Special AFOSR Session H-8
From Brain-inspired Networks for Multifunctional Systems to Neuromorphic Computing at the Edge of Biology
ABSTRACTS
H-8:IL01 Brain-inspired Synaptic Resistor Circuits for Multifunctional Intelligent Systems with Real-time Learning
YONG CHEN, University of California, Los Angeles, CA, USA
The human brain has long served as the inspiration for artificial intelligent (AI) systems. A neurobiological circuit concurrently executes the signal-processing and learning in parallel analog mode to formulates new tactics and adapt to changing environments in real time. The artificial neural network (ANN) run on computing circuits must first be trained in learning processes before they can execute signal-processing algorithms. A challenge remains in developing neuromorphic circuits that can perform learning and signal-processing concurrently in real-time, with adaptability to uncertain and mutable environments. In this work, we report a synaptic resistor (synstor) to emulate a synapse with memory, learning, and signal-processing functions. The voltage pulses triggered currents via synstors for signal-processing. Concurrently, synstor conductance is modified by following the synaptic learning algorithm. Without any prior training, the synstor circuits concurrently executed signal-processing and learning in real time to control multifunctional systems such as morphing wing and drone faster than human controllers, and with learning speed, performance, power consumption, and adaptability to the environment significantly superior to a trained ANN running on computers.
H-8:L02 Synstor-based Device Simulations and Learning Algorithms for Self-programming Neuromorphic Integrated Circuit
H.-T. Chien, Suin Yi, Texas A&M University, College Station, TX, USA
In-memory computing with analog memory devices including synaptic resistors (synstors) or memristors has occupied a significant portion of neuromorphic computing. This approach ideally leverages memory cells to replicate signal processing that occurs among synapses in biological neural networks. However, most of them currently relies on backpropagation, which is biologically implausible, to train neural networks. Furthermore, the use of backpropagation necessitates overwhelming memory buffer due to nonlocal nature until potentiation/depression of synapses. Here, we present a biologically plausible training algorithm, forward learning, to achieve energy-efficient and low memory-load computation. We demonstrate two-layer networks capable of classifying MNIST handwritten digit dataset with accuracies >90% within one epoch. Furthermore, forward learning is utilized for transfer learning, which also demonstrates high accuracy comparable to that of backpropagation. Finally, we compare other biologically plausible training algorithms to backpropagation in scaled artificial neural networks with >3 layers and >200 hidden neurons.
H-8:L03 Multimodal Actuators and Multifunctional Skins for Integrated Autonomous Systems
J.W. BOLEY, Boston University, Boston, MA, USA
Recently our group has realized a new palette of highly heterogenous 4D printing inks that can achieve lightweight multifunctional polymer composite structures with high stiffness (~40 GPa). With this platform we have realized autonomous robotic structures that can lift and carry unprecedented loads and maintain trajectories under large disturbances. However, our materials systems lack the additional functionality needed to achieve targeted applications. The overall objective of this research is to develop and integrate new multimodal lattice actuators with multifunctional skins to create new lightweight, multifunctional autonomous structures. To accomplish this objective, we (i) Develop new multifunctional elastomer skin materials to enhance the strain limit, toughness, and damage tolerance of new stiff multifunctional lattice materials, (ii) Develop new multimodal bending actuators that combine different types of actuator technologies for an unlimited number of operating configurations,, (iii) Integrate soft electromagnetic systems into the multifunctional skin, and (iv) Introduce new lattice designs for high force and volumetric shape change. We envision future autonomous aerial structures and exosuits that can significantly benefit the US Air Force warfighter.
H-8:L05 Grayscale Digital Light Processing 3D Printing for Multimaterial Additive Manufacturing
H. JERRY QI, The George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, USA
Additive manufacturing (AM, or also known as 3D printing) where materials are deposited in a layer-by-layer manner to form a 3D solid has seen significant advances in the recent decades. AM printing has the advantage in creating a part with complex geometry from a digit file, making them an ideal candidate for making multifunctional materials and structures. Multimaterial AM printing is an emerging field in recent years. It offers the advantage of placement of materials with different properties in the 3D space with high resolution, or controllable heterogeneity. It has also led to the emergence of 4D printing where the shape, properties, or functions can change after printing as a function of time. In this talk, we present a grayscale DLP (g-DLP) 3D printing method where we can print a part with gradient material properties that span four orders of magnitude using a single vat of resin. In addition, we present our modeling efforts on how light intensity and irradiation time can affect the printed quality. We also present our recent efforts of using evolutionary algorithm and machine learning to rapidly explore the design space and to obtain optimized design. This new design paradigm seamless connects design concepts to the 3D printed shape changing devices.
H-8:IL06 Hybrid Biomolecular Synapses for Sensing and Neuromorphic Computing at the Edge of Biology
S.A. SARLES, J. Maraj, M. Mansour, University of Tennessee, Knoxville, TN, USA; E. Schafer, V. Hu, N. Kamat, J. Rivnay, Northwestern University, USA
Despite recent progress in computing hardware that collocate brain-like plasticity and memory, most devices emulate only basic synapse and neuron functions and few are bio-compatible. Therefore, new materials and devices are needed to integrate sensing and neuromorphic computing in close proximity to living cells and tissues—at the edge of biology. These capabilities will enable new types of implantable or wearable, smart bioelectronics and transform how we monitor, predict, and control biological activities that benefit health, physical performance, and diagnosis and treatment of disease and injury. The goal of this AFOSR-sponsored project is to develop a hybrid biomolecular synapse capable of selective ionic and biochemical sensing, signal processing, learning, and memory in wet, cellular environments. The hybrid device pairs a biomembrane containing stimuli-responsive biomolecules with a soft semiconducting polymer. We hypothesize that this approach will: 1) lead to biocompatible sensing devices that transduce multiple types of physical stimuli; and 2) unlock new neuromorphic functionalities. I will discuss our work to fabricate, characterize, and model a prototype synapse that pairs a stimuli-responsive lipid membrane and a PEDOT:PSS-based organic electrochemical transistor.
H-8:L07 Drawing Inspiration from the Hippocampus for Next-generation Neuromorphic Computing
G.C. ADAM, Electrical and Computer Engineering Department, George Washington University, Washington, DC, USA
Modern deep neural networks have brittle behavior when dealing with new input types and hardware non-idealities. This poses a challenge when wanting to develop large capacity accelerators with continual learning capabilities based on emerging memory technologies, such as memristors. These novel devices have shown potential for the compact and efficient implementation of artificial synapses for neuro-inspired computing, but issues related to device non-idealities prevent practical adoption at scale. By comparison, work in neuroscience seems to indicate that the brain can perform astounding computation with unreliable synapses and in the presence of input noise. The hippocampus is a core region of the brain with a major role in learning and memory. It features a diversity of neuronal types with various connectivity patterns and also hosts a unique phenomenon in the healthy adult mammalian brain: the addition of new neurons via neurogenesis. In this work, the impact of neuronal diversity and neurogenesis-inspired neuronal age on artificial spiking neural network training will be reported. The efforts to develop neurogenesis-inspired networks and memristor-based hardware capable of continual learning will also be discussed.
H-8:L09 Biomolecular Materials and Networks Enabling Neuromorphic Computing at the Edge of Biology
J.S. NAJEM, N.X. Armendarez, A. Mohamed, The Pennsylvania State University, University Park, PA, USA; M.S. Hasan, University of Mississippi, Oxford, MS, USA
The rise of AI and IoT in biological domains has sparked interest in developing biocompatible, energy-efficient computing devices. Drawing inspiration from how the brain utilizes ion channels within lipid membranes for functions like generating action potentials and synaptic plasticity, we introduce biomolecular devices designed to emulate biological synapses. These devices consist of synthetic lipid membranes demonstrating volatile memcapacitance. When integrated with voltage-driven alamethicin ion channels, they exhibit volatile memristance, effectively replicating short-term synaptic plasticity in biological systems. We show that these biomolecular mem-devices find application in online learning and reservoir computing, particularly in tasks such as time series prediction and temporal data classification. Furthermore, we present a dynamic neuristor, comprising two biomolecular memristors in parallel, capable of generating Hodgkin-Huxley-like action potentials. This neuristor faithfully emulates biological neurons with remarkable power efficiency, making it suitable for nonlinear control and signal processing tasks. This research establishes the foundation for cost-effective, low-power biomolecular materials that can facilitate neuromorphic computing in biological environments.
H-8:IL11 High-precision Analog Computing with Memristors
J. Joshua YANG, University of Southern California, Los Angeles, CA, USA
In the era of the Internet of Things (IoT), the proliferation of ubiquitous sensors has led to a significant influx of analog data, commonly referred to as the "analog data deluge." This surge necessitates the development of a novel computing paradigm capable of efficiently processing analog data with high throughput. In-memory computing utilizing analog memristors has emerged as a promising solution for this purpose. However, existing efforts have primarily focused on low-precision domains like neural networks or have relied on frequent quantization, leading to notable energy and area efficiency drawbacks. To enhance the precision of analog computing, recent advancements have been made on two fronts. Firstly, memristors with an unprecedented precision of thousands of conductance levels have been developed, surpassing the precision of all other analog memory types. Secondly, a novel circuit architecture and programming protocol have been designed to achieve remarkably high precision while minimizing circuit overhead. These breakthroughs enable analog devices to perform high-precision computing, significantly expanding the range of applications for analog computing.
H-8:L12 Effect of Oxygen Vacancy and Si Doping on the Memristive Electrical Properties of Ta2O5
S. Islam, Spectral Energies; J. Lee, ARCTOS Technology Solutions; S. Ganguli, A.K. ROY, Air Force Research Laboratory, Wright-Patterson AFB. OH, USA
The resistive switching behavior in Ta2O5 based memristors is largely controlled by the formation and annihilation of conductive filaments (CFs) that are generated by the migration of oxygen vacancies (OVs). To gain a fundamental insight on the switching characteristics, we have investigated the electrical transport properties of two different Ta2O5 polymorphs (ϵ-Ta2O5 and λ-Ta2O5), using density functional theory, and associated vacancy induced electrical conductivity using Boltzmann transport theory. The projected band structure and DOS in a few types of OVs, (two-fold (O2fV), three-fold (O3fV), interlayer (OILV), and distorted octahedral coordinated vacancies (OεV)) reveal that the presence of OILV would cause Ta2O5 to transition from a semiconductor to a metal, leading to improved electrical conductivity, whereas the other OV types only create localized mid-gap defect states within the bandgap. On the combined effect of OVs and Si-doping, a reduction of the formation energy, OVs near Si atoms, and creation of defect states near the conduction band edge inferred to be advantageous to the uniformity of CFs produced by OVs. These findings can serve as guidance for further experimental work aimed at enhancing the uniformity and switching properties for Ta2O5-based memristors.
H-8:L13 Ferroelectrics for Emergent Silicon-integrated Optical Computing
A. DEMKOV, The University of Texas at Austin, Austin, TX, USA and La Luce Cristallina, Inc. Austin, TX, USA
Traditional computing based on CMOS technology is nearing physical limits in terms of miniaturization, speed, and power consumption. Consequently, alternative approaches are under investigation. The most promising is based on a “brain-like” or neuromorphic computation scheme, another is optical quantum computing. These approaches can be realized using silicon photonics (SiPh), and at the heart of both technologies is an efficient, ultra-low power broad band optical modulator. A complete or partial switch from electrons to photons would be revolutionary, the technology ultimately requires integration of active and passive photonic elements on a single chip. As silicon modulators suffer from relatively high-power consumption and large size, materials other than silicon are considered for the compact energy-efficient modulator. I will discuss recent progress in integrating ferroelectric oxides with SiPh for the purpose of fabricating modulators exploiting the linear electro-optic effect. These will enable neuromorphic circuit architectures that exploit shifting computational machine learning paradigms, while leveraging current manufacturing infrastructure. This will result in a new generation of computers that consume less power and possess a larger bandwidth.
H-8:L14 SWaP-Efficient System-on-a-Chip for Neuromorphic Computing
E. Yesil, C.-J. Tien, R. Hadi, H. Yong, D. Huang, Y. Chen, Mau-Chung Frank Chang, UCLA, Los Angeles, CA, USA
This paper will present our System-on-a-Chip (SoC) approach for achieving SWaP-efficient neuromorphic computing for future autonomous systems, aimed to replace the traditional offsite, power-hungry and latency-prone systems in remote regions with minimal to no-network connectivity. Central to this work is the introduction of RRAM-enabled neuromorphic compute units with long edge-learning retention time. As an example, we will detail the design and functionality of chip level circuits tailored towards flight controls, leveraging TSMC N40 (40nm) CMOS technology integrated with embedded RRAM arrays. This innovative design facilitates the implementation of AI tasks by dynamically reconfiguring the RRAM cell, adopting power-optimized algorithms to accomplish neuromorphic-like computing with low energy consumption, short latency, and highly compact on-chip solutions.