Researchers from University of Pennsylvania have developed an electrical circuit that mimics a neural network and can perform tasks like nonlinear regression. This circuit functions at low power levels and does not require a computer to be trained.

The circuit consists of a network of transistors that act as virtual variable resistors. It learns by adjusting the resistance values through a method known as coupled learning. The researchers demonstrate that this circuit can handle nonlinear functions, inclusive of XOR and nonlinear regression, indicating that larger configurations might perform arbitrary functions similar to those by deep learning networks. A key benefit of this circuit-based learning is scalability; as a physical system with parallel element updates, the training time does not increase with network size. Additionally, its computation during inference is quick and energy-efficient. According to the research team:

The circuit is durable, can be retrained within seconds, and executes learned tasks in microseconds while only using picojoules of energy per transistor. This shows promising potential for efficient, high-speed computing in localized systems such as sensors, robotic controllers, and medical devices, while also being manufacture-ready for large-scale applications pertaining to emergent learning.

The design includes two identical networks of MOSFET transistors, with corresponding transistors in each network connected to a single capacitor. During the training phase, both networks receive the same training input voltages. However, in one network, termed the clamped network, the output is fixed to a desired state, whereas the other network, or free network, does not have a predetermined output.

Machine Learning without a Processor: Emergent Learning in a Nonlinear Electronic Metamaterial

The difference in electrical states between two networks adjusts the voltages in capacitors linked to the transistors, representing the neural network’s “weights.” After training, these voltages can be frozen, allowing the system to perform inference tasks by applying new input voltages and observing the output voltage.

Although this technology presents benefits such as speed and reduced power consumption, there are several challenges noted by the researchers. The optimal network configuration for specific applications remains to be determined. Currently, they employ a square lattice topology, which is considered “likely too simple and sparse.” Further investigations are necessary to understand how training duration and power needs scale with the size of the system.

In a discussion on Hacker News, a participant raised a question:

Once the training is complete, one topic not covered in the document is the method for maintaining the charge on the gate capacitors, which resemble the weights in a traditional neural network. For any effective implementation, it’s vital to constantly refresh this so that the weights remain stable. I wonder if this was mentioned elsewhere that I might have missed?

Another user proposed the use of MOS capacitors, akin to a flash drive, for maintaining the charge, but noted that this could restrict the capacity of training the system could endure.

The intersection of hardware and machine learning has deep roots, dating back to the creation of the initial perceptron by Frank Rosenblatt in 1957. Building on this legacy, researchers at MIT recently developed an analog deep learning system that employs programmable resistors to operate a fast, energy-efficient neural network right on the hardware.

Welcome to DediRock, your trusted partner in high-performance hosting solutions. At DediRock, we specialize in providing dedicated servers, VPS hosting, and cloud services tailored to meet the unique needs of businesses and individuals alike. Our mission is to deliver reliable, scalable, and secure hosting solutions that empower our clients to achieve their digital goals. With a commitment to exceptional customer support, cutting-edge technology, and robust infrastructure, DediRock stands out as a leader in the hosting industry. Join us and experience the difference that dedicated service and unwavering reliability can make for your online presence. Launch our website.