With machine learning and artificial intelligence becoming increasingly widespread, we can notice their presence ranging from computer vision to text generation. However, the actions require complex neural networks and require exponentially high energy consumption. These technologies are becoming more and more unsustainable due to their rising energy demands for training. To reduce this, researchers found a way where neural networks can make machine learning more sustainable.
Researchers at the Max Planck Institute for Science of Light bring forward their research demonstrating a simple method to reduce energy consumption by AI. Considering the need for a more cost- and energy-efficient alternative of neuromorphic computing is required. Neuromorphic computing replaces the neural networks on our computers with physical neural networks. This performs mathematical operations physically faster and in a more energy-efficient way.
Challenges
- High laser powers are required for performing necessary complex mathematical computations.
- The unavailability of general training methods for such physical neural networks.
Director at the Institute, Florian Marquardt explains, “Normally, the data input is imprinted on the light field. However, in our new methods, we propose to imprint the input by changing the light transmission.”
Highlights
- Photonics and optics are being seen as promising platforms for neuromorphic computing as they keep energy consumption to a minimum.
- It is possible to perform computations in parallel at a very high speed.
- Both challenges can be overcome with the new proposed method by researchers as the input signal is processed in an arbitrary fashion.
The light field behaves in the simplest way where interfering with each other does not influence one another. Therefore, with the new approach, it is possible to avoid complicated physical interactions to perform required mathematical functions.
In the words of Clara Wanjura, the first author of the study said, “At the same time, this allows one to measure all relevant information for the training.”
Also, according to another research new computer chips can increase AI’s energy efficiency,
Conclusion
With various demonstrations, researchers are planning to collaborate with different experimental groups to explore the implementation of their method under different conditions. The approach that neural networks can make machine learning more sustainable, if successful, this system for neuromorphic devices can be used to physically train a broad range of platforms.
Source: Fully nonlinear neuromorphic computing with linear wave scattering
