Liquid machine learning system adapts to changing conditions

Liquid machine learning system- a system that trains itself to acclimatize to newer and newer conditions. Researchers at MIT have successfully fabricated a neural network that is capable of learning on the job contrary to the existing networks.  The credit for this goes to the supple algorithms that these networks use. Thus they are often referred to as liquid networks as they alter these equations involved based on newer data inputs. And that’s how these algorithms train themselves on the job and are not dependent on their prior training.

Also Read : Customized brains for robots

Liquid machine-learning

Ramin Hasani is the lead of the study for a liquid machine-learning system and believes strongly in its potential. (Hisami’s fellow researchers are Daniela Rus, Alexander Amini, Mathias Lencher, and Raduu Grosu). He proposes that the system would be extremely efficient in dealing with the processing of the data that involves streaming. This holds strong possibilities in robotics, processing of natural language, and even time-based data processing. Hasani adds that since our surroundings and the world is largely based on sequencing, i.e. we perceive things in a sequence, the liquid machine learning system can prove extremely beneficial. A crisp example of this is that we even see images in a sequence in the real world and not one at a time.

Promising future of liquid machine learning

The study of liquid machine learning systems can augment existing technologies like autonomous vehicles, diagnostic and indicative applications in the medical field, processing of financial data, and even video processing. The authors of the study maintain that data in the real world and otherwise is very uncertain. And to cater to these ever-changing real-world needs, this liquid neural network algorithm comes apt and powerful.

Inspiration from C.elegans:

Neural networks in the computer world are based on the functioning of the human brain. And Hasani’s model is greatly inspired by C.elegans’ brain’s mechanism. Now, the most impressive function of the C.elegans brain is that despite having only 302 neurons, the microscopic organism is extremely efficient in handling its real-world compound dynamics. Thus Hasani, after carefully observing C.elegans’ efficient brain mechanism, crafted his network that permitted changing of parameters. With time, these parameters would alter based on outcomes of differential equations.  This is a much-needed feature for the upcoming neural networks. And it is so because the existing neural networks have limitations in terms of adjustability and do not adapt well to the newer data. And thus Hasani’s work of flexible algorithms is recognizable.  

Striking features of liquid machine learning system:

Hasani’s model of liquid machine learning system is adaptable, more perceivable, and more robust. It outperformed other existing series based technologies in foretelling values from datasets. These datasets could vary from atmospheric chemistry to even noticing the patterns of traffic. Hasani thus feels jubilant and expresses his intent to have lesser but quality nodes. He further wants to scale down his network while everyone else is in rush to scale up. And Hasani at the same time wants to keep improving the quality of liquid machine learning systems to produce better and smarter future systems.