On-device research index

arXiv research

A locally-built, LLM-digested index of recent arXiv papers in quant finance, geometry/topology, and statistical ML — keyword search served straight from SQLite on this machine.

168,657 papers · 148 categories

Trend · papers per month

count % of corpus

7 results for “NPU”

S3NAS: Fast NPU-aware Neural Architecture Search Methodologycs.LG

S3NAS finds high-accuracy CNN architectures for NPUs in 3 hours.

problem Finding optimal CNN architectures for NPUs with high accuracy and low latency.

method Supernet design, Single-Path NAS, scaling, analytical latency estimation.

result 82.72% top-1 accuracy on ImageNet with 11.66 ms latency.

Neural Architecture Search NPU Convolutional Neural Networks Latency Estimation Image Classification

2020-09-04abs ↗pdf ↗

Neural Power Unitscs.LG

Neural Power Unit (NPU) learns arbitrary power functions on real numbers.

problem Neural Networks struggle with generalizing beyond seen data and arithmetic operations.

method Introduces Neural Power Unit (NPU) that operates on real numbers and learns arbitrary power functions.

result NPU outperforms competitors in accuracy and sparsity on arithmetic datasets and discovers governing equations from data.

Neural Networks Arithmetic Units Real Numbers Power Functions Generalization

2020-06-02abs ↗pdf ↗

Learning Division with Neural Arithmetic Logic Modulescs.NE

Paper tackles division difficulty, proposing new methods to improve accuracy.

problem Division is the most challenging arithmetic operation for both humans and computers.

method Proposes two novel approaches: Neural Reciprocal Unit (NRU) and Neural Multiplicative Reciprocal Unit (NMRU), and improves an existing division module.

result Improves division accuracy from 70.2% to 91.6%.

Neural Arithmetic Logic Modules Division Neural Reciprocal Unit Neural Multiplicative Reciprocal Unit Real Neural Power Unit

2021-10-11abs ↗pdf ↗

Invocation-driven Neural Approximate Computing with a Multiclass-Classifier and Multiple Approximatorscs.LG

Neural approximate computing gains enormous energy-efficiency at the cost of tolerable quality-loss. A neural approximator can map the input data to output while a classifier determines whether the input data are safe to approximate with quality guarantee. However, existing works cannot maximize the invocation of the a…

2018-10-19abs ↗pdf ↗

CompactNet: Platform-Aware Automatic Optimization for Convolutional Neural Networkscs.LG

Convolutional Neural Network (CNN) based Deep Learning (DL) has achieved great progress in many real-life applications. Meanwhile, due to the complex model structures against strict latency and memory restriction, the implementation of CNN models on the resource-limited platforms is becoming more challenging. This work…

2019-05-28abs ↗pdf ↗

Restructurable Activation Networkscs.CV

New activation networks improve model efficiency and performance.

problem Creating hardware-efficient deep learning models.

method Restructurable Activation Networks (RANs) with RAN-explicit and RAN-implicit methods.

result RANs achieve state-of-the-art results with improved hardware efficiency.

activation networks hardware efficiency model scaling deep learning CNNs

2022-08-17abs ↗pdf ↗

Multi-Scale Dequant: Eliminating Dequantization Bottleneck via Activation Decomposition for Efficient LLM Inferencestat.ML

MSD removes dequantization bottleneck in LLM inference by approximating high-precision activations.

problem Dequantization bottleneck in LLM inference on modern AI accelerators.

method MSD decomposes high-precision activations into multiple low-precision components for direct multiplication with quantized weights.

result MSD avoids INT8-to-BF16 weight conversion, reducing dequantization cycles and HBM traffic.

dequantization bottleneck large language models quantization activation decomposition efficient inference

2026-05-13abs ↗pdf ↗