Design of a CMOS memristor emulator-based, self-adaptive spiking analog-to-digital data conversion as the lowest level of a self-x hierarchy

Abstract

Abstract. The number of sensors used in modern devices is rapidly increasing, and the interaction with sensors demands analog-to-digital data conversion (ADC). A conventional ADC in leading-edge technologies faces many issues due to signal swings, manufacturing deviations, noise, etc. Designers of ADCs are moving to the time domain and digital designs techniques to deal with these issues. This work pursues a novel self-adaptive spiking neural ADC (SN-ADC) design with promising features, e.g., technology scaling issues, low-voltage operation, low power, and noise-robust conditioning. The SN-ADC uses spike time to carry the information. Therefore, it can be effectively translated to aggressive new technologies to implement reliable advanced sensory electronic systems. The SN-ADC supports self-x (self-calibration, self-optimization, and self-healing) and machine learning required for the internet of things (IoT) and Industry 4.0. We have designed the main part of SN-ADC, which is an adaptive spike-to-digital converter (ASDC). The ASDC is based on a self-adaptive complementary metal–oxide–semiconductor (CMOS) memristor. It mimics the functionality of biological synapses, long-term plasticity, and short-term plasticity. The key advantage of our design is the entirely local unsupervised adaptation scheme. The adaptation scheme consists of two hierarchical layers; the first layer is self-adapted, and the second layer is manually treated in this work. In our previous work, the adaptation process is based on 96 variables. Therefore, it requires considerable adaptation time to correct the synapses' weight. This paper proposes a novel self-adaptive scheme to reduce the number of variables to only four and has better adaptation capability with less delay time than our previous implementation. The maximum adaptation times of our previous work and this work are 15 h and 27 min vs. 1 min and 47.3 s. The current winner-take-all (WTA) circuits have issues, a high-cost design, and no identifying the close spikes. Therefore, a novel WTA circuit with memory is proposed. It used 352 transistors for 16 inputs and can process spikes with a minimum time difference of 3 ns. The ASDC has been tested under static and dynamic variations. The nominal values of the SN-ADC parameters' number of missing codes (NOMCs), integral non-linearity (INL), and differential non-linearity (DNL) are no missing code, 0.4 and 0.22 LSB, respectively, where LSB stands for the least significant bit. However, these values are degraded due to the dynamic and static deviation with maximum simulated change equal to 0.88 and 4 LSB and 6 codes for DNL, INL, and NOMC, respectively. The adaptation resets the SN-ADC parameters to the nominal values. The proposed ASDC is designed using X-FAB 0.35 µm CMOS technology and Cadence tools.

On-Chip Adaptive Implementation of Neuromorphic Spiking Sensory Systems with Self-X Capabilities

“…The second stage, the self-adaptive spike-to-digital converter (SA-SDC), generates a digital code corresponding to the TD value. The three core principles of Jeffress's theory are integrated into the SA-SDC through synapse weights, an array of adaptable coincidence detection, and a winner-takes-all (WTA) mechanism with memory [18]. In our chip, we addressed the aforementioned issues by making two contributions to the universal-sensor-interface with self-X properties (USIX), as described in our earlier publication [16].…”

“…By adding more ACDs in a cascade, the ability to measure time intervals between in1 and in2 can be enhanced, resulting in an increased resolution of the SA-SDC. The number of ACDs needed to achieve a specific number of bits (NOB) can be determined using Equation (1), which was developed in [18].…”

“…Each ACD unit consists of one neuron (N) and two adaptive synapses (AS). In [18], we suggested an adaptive synapse that utilizes the CMOS memristor to imitate the short-term plasticity (STP) and long-term plasticity (LTP) of a biological synapse. Figure 5 displays the schematic of the adaptive synapse we proposed.…”

“…The first is the normal mode, in which both synapses of the ACD are connected to the neuron. The second mode is the adaptation mode, which was created in [18] by implementing a two-layer adaptation hierarchy. The first layer operates at the ACD level, while the second layer operates at the SA-SRC level.…”

“…This adaptive approach carries the promise to compensate both static and dynamic degradations, reversible and irreversible ones in both cases. In earlier work [17][18][19][20], the concept of using spike timing was introduced, to convey information and emulate acoustic localization.…”