Utilizing electronic, ionic and photonic coupling for neuromorphic devices
Inspired by neural computing, the pursuit of ultralow power neuromorphic architectures with highly distributed memory and parallel processing capability has recently gained more traction. Memristive devices emulating brain-like signal processing forms the foundation for development of efficient lear...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Doctor of Philosophy |
| Language: | English |
| Published: |
Nanyang Technological University
2019
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/102448 http://hdl.handle.net/10220/48564 |
| Summary: | Inspired by neural computing, the pursuit of ultralow power neuromorphic architectures with highly distributed memory and parallel processing capability has recently gained more traction. Memristive devices emulating brain-like signal processing forms the foundation for development of efficient learning circuitry, but few devices offer analog-like switching transitions, symmetry, linearity and wide range of dynamic conductance states, necessary for efficient computation. Conventional drift-based memristors relying on stochastic metallic/defect filament formation fall behind in this regard due to their abrupt switching transitions. Very recently, diffusive memristors and second-order drift memristors have been engineered to approximate the biological synaptic dynamics based on metal atom diffusion, thermal dissipation, mobility decay and spontaneous nanoparticle formation. A common feature in all these systems is the insertion/extraction of mobile ions, directly or indirectly altering the electronic conductivity in a continuous analog-like manner. Hence, a hybrid ionic-electronic (ionotronic) conduction mechanism could be considered vital to mimic the biological neural dynamics. This calls for the need of material systems and device configurations where ionic and electronic conductivities co-exists intimately and which could be synergistically controlled in a manner, temporally similar to neural dynamics.
Drawing inspiration from this, the dissertation tackles the hardware switching requirements for bio-inspired computations by adopting novel memristive device configurations encompassing ionically and optically-active semiconductors and dielectrics. Ionic semiconducting properties of halide perovskites are initially investigated in a two-terminal memristive configuration, followed by ionic gating of metal oxide and two-dimensional (2D) transition metal di-chalcogenides (TMDCs) in a three-terminal field-effect configuration. Finally, optical modulation of carrier concentration is exploited to create optoelectronic synapses with multi-modal programmability and higher plasticity. All material sets and device configurations are investigated and compared based on the degree of achievable plasticity -specifically short and long-term plasticity, conductance linearity, symmetry of weight changes and energy consumption.
An ionic semiconductor which couples fast electronic transitions with slow drift-diffusive ionic kinetics would enable energy-efficient analog-like switching of metastable conductance states. Here, the intimate ionic-electronic (ionotronic) coupling in halide perovskite semiconductors is utilized to create memristive synapses with a dynamic continuous transition of conductance states. Co-existence of carrier injection barriers and ion migration in the perovskite films defines the degree of synaptic plasticity, more notable for the larger organic ammonium and formamidinium cations than the inorganic cesium counterpart. Optimized pulsing schemes facilitate a balanced interplay of short and long-term plasticity rules like paired-pulse facilitation and spike time dependent plasticity, cardinal for learning and computing. Trained as a memory array, halide perovskite synapses demonstrate reconfigurability, learning, forgetting and fault tolerance analogous to the human brain. Network-level simulations of unsupervised learning of handwritten digit images utilizing experimentally-derived device parameters, validates the utility of these memristors for energy-efficient neuromorphic computations, paving way for novel ionotronic neuromorphic architectures with halide perovskites as the active material.
In contrast to the abrupt switching transitions in a conventional two-terminal memristor configuration, a three-terminal field-effect transistor (FET) configuration exhibiting time-dependent hysteresis would result in much smoother conductance transitions. A gated control harnessing slow ion diffusive kinetics would enable precise modulation of electronic transitions and an analog programming of channel conductance. Combining multiple gating controls that can induce temporally distinct modulations on the carrier concentration would increase the weight plasticity and memory storage capacity, enabling emulation of more complex neuronal behaviour. Herein, artificial synapses with intimate electronic-ionic (ionotronic) coupling are realized in a dual-gated electric-double-layer neuristor configuration with high-mobility amorphous metal oxide semiconducting channels, a solid-state ionic top dielectric and a bottom electronic insulator. This multi-gated architecture allows one gate to capture the effect of local activity correlations and the second gate to represent global neuromodulations, enabling higher-order temporal correlations like heterosynaptic plasticity, homeostasis and association at a unitary level. The dual-gate operation extends the available dynamic range of synaptic conductance while maintaining symmetry in the weight-update operation, expanding the number of accessible memory states. Finally, operating neuristors in the sub-threshold regime enables synaptic weight changes with high gain, while maintaining ultralow power consumption of the order of femto-Joules.
Information transmission via light-pulses as opposed to voltage-packets could unlock exceptionally fast transmission speeds with less cross-talk, high parallelism, low noise and nearly unlimited bandwidth. Neuromorphic transistors programmable via multimodal electrical and optical pulses would pave way for novel synergistic architectures encompassing advantages of both electron and photon-based computing platforms. Here, temporal plasticity of 2D TMDC-based neuristors are investigated via synergistic electronic, ionotronic and photoactive gating controls, addressing different charge-trapping probabilities to finely modulate the synaptic weights. The slow recombination of photo-generated carriers augments the ionotronic modulation of channel conductance in such configurations, resulting in orthogonally programmable (electrical + optical) artificial synapses with enhanced plasticity. Synergistic gating controls amalgamate neuromodulation schemes to achieve “plasticity of plasticity-metaplasticity” via dynamic control of Hebbian spike-time dependent plasticity and homeostatic regulation. The optoelectronic gating approach facilitates precise programming of conductance states with high temporal plasticity, wide dynamic conductance range and high linearity, compatible with both spiking and deep neural networks. Photons pulses are utilized for potentiation and electrical pulses for depression of weights, allowing facile emulation of basic arithmetic operations like an abacus, in addition to the bio-inspired plasticity models. Finally, photons are utilized as global, virtual interconnects in an array composed of light sensitive and insensitive elements to selectively modulate plasticity. Global neuromodulations are encompassed as photon packets, regulating the network dynamics. Drawing inspiration from optogenetics, these devices are portrayed as optogenetic actuators in combination with dielectric elastomeric actuators to create an optogenetic toolbox for artificially intelligent behavioural soft robotics. |
|---|