaged in the true practical inference work completed within the area.
Normally, following the training stage, through which neural network calculations are supplied with a large volume of relevant objective data to enhance their inference capacities, and rewarded for correct responses to be able to optimize functionality, they’re basically fixed.
However, Hasani’s team developed a way by which his”liquid” neural net could accommodate the parameters to get”achievement” over time in reaction to new data, meaning that when a neural net tasked with understanding on a self-driving automobile goes from clear heavens into the snow, for example, it is better able to take care of the change in the situation and keep up a high degree of functionality.
The most important difference from the process introduced by Hasani and his collaborators is the fact that it concentrates on time-series adaptability, which means that instead of being assembled on training information that’s basically composed of quite a few snapshots, or static minutes fixed in time, the liquid networks essentially believes time string data — or even sequences of pictures instead of isolated pieces.
Due to how in which the system was created, it is really also more amenable to research and observation from researchers, compared to conventional neural networks. This sort of AI is generally known as a”black box,” since while people growing the calculations understand the inputs and the criteria for discovering and encouraging successful behavior, they can not typically determine what precisely is happening inside the neural networks that contribute to achievement.
This”liquid” version provides more transparency, and it is less expensive when it comes to computing since it depends on fewer, but more complex, computing nodes.
Meanwhile, performance results indicate that it’s better than other alternatives for accuracy in predicting the future values of known data sets. The next step for Hasani and his team is to determine how best to make the system even better, and ready it for use in actual practical applications.