A5010462308- Advanced Memory and Neural Computing
- Neural dynamics and brain function
- Ferroelectric and Negative Capacitance Devices
- Neural Networks and Reservoir Computing
- Neuroscience and Neural Engineering
- CCD and CMOS Imaging Sensors
- Advanced Neural Network Applications
- Advanced Chemical Sensor Technologies
- Neural Networks and Applications
- Adversarial Robustness in Machine Learning
- Generative Adversarial Networks and Image Synthesis
- Analytical Chemistry and Sensors
- Electrochemical Analysis and Applications
- Molecular Communication and Nanonetworks
- Physical Unclonable Functions (PUFs) and Hardware Security
- Insect Pheromone Research and Control
- Non-Destructive Testing Techniques
- Speech and Audio Processing
Syngenta (Switzerland)
2023
University of California, San Diego
2014-2018
Intel (United States)
2016
University of Zurich
2010-2014
ETH Zurich
2011-2014
SIB Swiss Institute of Bioinformatics
2011-2014
Abstract By mimicking the neurons and synapses of human brain employing spiking neural networks on neuromorphic chips, computing offers a promising energy-efficient machine intelligence. How to borrow high-level dynamic mechanisms help achieve energy advantages is fundamental issue. This work presents an application-oriented algorithm-software-hardware co-designed system for this First, we design fabricate asynchronous chip called “Speck”, sensing-computing chip. With low processor resting...
Spiking neural networks and neuromorphic hardware platforms that simulate neuronal dynamics are getting wide attention being applied to many relevant problems using Machine Learning. Despite a well-established mathematical foundation for dynamics, there exists numerous software solutions stacks whose variability makes it difficult reproduce findings. Here, we establish common reference frame computations in digital systems, titled Neuromorphic Intermediate Representation (NIR). NIR defines...
In the last few years, spiking neural networks (SNNs) have been demonstrated to perform on par with regular convolutional networks. Several works proposed methods convert a pre-trained CNN Spiking without significant sacrifice of performance. We demonstrate first that quantization-aware training CNNs leads better accuracy in SNNs. One benefits converting is leverage sparse computation SNNs and consequently equivalent at lower energy consumption. Here we propose an optimization strategy train...
The field of neuromorphic computing holds great promise in terms advancing efficiency and capabilities by following brain-inspired principles. However, the rich diversity techniques employed research has resulted a lack clear standards for benchmarking, hindering effective evaluation advantages strengths methods compared to traditional deep-learning-based methods. This paper presents collaborative effort, bringing together members from academia industry, define benchmarks computing:...
Spike generation and routing is typically the most energy-demanding operation in neuromorphic hardware built using spiking neurons. Spiking neural networks running on hardware, however, often use rate-coding where neurons spike rate treated as information-carrying quantity. Rate-coding a highly inefficient coding scheme with minimal information content each spike, which requires transmission of large number spikes. In this paper, we describe an alternative type based temporal neuron activity...
Embedded, continual learning for autonomous and adaptive behavior is a key application of neuromorphic hardware. However, implementations embedded at large scales that are both flexible efficient have been hindered by lack suitable algorithmic framework. As result, most hardware trained off-line on clusters dedicated processors or GPUs transferred post hoc to the device. We address this introducing neural synaptic array transceiver (NSAT), computational framework facilitating matching...
Axonal delays are used in neural computation to implement faithful models of biological systems, and spiking networks solve computationally demanding tasks. While there is an increasing number software simulations that make use axonal delays, only a small fraction currently existing hardware neuromorphic systems supports them. In this paper we demonstrate strategy temporal distributed across multiple Very Large Scale Integration (VLSI) chips. This achieved by exploiting the inherent device...
Spike-Timing-Dependent Plasticity (STDP) is a bio-inspired local incremental weight update rule commonly used for online learning in spike-based neuromorphic systems. In STDP, the intensity of long-term potentiation and depression synaptic efficacy (weight) between neurons expressed as function relative timing pre- post-synaptic action potentials (spikes), while polarity change dependent on order (causality) spikes. Online STDP updates causal acausal spike times are activated at onset post-...
We demonstrate an event-driven Deep Learning (DL) hardware software ecosystem. The user-friendly tools port models from Keras (popular machine learning libraries), automaticaly convert DL to Spiking equivalents, i.e. Convolutional Neural Networks (SCNNs) and run spiking simulations of the converted on emulator for testing prototyping. More importantly, ports onto a novel, ultra-low power, real-time, ASIC SCNN Chip: DynapCNN. An interactive demonstration real-time face recognition system...
Introduction Spiking Neural Networks (SNNs) are gaining significant traction in machine learning tasks where energy-efficiency is of utmost importance. Training such networks using the state-of-the-art back-propagation through time (BPTT) is, however, very time-consuming. Previous work employs an efficient GPU-accelerated backpropagation algorithm called SLAYER, which speeds up training considerably. does not take into account neuron reset mechanism while computing gradients, we argue to be...
Many sounds of ecological importance, such as communication calls, are characterized by time-varying spectra. However, most neuromorphic auditory models to date have focused on distinguishing mainly static patterns, under the assumption that dynamic patterns can be learned sequences ones. In contrast, emergence feature sensitivity through exposure formative stimuli has been recently modeled in a network spiking neurons based thalamo-cortical architecture. The proposed effect lateral and...
Neuromorphic hardware offers an electronic substrate for the realization of asynchronous event-based sensory-motor systems and large-scale spiking neural network architectures. In order to characterize these systems, configure them, carry out modeling experiments, it is often necessary interface them workstations. The software used this purpose typically consists a large monolithic block code which highly specific setup used. While approach can lead integrated hardware/software hampers...
Edge computing solutions that enable the extraction of high-level information from a variety sensors is in increasingly high demand. This due to increasing number smart devices require sensory processing for their application on edge. To tackle this problem, we present vision sensor System Chip (SoC), featuring an event-based camera and low-power asynchronous spiking Convolutional Neural Network (sCNN) architecture embedded single chip. By combining both die, can lower unit production costs...
Stimulus-specific adaptation (SSA) is a phenomenon observed in neural systems which occurs when the spike count elicited single neuron decreases with repetitions of same stimulus, and recovers different stimulus presented. SSA therefore effectively highlights rare events sequences, suppresses responses to repetitive ones. In this paper we present model based on synaptic depression describe its implementation neuromorphic analog very-large-scale integration (VLSI). The hardware system...
In the past recent years several research groups have proposed neuromorphic Very Large Scale Integration (VLSI) devices that implement event-based sensors or biophysically realistic networks of spiking neurons. It has been argued these can be used to build systems, for solving real-world applications in real-time, with efficiencies and robustness cannot achieved conventional computing technologies. order complex systems it is necessary interface VLSI among each other, robotic platforms,...
Event-based dynamic vision sensors provide very sparse output in the form of spikes, which makes them suitable for low-power applications. Convolutional spiking neural networks model such event-based data and develop their full energy-saving potential when deployed on asynchronous neuromorphic hardware. being a nascent field, sensitivity to potentially malicious adversarial attacks has received little attention so far. We show how white-box attack algorithms can be adapted discrete nature...
Spike-timing-dependent plasticity (STDP) incurs both causal and acausal synaptic weight updates, for negative positive time differences between pre-synaptic postsynaptic spike events. For realizing such updates in neuromorphic hardware, current implementations either require forward reverse lookup access to the connectivity table, or rely on memory-intensive architectures as crossbar arrays. We present a novel method using only of permitting memory-efficient implementation. A simplified...
We have recently demonstrated the emergence of dynamic feature sensitivity through exposure to formative stimuli in a real-time neuromorphic system implementing hybrid analog/digital network spiking neurons. This network, inspired by models auditory processing mammals, includes several mutually connected layers with distance-dependent transmission delays and learning form spike timing dependent plasticity, which effects stimulus-driven changes connectivity. Here we present results that...
Several learning rules for synaptic plasticity, that depend on either spike timing or internal state variables, have been proposed in the past imparting varying computational capabilities to Spiking Neural Networkss (SNNs). Due design complications these are typically not implemented neuromorphic devices leaving be only capable of inference. In this work we propose a unidirectional post-synaptic potential dependent rule is triggered by pre-synaptic spikes, and easy implement hardware. We...
Identifying specific keywords in speech is a common task present current mobile devices. In this context, Keyword spotting (KWS) the triggering cue for device to start "listening in" on uttered command. Since can originate at any moment, KWS must be continuously performed. Small-footprint typically designed directly detect entire keyword (as opposed sub-word components, such as phonemes), with low-power solutions relying heavily artificial neural networks (ANNs). As power-efficient...