A5084842768- Advanced Memory and Neural Computing
- Ferroelectric and Negative Capacitance Devices
- Neural Networks and Reservoir Computing
- CCD and CMOS Imaging Sensors
- Neural dynamics and brain function
- Neural Networks and Applications
- Advanced Vision and Imaging
- Cellular Automata and Applications
- Photoreceptor and optogenetics research
- Human Pose and Action Recognition
- Video Analysis and Summarization
University of Tennessee at Knoxville
2022-2025
Knoxville College
2022-2025
University of Tennessee System
2023
The cart-pole application is a well-known control that often used to illustrate reinforcement learning algorithms with conventional neural networks. An implementation of the from OpenAI Gym ubiquitous and popular. Spiking networks are basis brain-based, or neuromorphic computing. They attractive, especially as agents for applications, because their very low size, weight power requirements. We motivated help researchers in computing be able compare work common benchmarks, this paper we...
Abstract Neuromorphic computing is a novel style of that features low-power spiking neural networks as the main compute components. It an event-driven computational paradigm naturally pairs with event-based cameras and their asynchronous event output. In this work, we present NeuroPong, closed-loop neuromorphic hardware system composed camera, system, Atari 2600 console. The facilitates implementation network agents capable playing games in real time using camera capture input. We perform...
The synapse is a key element of neuromorphic computing in terms efficiency and accuracy. In this paper, an optimized current-controlled memristive circuit proposed. Our proposed demonstrates reliability the face process variation inherent stochastic behavior memristors. Up to 82% energy optimization can be seen during SET operation over prior work. addition, READ shows up 54% savings. approach also provides more reliable programming traditional methods. This design demonstrated with 4-bit...
In this work, we address the problem of training a neuromorphic agent to work on data from event-based cameras. Although camera is much sparser than standard video frames, sheer number events can make observation space too complex effectively train an agent. We construct multiple networks that downsample so as more effective. then perform case study play Atari Pong game by converting each frame and downsampling them. The final network combines both discuss some practical considerations well.
Neuromorphic computing is a leading option for non von-Neumann architectures. With it, neural networks are developed that derive architectural inspiration from how the brain operates with neurons, synapses, and spikes. These often implemented in either software or hardware based neuroprocessors designed to handle specific tasks efficiently. Even if hardware, emulation instrumental determining worthwhile features capabilities of architecture. In this work two novel introduced: software-based...
A challenge associated with effectively using spiking neuromorphic systems is how to communicate data and from the implementation. Unless a or event-based sensing system used, has be converted into spikes processed as input by system. The output produced have turned back value decision. There are variety of commonly used encoding approaches, such rate coding, temporal population well several voting first-to-spike. However, it not clear which most appropriate approach use whether choice...
Spiking neural networks are powerful computational elements that pair well with event-based cameras (EBCs). In this work, we present two spiking network architectures process events from EBCs: one isolates and filters out based on their speeds, another clusters the DBSCAN algorithm.
Recurrent, sparse spiking neural networks have been explored in contexts such as reservoir computing and winner-take-all networks. However, we believe there is the opportunity to leverage recurrence for other tasks, particularly control. In this work, show that evolved recurrent perform significantly better than feed-forward counterparts. We give two examples of types are demonstrate they highly unlike traditional, more structured used deep learning literature.
DBSCAN is an algorithm that performs clustering in the presence of noise. In this paper, we provide two constructions allow to be implemented neuromorphically, using spiking neural networks. The first construction termed "flat," resulting large networks compute quickly, five timesteps. Moreover, pipelining, so a new calculation may performed every timestep. second "systolic", and generates much smaller networks, but requires inputs spiked over several timesteps, column by column. We precise...
The cart-pole application is a well-known control that often used to illustrate reinforcement learning algorithms with conventional neural networks. An implementation of the from OpenAI Gym ubiquitous and popular. In this paper, we explore using as benchmark for spiking We propose four parameter settings scale in difficulty, particular beyond default which do not pose difficult test AI agents. achievement levels agents are trained on these settings. Next, perform an experiment employs its...
Event-based cameras are with high dynamic range that measure changes in light intensity at each pixel instead of capturing frames like traditional cameras. There several event-based camera simulation software libraries can convert videos or collections to a stream simulated events. To the authors' knowledge, exception vehicle control projects, there no simulate activity online during an agent's training phase on other various applications. This work introduces ALE_EBC, wrapper around Arcade...
Event-based cameras and classification datasets pair nicely with neuromorphic computing. Furthermore, it is attractive from a SWaP perspective to have fully pipeline event-based camera output instead of having preprocess the data prior classification. In this work, we examine how two observation space reduction techniques impact performance on DVSGesture dataset. The can be implemented as spiking neural networks so that no preprocessing required, instead, only routing events proper input...