Recent advances in nanotechnology have highlighted the potential of neural networks built from coupled nano-oscillators, technology which takes advantage of processes that mimic the nonlinear yet efficient abilities of the human brain’s neurons. Using existing machine learning techniques to train these networks for solving complex, real-life tasks is still hampered by the need to add energy-inefficient heavy circuitry. New work demonstrates a pattern classifier architecture that enables the use of nonlinear dynamics and synchronization technology of energy-efficient coupled nano-oscillators that can be trained using standard machine learning algorithms.

To train nano-oscillators’ classifiers, algorithms adjust the frequencies of oscillator pairs to reinforce expected synchronizations, pulling the natural frequencies closer, and weaken unexpected synchronizations that drive their natural frequencies apart. The existing training method has two drawbacks: The currents of all oscillators are fixed independently, leading to a reduced leverage to increase computational power, and it does not deal well with data with strong asymmetries. The proposed classification architecture allows overcoming these drawbacks.

Vodenicarevic et al. introduce the innovative idea of a classifier where the oscillator frequencies are learned linear combinations of the inputs. This allows them to utilize a training process based on gradient descent to iteratively minimize the total error function.

The team compared their new architecture to that of an existing reference classifier; it correctly classified a larger percentage of off-diagonal circular and concave target regions. The new architecture also correctly categorized a larger percentage of the standard data set known as Iris than the reference classifier did. Comparison of results on the Iris data set further highlights the power of the nonmonotonic and interunit interactions of the oscillator-based architecture.

Source: “Nano-oscillator-based classification with a machine learning-compatible architecture,” by Damir Vodenicarevic, Nicolas Locatelli, Julie Grollier, and Damien Querlioz, Journal of Applied Physics (2018). The article can be accessed at