A5084220043- Stochastic Gradient Optimization Techniques
- Advanced Neural Network Applications
- Advanced Memory and Neural Computing
- Error Correcting Code Techniques
- Neural Networks and Applications
- Transportation Systems and Safety
- Ferroelectric and Negative Capacitance Devices
- Cryptographic Implementations and Security
- Green IT and Sustainability
- Chaos-based Image/Signal Encryption
- Anomaly Detection Techniques and Applications
- Network Packet Processing and Optimization
- Mobile Ad Hoc Networks
- Machine Learning and ELM
- Coding theory and cryptography
- Age of Information Optimization
- Network Time Synchronization Technologies
- CCD and CMOS Imaging Sensors
- Mining and Industrial Processes
- Power Line Communications and Noise
- Advanced Data Storage Technologies
- Radiation Effects in Electronics
- Engine and Fuel Emissions
- Brain Tumor Detection and Classification
- Real-Time Systems Scheduling
University of California, Los Angeles
2020-2025
Amazon (United States)
2023
As privacy and latency requirements force a move towards edge Machine Learning inference, resource constrained devices are struggling to cope with large computationally complex models. For Convolutional Neural Networks, those limitations can be overcome by taking advantage of enormous data reuse opportunities amenability reduced precision. To do that however, level compute density unattainable for conventional binary arithmetic is required. Stochastic Computing deliver such density, but it...
Deep learning has grown in capability and size recent years, prompting research on alternative computing methods to cope with the increased compute cost. Stochastic (SC) promises higher efficiency its compact units, but accuracy issues have prevented wide adoption, accuracy-improving techniques sacrificed runtime or training performance. In this work, we propose extended range SC—Range-Extended SC Accumulation deal of SC. By modifying functionality OR-based accumulation, increase computation...
We present the first programmable and precision-tunable Stochastic Computing (SC) neural network (NN) inference accelerator. The use of SC makes it possible to achieve multiply-accumulate (MAC) density 38.4k MAC/mm2, enabling a level spatial data reuse unachievable conventional, fixed-point architectures. This extensive amortizes cost conversion reduces number memory accesses, which can otherwise consume significant energy latency. Our accelerator is stand-alone architecture, with custom...
Stochastic computing (SC) has seen a renaissance in recent years as means for machine learning acceleration due to its compact arithmetic and approximation properties. Still, SC accuracy remains an issue, with prior works either not fully utilizing the computational density or suffering from significant losses. In this work, we propose GEO - Generation Execution Optimized Computing Accelerator Neural Networks, which optimizes stream generation execution components of SC, bridges gap between...
As the adoption of Neural Networks continues to proliferate different classes applications and systems, edge devices have been left behind. Their strict energy storage limitations make them unable cope with sizes common network models. While many compression methods such as precision reduction sparsity proposed alleviate this, they don’t go quite far enough. To push size its absolute limits, we combine binarization in Pruned-Permuted-Packed XNOR (3PXNet), which can be efficiently implemented...
Stochastic computing (SC) has recently emerged as a promising method for efficient machine learning acceleration. Its high compute density, affinity with dense linear algebra primitives, and approximation properties have an uncanny level of synergy the deep neural network computational requirements. However, there is conspicuous lack works trying to integrate SC hardware sparsity awareness, which brought significant performance improvements conventional architectures. In this work, we...
High compute density improves the data reuse and is key to reducing off-chip memory access achieving high energy efficiency in ML accelerators. Compute-in-Memory (CIM) promises but requires ADCs, DACs that add macro's area [1] [2] limiting its density. Besides, CIM's analog sensitive process variability mismatches. The transistor nonlinearity also significantly degrades accuracy. Stochastic Computing (SC), which represents numbers as probability of 1s random binary streams, a digital-compute...
Researchers have long touted a vision of the future enabled by proliferation internet-of-things devices, including smart sensors, homes, and cities. Increasingly, embedding intelligence in such devices involves use deep neural networks. However, their storage processing requirements make them prohibitive for cheap, off-the-shelf platforms. Overcoming those is necessary enabling widely-applicable devices. While many ways making models smaller more efficient been developed, there lack...
Quantization is spearheading the increase in performance and efficiency of neural network computing systems making headway into commodity hardware. We present SWIS - Shared Weight bIt Sparsity, a quantization framework for efficient inference acceleration delivering improved storage compression through an offline weight decomposition scheduling algorithm. can achieve up to 54.3% (19.8%) point accuracy improvement compared truncation when quantizing MobileNet-v2 4 (2) bits post-training (with...