Neuromorphic Reservoir Computing with Memristive Nanofluidic Diodes

Abstract

Memristive systems show conductance states modulated by past electrical stimuli acting as artificial synapses. Most neuromorphic computing systems are based on solid-state memristive devices that use physical environments and electrical carriers different from the ionic solutions characteristic of biochemical and bioengineering applications. Here, we use membranes with multiple nanopores showing different conductance states in an aqueous electrolyte as a model for reservoir computing (RC). To this end, the different membrane conductances obtained with distinct sequences of voltage pulses in the millisecond range are used for the identification of 10-digit inputs in the case of both correct and corrupted inputs. Using the current rectification of the nanofluidic conical diodes, we explore two additional options: (i) the use of the current and its sign instead of the conductance in the digit identification and (ii) the use of an antiparallel arrangement of two membranes instead of the single-membrane unit.

Keywords: memristor; nanofluidics; nanopores; neuromorphic; reservoir computing.

1

Mapping of the positive and negative voltage pulses (input) into the distinct reservoir states (read-out) permits to identify the patterns of 10 digits (output). The ionic accumulation (voltage V _p > 0, high membrane conductance) and depletion (voltage V _n < 0, low membrane conductance) in the nanofluidic conical pores provide the short-term plasticity of the membrane conductance needed for RC. The final conductance states corresponding to the digits i = 0, 1, ..., 9 and the assumed identical membranes j = 1, 2, ..., 5 are grouped in the conductance matrix G _ij . These reservoir states are then processed by the read-out layer, which associates the different pulse inputs patterns to the distinct digit outputs, allowing an optimum separation between them.

2

(a) Membrane final conductance values for the 2⁴ = 16 voltage pulse sequences. (b) Signature G _ij patterns for the binary 5 × 4 images of the digits from “0” to “9”. (c) Classification accuracy per training epoch and confusion matrices after 10, 30, and 41 training epochs. The predicted label for each digit is represented by a dark-blue square. The training of the read-out layer is based on the gradient descent method for the digit classification task.

3

(a) Two patterns of a particular set of 1-bit corrupted digits and the confusion matrix. (b) Case of a particular set of 2-bit corrupted digits. (c) Average confusion matrix obtained over 1000 sets for the 1-bit corrupted case. (d) Case of 2-bit corrupted digits. While the digit identification worsens when the number of corrupted bits increases, the confusion matrix still shows significant average accuracies.

4

(a) Number of training epochs needed for a 100% classification accuracy over 10000 different realizations assuming a 10% maximum variability in the membrane conductance. (b) 20% maximum variability. (c) 25% maximum variability. (d) 35% maximum variability. The mean value of the needed epochs, ⟨N _epochs⟩, for each distribution is shown.

5

(a) Scheme of an individual conical pore in the membrane, the electrodes, and the experimental memristive (I)–voltage (V) curve obtained with a sinusoidal voltage wave of frequency 20 Hz. (b) Patterns of final currents obtained for each digit identification. A multipore membrane is considered for each of the five rows in the digit identification of Figure . (c) Accuracy vs training epoch using the current instead of the membrane conductance. (d) Case of an antiparallel arrangement of two identical membranes instead of only one membrane for each of the five rows forming each digit. The scheme of two individual conical pores, one in each membrane, the electrodes, and the experimental steady-state I–V memristive curve are shown. (e) Final currents obtained for each digit identification. (f) Accuracy vs training epoch using the current instead of the membrane conductance.