Power/Performance Bits: March 16

Adaptable neural nets
Neural networks go through two phases: training, when weights are set based on a dataset, and inference, when new information is assessed based on those weights. But researchers at MIT, Institute of Science and Technology Austria, and Vienna University of Technology propose a new type of neural network that can learn during inference and adjust its underlying equations to continuously adapt to new data inputs.

Called ‘liquid’ networks by the researchers, they could be especially useful for data streams that change over time. “The real world is all about sequences. Even our perception — you’re not perceiving images, you’re perceiving sequences of images,” said Ramin Hasani, a postdoc in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). “This is a way forward for the future of robot control, natural language processing, video processing — any form of time series data processing. The potential is really significant.”

In designing the network, inspiration was taken from how neurons of the nematode C. elegans activate and communicate with each other. “It only has 302 neurons in its nervous system,” said Hasani, “yet it can generate unexpectedly complex dynamics.” In the equations used to structure the neural network, parameters are allowed to change over time based on the results of a nested set of differential equations.

Hasani said that the ‘liquid’ network is more resilient to unexpected or noisy data and is more interpretable, too. “Just changing the representation of a neuron, you can really explore some degrees of complexity you couldn’t explore otherwise. The model itself is richer in terms of expressivity.” The small number of highly expressive neurons make it easier to tell why the network made a certain characterization.

The team said the network performed well against other time series algorithms, accurately predicting future values in datasets ranging from atmospheric chemistry to traffic patterns. “In many applications, we see the performance is reliably high,” Hasani said, noting that the small network size also meant lower compute resources were needed. “Everyone talks about scaling up their network. We want to scale down, to have fewer but richer nodes.”

The researchers plan to keep improving the system and make it ready for commercial applications.

Photonic DAC
Researchers at George Washington University and University of California Los Angeles developed a photonic digital-to-analog converter (DAC) that does not require the signal to be converted in the electrical domain.

The researchers argue that power consumption and throughput bottlenecks in networks arise from electronic-to-optical and optical-to-electronic conversion.

The team used a silicon photonic chip platform for their 4-bit prototype of the coherent parallel photonic binary-weighted DAC and say the result has the potential to satisfy the demand for high data-processing capabilities while acting on optical data, interfacing to digital systems, and performing in a compact footprint, with both short signal delay and low power consumption.

“We found a way to seamlessly bridge the gap that exists between these two worlds, analog and digital,” said Volker J. Sorger, an associate professor of electrical and computer engineering at George Washington University. “This device is a key stepping stone for next-generation data processing hardware.”

The DAC’s operating speed and power consumption are limited by the performance of the modulators responsible for encoding the digital bit onto the optical domain, but the team sees future improvements to this on the horizon.

In the paper, the researchers add, “In contrast to other parallel photonic DAC implementations, this coherent photonic DAC does not require the signal to be converted in the electrical domain and therefore could support data I/O interfaces of high-throughput applications such as in domain-specific compute accelerations, emerging photonic integrated neuromorphic signal processing engines, and network-edge processing and data-routing platforms.”

Thermomagnetic generators
Karlsruhe Institute of Technology (KIT) and Tohoku University researchers are working to harvest waste heat at room temperature by increasing the electrical power per footprint of thermomagnetic generators.

When harvesting waste heat, the higher the temperature of it, the easier it is to capture its energy. Thermoelectric generators have been used, but the researchers said that they can be expensive and involve toxic materials while having low efficiencies.

Instead, they turned to thermomagnetic generators. The thermomagnetic generators are based on alloys that have temperature-dependent magnetic properties. Alternating magnetization induces an electrical voltage in a coil applied.

The team’s device uses a nickel-manganese-gallium Heusler alloy, which are magnetic intermetallic compounds. These are applied in the form of thin films in thermomagnetic generators and provide for a big temperature-dependent change of magnetization and quick heat transfer. Resonant vibrations are induced, and can be converted efficiently into electrical power.

The thermomagnetic generators are based on magnetic thin films with highly temperature-dependent properties. (Photo: IMT/KIT)

They found that alloy film thickness and the device footprint influence electrical power in opposite directions, and were able to increase power per footprint by 3.4 by increasing the thickness of the alloy film from five to 40 micrometers.

“Based on the results of our work, thermomagnetic generators are now competitive with established thermoelectric generators for the first time. With this, we have come a lot closer to the goal of converting waste heat into electrical power at small temperature differences,” said Professor Manfred Kohl, head of the Smart Materials and Devices Group at KIT’s Institute of Microstructure Technology.

The thermomagnetic generators reached a maximum electrical power of 50 microwatts per square centimeter at a temperature change of three degrees Celsius. “These results pave the way to the development of customized thermomagnetic generators connected in parallel for potential use of waste heat close to room temperature,” said Kohl.

Next, the researchers said that further material development and engineering is needed to improve the power generated.