Анатомия конволюционных уровней высокопроизводительного углубленного изучения на архитектуре SIMD

Познакомьтесь с новым методом построения конволюционных уровней для архитектуры x86 (например, для процессоров Intel® Xeon Phi™ и Intel® Xeon®), который обеспечивает высокоэффективную работу уровней в мультикомпьютерных системах.


Результаты поиска: 26

Distributed Deep Reinforcement Learning

This study explores the scalability of deep reinforcement learning and shows how you can use 768 CPU cores to cut training time down from 10 hours to 21 minutes—enough time to master multiple classic Atari* games.

HeNet feature visualization

HeNet: A Deep Learning Approach on Intel® Processor Trace for Effective Exploit Detection

This paper presents a novel technique for detecting malware through classifying hardware-generated control flow traces using a deep neural network.

Caffe Logo

Highly Efficient 8-Bit, Low-Precision Inference of Convolutional Neural Networks

Get a detailed explanation of Intel® Optimization for Caffe, a deep learning framework supportive of 8-bit models to be used on Intel® Xeon® Scalable processors.

linear classifiers versus non-linear classifiers table

Integrate Multiplicative Features into Supervised Distributional Methods for Lexical Entailment

By adding multiplicative features into word-pair representations, linear and non-linear classifiers were found to have increased performance.

image segmentation examples

Interactive Image Segmentation with Latent Diversity

Bridging the gap between image segmentation and user-intended regions for applications, this method outperforms all prior approaches.

Learn to See in the Dark

This research uses convolutional neural networks to teach machines how to convert images taken in the dark into images in full light.

question answering example

Knowledge Memory Networks for Visual Question Answering

Working to overcome visual question answering—one of the most difficult tasks for AI—this network outperforms current methods in the task of answering textual questions about an image.

example of a matryoshka network

Matryoshka Networks: Predict 3D Geometry via Nested Shape Layers

Read about a novel 2D encoding method for 3D geometry that reconstructs 3D shapes from an image by using image encoders, voxel tube, and shape layer decoder.

Robot in crowd of people

Motion Planning Among Dynamic, Decision-Making Agents with Deep Reinforcement Learning

This research proposes an algorithm that improves performance of robotic navigation through crowds of people using deep learning instead of 3D Lidar.

graph of noun relations

Paraphrase to Explicate: Revealing Implicit Noun-Compound Relations

Using a neural model to generalize word phrases, this technique improves the performance of natural language processing (NLP) classification and paraphrasing tasks.

results of image synthesis techniques

Semi-Parametric Image Synthesis

Using encoder and decoder neural networks, this research details a new approach to creating photo-realistic synthetic images from an input label layout.

train loss without/without normalization table

Single Image Reflection Separation with Perceptual Losses

Using convolutional neural networks, this new approach outperforms previous methods in identifying and removing reflections within images.