Josephson Junction Circuits For Single-Photon Optoelectronic Neurons And Synapses

patent description

We recently invented and patented a new type of hardware for neuromorphic computing (Phys. Rev. Applied, 7, 034013 (2017) (attached to email). The key ideas are to use integrated photonic devices for massive connectivity nd superconducting electronics for efficient memory and computation. The present invention disclosure regards new circuit designs using Josephson junctions that enables several of the key functions required to realize the proposed optoelectronic neurons. Here we introduce circuits to achieve three functions: 1) convert a photonic signal to a supercurrent in a loop, which is essentially photon-to-fluxon transduction during a synaptic event, 2) setting and dynamically reconfiguring the synaptic weight, which is essentially changing the amount of flux which gets added to the loop during a synaptic event, and 3) detecting a supercurrent threshold to initiate a neuronal firing event.

For the first of these functions, a superconducting-nanowire single-photon detector (SNSPD) is in parallel with a Josephson junction (JJ). When the SNSPD detects a photon, it diverts current across the JJ (henceforth referred to as the receiver JJ), which causes it to produce single-flux quantum (SFQ) pulses. These SFQ pulses are then shuttled across another JJ and into an integrating  loop. Using a photonic receiver with a JJ to produce SFQ pulses has been demonstrated before. To our knowledge, it has not been done in conjunction with a JJ in a loop to provide a summing element which is crucial for the neuromorphic application under consideration. It is also an excellent way to produce a single-photon camera which integrates photon counts and converts them to electrical flux in much the same way a CCD converts photons to electrical charge. This could be used to produce a flux-coupled device (FCD) camera sensitive to single photons with spectral responsivity from the UV to the mid-IR. The problem this solves in the neuromorphic  application is to convert optical signals to the electronic domain and sum them so that multiple synaptic events from one or many other neurons can combine their inputs in a single superconducting loop. The limitations are that JJ are required, which can be difficult to fabricate reliably, and that there is some dynamic range over which the devices can operate (JJ le, number of photon events that can be summed, number of neurons which can be connected). These performance metrics are currently being investigated.

For the second of these functions, the synaptic weight is dynamically reconfigured by changing the current bias across the JJ which receives the SNSPD current (receiver JJ). This bias is reconfigured by adding flux to a superconducting loop which is inductively coupled to a wire biasing the receiver JJ. This loop is referred to as the excitatory synaptic storage loop, not to be confused with the neuronal integration loop described above. To increase the synaptic weight, an SFQ pulse is added to the synaptic storage loop, and to decrease the synaptic weight, a counter-circulating SFQ pulse is added to the loop. The amount of current which can be added to the synaptic storage loop determines the dynamic range of the synapse, and the number of SFQ pulses which saturate this dynamic range determines the bit depth of the synapse. The dynamic range is determined by the design of the JJ circuits comprising the synapse, and the bit depth is controlled  by the inductance of the synaptic storage loop. Flux can be added to the synaptic storage loop with photon detection events by directing the current from an SNSPD to a DC-to­ SFQ converter which will add a single SFQ pulse to the synaptic storage loop. This allows for strengthening or weakening of the synaptic weight based on firing activity in the superconducting optoelectronic network, which is crucial for unsupervised  learning. Single-photons can modify the synaptic weight; this is the fundamental limit for photonic circuit reconfiguration. To our knowledge, nobody has ever designed single-photon optoelectronic neurons which use JJ circuits to change the synaptic weights, so this entire concept is new. This solves two problems 1) modifying the effect of a single-photon synapse event on the integrated signal (synaptic weight) and 2) using single photons to update the synaptic weight. The limitations are again related to bit depth and dynamic range, and the parameter space is presently being explored. So far a synapse with 10 µA dynamic range with can be incremented with 1024 levels (10 bits) has been demonstrated with WRSpice simulations, and a circuit has been conceived to increment the synaptic weight with single photons or two photons, as is required for a Hebbian learning rule. This performance is excellent, far surpassing the required performance for these components.

The third of these functions is achieved by placing a JJ in the neuronal integration loop, referred to as the thresholding JJ. When the integrated current in the neuronal integration loop reaches the le of this JJ, subsequent single-photon synaptic SFQ pulses begin to be directed to an alternative current path, which initiates a neuronal firing event. A similar approach of using a JJ to detect a current threshold was introduced in 2007 to produce JJ neurons working entirely with SFQ pulses. What is new here is that it is integrated with photonic devices to detect a threshold number of photonic synaptic events, and it is used to drive a resistive heater which produces 1 V to drive a semiconducting light emitter. This solves the problem of establishing a critical current threshold which is a necessary element of neuronal operation. The limitations are again related to dynamic range and are being investigated. It is important to note that this threshold JJ could be replaced with another thresholding device such as a simple constriction, an nTron, or a yTron. The actual device used in practice will depend on the context, and all approaches are currently being investigated.