SEARCH · Engineering Papers
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Models for simulating some aspects of neural intelligence are presented and discussed. Special attention is given to terminal neurodynamics as a particular architecture of terminal dynamics suitable for modeling information flows. Applications of terminal chaos to information fusion as well as to planning and modeling coordination among neurons in biological systems are disussed.
Artificial neurons exhibiting volatile threshold switching and action potential‐like oscillations are crucial for brain‐inspired computing. While Complimentary Metal‐Oxide‐Semiconductor (CMOS)‐based strategies require hundreds of transistors to simulate each neuron, neuronal oscillations arise spontaneously in individual electro‐thermal devices due to nonlinearities like the Mott transition in VO 2 . Despite improved understanding of the physics, quantitative connections between neuronal performance and material properties remain under‐explored, preventing predictive neuron design and rational materials selection. In this work, a physics‐aware forward design methodology is developed for interrogating a wide palette of materials with properties varying by orders of magnitude, and their performance (high frequency, high dynamical reconfigurability and low power) under external circuit and device geometry constraints is assessed. The space of viable materials is identified to be much larger than previously recognized, with candidates from a range of materials classes, including Ge, GaP and MoS 2 . CMOS‐compatible performance (such as 100 GHz oscillating frequencies) can be achieved with CMOS‐compatible node sizes (≈10 nm). Finally, combinations of material properties yielding desired neuronal performance under uncertain design constraints are considered. This work solidifies forward design principles for electro‐thermal neuron devices, a necessary pre‐condition for inverse design from desired neuronal performance to required materials properties.
An improved algorithm has been devised for training a recurrent multilayer perceptron (RMLP) for optimal performance in predicting the behavior of a complex, dynamic, and noisy system multiple time steps into the future. [An RMLP is a computational neural network with self-feedback and cross-talk (both delayed by one time step) among neurons in hidden layers]. Like other neural-network-training algorithms, this algorithm adjusts network biases and synaptic-connection weights according to a gradient-descent rule. The distinguishing feature of this algorithm is a combination of global feedback (the use of predictions as well as the current output value in computing the gradient at each time step) and recursiveness. The recursive aspect of the algorithm lies in the inclusion of the gradient of predictions at each time step with respect to the predictions at the preceding time step; this recursion enables the RMLP to learn the dynamics. It has been conjectured that carrying the recursion to even earlier time steps would enable the RMLP to represent a noisier, more complex system.
Impulse neural networks use pulse trains to communicate neuron activation levels. Impulse neural circuits emulate natural neurons at a more detailed level than that typically employed by contemporary neural network implementation methods. An impulse neural circuit which realizes short term memory dynamics is presented. The operation of that circuit is then characterized in terms of pulse frequency modulated signals. Both fixed and programmable synapse circuits for realizing long term memory are also described. The implementation of a simple and useful unsupervised learning law is then presented. The implementation of a differential Hebbian learning rule for a specific mean-frequency signal interpretation is shown to have a straightforward implementation using digital combinational logic with a variation of a previously developed programmable synapse circuit. This circuit is expected to be exploited for simple and straightforward implementation of future auto-adaptive neural circuits.
Postmortem brains of patients diagnosed with HIV-1-associated neurocognitive disorders (HAND) exhibit loss of dendrites. However, the mechanisms by which synapses are damaged are not fully understood. Dendrite length and remodeling occurs via microtubules, the dynamics of which are regulated by microtubule-binding proteins, including microtubule-associated protein 2 (MAP2). The HIV protein gp120 is neurotoxic and interferes with neuronal microtubules. We measured MAP2 concentrations in human cerebrospinal fluid (CSF) and MAP2 immunoreactivity in rat cortical neurons exposed to HIV and gp120. First, we examined whether HIV affects MAP2 levels by analyzing the CSF of 27 persons living with HIV (PLH) whose neurocognitive performance had been characterized. We then used rat cortical neurons to study the mechanisms of HIV-mediated dendritic loss. PLH who had HAND had greater MAP2 concentrations within the CSF than cognitive normal PLH. In cortical neurons, the deleterious effect of HIV on MAP2-positive dendrites occurred through a gp120-mediated mechanism. The neurotoxic effect of HIV was blocked by a CCR5 antagonist and prevented by Helix-A, a peptide that displaces gp120 from binding to microtubules, conjugated to a nanolipoprotein particle delivery platform. Our findings support that HIV at least partially effects its neurotoxicity via neuronal cytoskeleton modifications and provide evidence of a new therapeutic compound that could be used to prevent the HIV-associated neuropathology.
A device physics model is developed to analyze spontaneous neuron-like spike train generation in current driven silicon p(+)-n-n(+) devices in cryogenic environments. The model is shown to explain the very high dynamic range (0 to the 7th) current-to-frequency conversion and experimental features of the spike train frequency as a function of input current. The devices are interesting components for implementation of parallel asynchronous processing adjacent to cryogenically cooled focal planes because of their extremely low current and power requirements, their electronic simplicity, and their pulse coding capability, and could be used to form the hardware basis for neural networks which employ biologically plausible means of information coding.
Graph algorithms enable myriad large-scale applications including cybersecurity, social network analysis, resource allocation, and routing. The scalability of current graph algorithm implementations on conventional computing architectures are hampered by the demise of Moore’s law. We present a theoretical framework for designing and assessing the performance of graph algorithms executing in networks of spiking artificial neurons. Although spiking neural networks (SNNs) are capable of general-purpose computation, few algorithmic results with rigorous asymptotic performance analysis are known. SNNs are exceptionally well-motivated practically, as neuromorphic computing systems with 100 million spiking neurons are available, and systems with a billion neurons are anticipated in the next few years. Beyond massive parallelism and scalability, neuromorphic computing systems offer energy consumption orders of magnitude lower than conventional high-performance computing systems. We employ our framework to design and analyze new spiking algorithms for shortest path and dynamic programming problems. Our neuromorphic algorithms are message-passing algorithms relying critically on data movement for computation. For fair and rigorous comparison with conventional algorithms and architectures, which is challenging but paramount, we develop new models of data-movement in conventional computing architectures. This allows us to prove polynomial-factor advantages, even when we assume a SNN consisting of a simple grid-like network of neurons. To the best of our knowledge, this is one of the first examples of a rigorous asymptotic computational advantage for neuromorphic computing.
The Locally Competitive Algorithm (LCA) uses local competition between non-spiking leaky integrator neurons to infer sparse representations, allowing for potentially real-time execution on massively parallel neuromorphic architectures such as Intel's Loihi processor. Here, we focus on the problem of inferring sparse representations from streaming video using dictionaries of spatiotemporal features optimized in an unsupervised manner for sparse reconstruction. Non-spiking LCA has previously been used to achieve unsupervised learning of spatiotemporal dictionaries composed of convolutional kernels from raw, unlabeled video. We demonstrate how unsupervised dictionary learning with spiking LCA (\hbox{S-LCA}) can be efficiently implemented using accumulator neurons, which combine a conventional leaky-integrate-and-fire (\hbox{LIF}) spike generator with an additional state variable that is used to minimize the difference between the integrated input and the spiking output. We demonstrate dictionary learning across a wide range of dynamical regimes, from graded to intermittent spiking, for inferring sparse representations of both static images drawn from the CIFAR database as well as video frames captured from a DVS camera. On a classification task that requires identification of the suite from a deck of cards being rapidly flipped through as viewed by a DVS camera, we find essentially no degradation in performance as the LCA model used to infer sparse spatiotemporal representations migrates from graded to spiking. We conclude that accumulator neurons are likely to provide a powerful enabling component of future neuromorphic hardware for implementing online unsupervised learning of spatiotemporal dictionaries optimized for sparse reconstruction of streaming video from event based DVS cameras.
Electronic neural networks made to learn faster by use of terminal teacher forcing. Method of supervised learning involves addition of teacher forcing functions to excitations fed as inputs to output neurons. Initially, teacher forcing functions are strong enough to force outputs to desired values; subsequently, these functions decay with time. When learning successfully completed, terminal teacher forcing vanishes, and dynamics or neural network become equivalent to those of conventional neural network. Simulated neural network with terminal teacher forcing learned to produce close approximation of circular trajectory in 400 iterations.
Although the potency of conventional anesthetics correlates with lipophilicity, an affinity to water also is essential. It was recently found that compounds with very low affinities to water do not produce anesthesia regardless of their lipophilicity. This finding implies that clinical anesthesia might arise because of interactions at molecular sites near the interface of neuronal membranes with the aqueous environment and, therefore, might require increased concentrations of anesthetic molecules at membrane interfaces. As an initial test of this hypothesis, we calculated in molecular dynamics simulations the free energy profiles for the transfer of anesthetic 1,1,2-trifluoroethane and nonanesthetic perfluoroethane across water-membrane and water-hexane interfaces. Consistent with the hypothesis, it was found that trifluoroethane, but not perfluoroethane, exhibits a free energy minimum and, therefore, increased concentrations at both interfaces. The transfer of trifluoroethane from water to the nonpolar hexane or interior of the membrane is accompanied by a considerable, solvent-induced shift in the conformational equilibrium around the C-C bond.
Most neuromorphic chips implement the single-compartment point neuron model where synapse circuits connect directly to a leaky integrate and fire (LIF) soma circuit. However, when using a biologically plausible soma circuit (e.g., Hodgkin-Huxley neuron model), an interface circuitry, such as a current conveyor circuit, is needed to transmit synaptic current to the soma circuit. This is especially true for ultra-low power neuron circuits, where membrane capacitance is on the order of 20 fF. This need for an interface circuit arises because the parasitic capacitance and leakage current caused by fabrication mismatch and second-order effects of the output transistors in the synapse circuits can disturb the spiking dynamics of the soma circuit if connected without an interface. Using an interface circuit to isolate the soma’s membrane capacitor from synapses resolves this issue. We propose to use a unidirectional resistor (a transconductance circuit) to connect the synapse and soma circuits instead of conventional current conveyor circuits. Using a biologically plausible spike pattern detection model, we show that the on-chip spike-timing-dependent plasticity (STDP) learning performance of the proposed unidirectional two-compartment neuron circuit is similar to a single-compartment circuit (with a current conveyor as an interface) and additionally, it is more power-efficient and biologically plausible. The chip is fabricated in the Taiwan Semiconductor Manufacturing Company (TSMC) 250 nm technology node and comprises a single neuron circuit.
Various technologies pertaining to allocating computing resources of a neuromorphic computing system are described herein. Subgraphs of a neural algorithm graph to be executed by the neuromorphic computing system are identified. The subgraphs are each executed by a group of neuron circuits serially. Output data generated by execution of the subgraphs are provided to the same or a second group of neuron circuits at a same time or with associated timing data indicative of a time at which the output data was generated. The same or second group of neuron circuits performs one or more processing operations based upon the output data.
A primary goal of molecular physiology is to understand how conformational changes of proteins affect the function of cells, tissues, and organisms. Here, we describe an imaging method for measuring the conformational changes of the voltage sensors of endogenous ion channel proteins within live tissue, without genetic modification. We synthesized GxTX-594, a variant of the peptidyl tarantula toxin guangxitoxin-1E, conjugated to a fluorophore optimal for two-photon excitation imaging through light-scattering tissue. We term this tool EVAP (Endogenous Voltage-sensor Activity Probe). GxTX-594 targets the voltage sensors of Kv2 proteins, which form potassium channels and plasma membrane–endoplasmic reticulum junctions. GxTX-594 dynamically labels Kv2 proteins on cell surfaces in response to voltage stimulation. To interpret dynamic changes in fluorescence intensity, we developed a statistical thermodynamic model that relates the conformational changes of Kv2 voltage sensors to degree of labeling. We used two-photon excitation imaging of rat brain slices to image Kv2 proteins in neurons. We found puncta of GxTX-594 on hippocampal CA1 neurons that responded to voltage stimulation and retain a voltage response roughly similar to heterologously expressed Kv2.1 protein. Our findings show that EVAP imaging methods enable the identification of conformational changes of endogenous Kv2 voltage sensors in tissue.
A phasic excitation-activation (PEXA) model is presented of the process of motoneuron excitation and the resultant activation and force development of a motor unit. The model input is an amount of depolarizing current (as when injected with an intracellular electrode), and the model output is muscle force. The model includes dynamics and nonlinearities similar to phenomena discovered experimentally by others: the firing rate response of motoneurons to steps of depolarizing current and the catch-like enhancement of force produced by overlapping motor neuron action potentials. The parameter values used in this model are derived from experimentally measured data and are expressed in physical units. Model predictions extend to published data beyond those used in generating the model parameter values.
In this work, we present Lift & Learn, a physics-informed method for learning low-dimensional models for large-scale dynamical systems. The method exploits knowledge of a system’s governing equations to identify a coordinate transformation in which the system dynamics have quadratic structure. This transformation is called a lifting map because it often adds auxiliary variables to the system state. The lifting map is applied to data obtained by evaluating a model for the original nonlinear system. This lifted data is projected onto its leading principal components, and low-dimensional linear and quadratic matrix operators are fit to the lifted reduced data using a least-squares operator inference procedure. Analysis of our method shows that the Lift & Learn models are able to capture the system physics in the lifted coordinates at least as accurately as traditional intrusive model reduction approaches. This preservation of system physics makes the Lift & Learn models robust to changes in inputs. Numerical experiments on the FitzHugh–Nagumo neuron activation model and the compressible Euler equations demonstrate the generalizability of our model.
A combinatorial optimization methodology is developed, which enables the efficient use of hypercube multiprocessors onboard mobile intelligent robots dedicated to time-critical missions. The methodology is implemented in terms of large-scale concurrent algorithms based either on fast simulated annealing, or on nonlinear asynchronous neural networks. In particular, analytic expressions are given for the effect of single-neuron perturbations on the systems' configuration energy. Compact neuromorphic data structures are used to model effects such as precedence constraints, processor idling times, and task-schedule overlaps. Results for a typical robot-dynamics benchmark are presented.
A semiautomated method for reconstructing objects from serial thin sections has been developed in the Biocomputation Center. The method is being used to completely, for the first time, type II hair cells and their innervations. The purposes are to learn more about the fundamental circuitry of the macula on Earth and to determine whether changes in connectivities occur under space flight conditions. Data captured directly from a transmission electron microscope via a video camera are sent to a graphics workstation. There, the digitized micrographs are mosaicked into sections and contours are traced, registered and displayed by semiautomated methods. Current reconstructions are of type II cells from the medial part of rat maculas collected in-flight on the Space Life Sciences-2 mission, 4.5 hrs post-flight, and from a ground control. Results show that typical type II cells receive processes from tip to six nearby calyces or afferents. Nearly all processes are elongated and have bouton-like enlargements; some have numerous vesicles. Multiple (2 to 4) processes from a single calyx to a type II cell are common, and approximately 1/3 of the processes innervale 2 or 3 type II cells or a neighboring cluster. From 2% to 6% of the cells resemble type I cells morphologically but have demi-calyces. Thus far, increments in synaptic number in type II cells of flight rats are prominent along processes that supply two hair cells. It is clear that reconstruction methods provide insights into details of macular circuitry not obtainable by other techniques. The results demonstrate a morphological basis for interactions between adjacent receptive fields through feed back-feed forward connections, and for dynamic alterations in receptive field range and activity during preprocessing of linear acceleratory information by the maculas. The reconstruction method we have developed will find further applications in the study of the details of neuronal architecture of more complex systems, to seek out shared organizational properties or neuronal networks and to understand better localization of synaptic changes in altered environments.
Male white mice were subjected to two and three week hypodynamia and then decapitated. Cytoplasmic RNA content per cell was measured by means of ultraviolet cytospectrometry. Changes in RNA content are shown, and the dynamics of the reparative processes of cells are discussed.