Optoelectronic neural networks comprise a system of interconnected processing units (neurons) interconnected by integrated photonic waveguides. The processing units receive photonic signals from other units. Each unit sums the received signals on a waveguide-integrated photon detector, and when the signal exceeds a threshold, a current pulse is delivered to a waveguide-integrated photon source which delivers its signal downstream to the other processing units to which it is connected. As in a biological neural system, the strength of connection between two neurons can be varied either in hardware or dynamically. Such optoelectronic neural networks can be implemented either with classical light levels using reverse-biased p-i-n photodetectors and standard digital electronics components to achieve thresholding, or with few-photon signals and superconducting single photon detectors to perform thresholding.
We propose a NC system based on superconducting detectors and electronics working with waveguide-integrated nano-LED emitters to behave as complete spiking neurons. Optical signals are communicated through reconfigurable nanophotonic waveguides, capturing the interconnectivity of the dendritic arbor. Firing thresholds and neuron gain are controlled by a dynamic superconducting network, and neuron-generated photonic signals can reconfigure this current-distribution network.
As mentioned in the abstract, similar functionality can be achieved with purely semiconducting components, albeit with reduced energy efficiency.
The invention comprises an optoelectronic circuit family based on superconducting photon detectors and superconducting electronics with semiconducting waveguide-integrated few-photon light-emitting diodes (LEDs). The components can be configured in myriad configurations to achieve diverse functionality suitable for neuromorphic computing.
The combination of efficient faint-light sources and superconducting-nanowire single-photon detectors interacting in an integrated-photonics environment enables neuronal operation with excellent energy efficiency, enormous intra- and inter-chip communication bandwidth, light speed-limited latency, compact footprint, and relatively simple fabrication.
The optoelectronic hardware platform achieves -20 al/synapse event, an improvement of 105 over recent CMOS systems.
Memory can be implemented in several ways including temporally fixed; synaptic weight variation via the actuation of locally suspended waveguides; or through the use of magnetic Josephson junctions. The suspended waveguides are reconfigurable on a time scale of lus. All of these approaches draw zero power in the steady state.
Enabled by the use of light rather than electric signals, massive interconnectivity is achievable with no need for time-multiplexing schemes that have limited CMOS systems to kHz event rates. Each unit in our system is capable of event rates of at least 20 MHz and capable of fully asynchronous activity.
Further, the use of optical signals introduces the capability to utilize the frequency degree of freedom. This can, for example, be employed to route one color to an inhibitory port of a receiving neuron and another color to an excitatory port, thereby emulating the behavior of GABA and glutamate neurotransmitters.
In addition to the use of photonic signals, the integration with a superconducting electronics platform enables the utilization of Josephson-junction-based circuits. In particular, this is crucial for DC biasing with zero static power dissipation. Additionally, simple circuits are presented to transduce fluxons to photons and vice versa.
This system can encode information in DC currents (threshold setting), DC voltages (synapse configuration), photon number (pulse height), photon frequency (color), photon pulse rate (firing rate), fluxon number, and magnetic synapse configuration. Because these processing units are each capable of representing a large number of bits, because these degrees of freedom are mutually interacting, and because a large number of processing units can be interconnected, the mutual information between this system and the applied stimulus can be made exceptionally large.
Combined with the ultra-low power densities of superconducting electronics, this technology appears capable of achieving states of complexity rivaled only by biological systems.