Cornell researchers train physical systems and revolutionize machine learning
A Cornell research group led by Professor Peter McMahon, Applied Physics and Engineering, has successfully trained various physical systems to perform machine learning calculations much like a computer. The researchers achieved this by transforming physical systems, such as an electrical circuit or a Bluetooth speaker, into a physical neural network – a series of algorithms similar to the human brain, allowing computers to recognize patterns of intelligence artificial.
Machine learning is at the forefront of scientific endeavors today. It’s used for a whole host of real-world applications, from Siri to search optimization to Google Translate. However, the energy consumption of chips is a major issue in this area, because the execution of neural networks, the basis of machine learning, consumes an immense amount of energy. This inefficiency severely limits the expansion of machine learning.
The research group has taken the first step towards solving this problem by focusing on the convergence of physical sciences and computation.
The physical systems trained by McMahon and his team – consisting of a simple electrical circuit, a loudspeaker and an optical network – identified handwritten numbers and spoken vowels with a high degree of accuracy and more efficiency than conventional computers.
According to the recent Nature.com article, “Deep Physical Neural Networks Trained with Backpropagation,” conventional neural networks are typically built by applying layers of mathematical functions. This relates to a subset of machine learning known as deep learning, in which algorithms are modeled on the human brain and networks are expected to learn the same way the brain does.
“Deep learning is usually driven by mathematical operations. We set out to make a physical system do what we wanted it to do – more directly,” said co-author and postdoctoral researcher Tatsuhiro Onodera.
This new approach results in a much faster and more energy-efficient method of performing machine learning operations, providing an alternative to the power-intensive requirements of conventional neural networks.
It might seem that this advantage of energy efficiency would be limited to small calculations, which would not require a significant amount of energy to start. However, larger calculations lead to greater energy efficiency, according to Onodera.
The potential of these physical neural networks goes beyond saving energy. According to McMahon, larger and more complex physical systems would have the ability to work with much larger data sets and with greater precision.
In addition, it is possible to connect together a series of different physical systems. For example, a loudspeaker could be connected to an electrical circuit to obtain a more complex system with greater potential.
“As you make the system bigger, it gets smarter,” Onodera said. “The range of things he can do is more versatile.”
Most of these physical systems can perform all the functions needed for machine learning calculations on their own in the same way as conventional systems. For example, when fed with handwritten numbers for image classification, physical networks can extract spatial features and determine the number on their own in much the same way as conventional neural networks.
The team also theorizes that many of the problems associated with the formation of conventional networks – such as the unintended decrease or increase in loss calculation in the feedback process – would disappear in the case of physical networks.
“If you look at each individual component [of the physical system], it could do something completely different,” said co-author and postdoctoral researcher Logan Wright. “It goes from point A to point B, but the trajectory is potentially completely different.”
Even if physical systems experience some form of wear and tear, which disrupts their computational capabilities, they can still be recycled, negating the adverse effects of any physical damage.
Currently, physical neural networks are only capable of an anticipation process. This means that they cannot train and retrain in the same way as recurrent neural networks – which have a constant feedback mechanism and can update their parameters as needed. Onodera, however, expressed optimism about training these systems to perform a recurrent feedback process.
Even though physical neural networks present a new approach to machine learning, they could potentially change the face of the field in the future. Wright wrote that one of the main reasons for this potential is that these systems replicate our brains more closely than other types.
Different types of physical systems are suitable for different types of learning operations and calculations. However, it may take some time for these physical networks to become largely integrated into the machine learning ecosystem, largely driven by conventional neural networks.
“The brain has evolved, [to the point] where physics and algorithms are all intertwined,” Wright said. “That’s what we’re getting closer to – physical algorithms instead of just hardware or software.”