This article discusses a groundbreaking neural network that operates using sound instead of traditional digital methods. The researchers have developed a system that uses vibrations on a titanium plate to classify handwritten digits, with the ability to adapt its behavior through trial and error. The article explores the potential of physical systems, such as this one, to revolutionize computing by leveraging the efficiency of natural computation. The challenges lie in building systems that can both think and learn, but progress has been made in developing physical systems that can perform both tasks. While these systems are not yet as efficient as digital neural networks, researchers believe they have the potential to surpass them in the future.
| Signal | Change | 10y horizon | Driving force |
|---|---|---|---|
| Neural network using sound as input and output | Technological | More advanced devices using unconventional inputs and outputs | Utilizing the unique properties of physical systems for efficient computation |
| Physicists developing physical systems for computation | Paradigm shift | Physical systems used as neural networks for both thinking and learning | Leveraging the universe’s ability to compute without explicit mathematical calculations |
| Physicists exploring the “thinking” half of the puzzle | Advancement in physical systems | Development of physical systems that can act as neural networks for thinking | Sidestepping the limitations of digital neural networks and harnessing the potential of physical systems |
| Physicists exploring the “learning” half of the puzzle | Advancement in learning algorithms | Development of alternative learning methods that don’t rely on backpropagation | Seeking more efficient and biologically inspired learning approaches |
| Combination of physical systems and digital models | Integration of digital and physical | Hybrid systems that combine the strengths of physical systems and digital models | Overcoming the limitations of physical systems by using digital models for training |
| Development of self-learning analog circuits | Advancement in circuit technology | Self-learning circuits with synaptic weights that adapt through coupling | Emulating the brain’s ability to learn without a centralized control structure |