Lambda-Reservoir Computing

Summary

UCLA researchers in the Department of Electrical and Computer Engineering have developed a Spectral Reservoir Computer that processes data using nonlinear optical interactions.

Background

Computer vision and natural language processing (NLP) are being revolutionized by deep learning networks (DNN) and recurrent neural networks (RNN), which are sets of algorithms that use artificial intelligence to cluster and classify data. These networks are trained offline before being deployed in the intended application, which can take anywhere from tens of hours to a few weeks. However, many in-real-time applications such as autonomous vehicles and drones demand fast inference beyond what is achievable with conventional neural networks. Moreover, these neural networks assume that the distribution of input and output data does not change significantly over time, which is not the case in real world applications such as cybersecurity, where underlying data generating mechanisms are quickly evolving. There is great interest in replacing DNN or RNN with a dynamic system that has the potential to offer fast, real time learning and inference with potentially lower power consumption.

Innovation

UCLA researchers have developed a Spectral Reservoir Computer that processes data using nonlinear optical interactions. This “Lambda-Reservoir” approach accesses millions of lambda-nodes in a single physical node without sacrificing complexity and speed. This novel single-node reservoir computer obviates the need for physical feedback, hence greatly simplifying the hardware. Moreover, the Lambda-Reservoir can be integrated with time stretch data acquisition to capture the output of the reservoir in real time at THz bandwidths. This system is unique in that it serializes multidimensional data into a 1D time series.

Applications

• Artificial Intelligence
• Autonomous Vehicles
• Drones
• Cybersecurity

Advantages

• Accesses millions of lambda-nodes in a single physical node without sacrificing complexity and speed
• Avoids physical feedback, thus simplifying the hardware
• Can be integrated with time stretch data acquisition to capture outputs in THz bandwiths

Related Materials

• Goda, K., K. K. Tsia, and B. Jalali. "Serial time-encoded amplified imaging for real-time observation of fast dynamic phenomena." Nature 458.7242 (2009): 1145.
• Goda, Keisuke, and Bahram Jalali. "Dispersive Fourier transformation for fast continuous single-shot measurements." Nature Photonics 7.2 (2013): 102.
• Mahjoubfar, Ata, et al. "Time stretch and its applications." Nature Photonics 11.6 (2017): 341.

Patent Status

Patent Pending