HIGH PERFORMANCE COMPUTING
Exploring machine learning with FPGAs
BILL JENKINS, OF INTEL PROGRAMMABLE SYSTEMS GROUP, LOOKS AT THE ROLE OF FPGA TECHNOLOGY IN MACHINE LEARNING
Machine learning is not a new idea: computer scientists have been developing algorithms
based on models of the way in which the human brain works for decades. But if you’ve recently been tagged in a photo on social media, targeted by ads on an e-commerce site, or curious about just how autonomous vehicles are going to drive themselves, then you’ll know that machine learning is rapidly moving out of the lab and into the real world. This is driving a period of feverish
development in machine learning, within which field-programmable gate arrays (FPGAs) are already taking a starring role. One approach to machine learning that
is currently getting a lot of attention is the convolutional neural network (CNN), widely used in image-processing tasks such as object detection and classification. A CNN is loosely based on the way that human vision works, and so is made up of artificial neurons, each with multiple inputs and outputs. Each neuron takes its inputs, applies a transform to them, and delivers an output value. The value of each input signal can be modified by a weighting factor, and the value that each neuron needs to reach before it will deliver an output can be altered by a biasing factor. The neurons are arranged into a series
of layers: an input layer that takes in the raw data, an output layer that delivers the network’s conclusion, and multiple ‘hidden’ layers in between. These hidden layers create intermediate results that help the network formulate a conclusion to the question it has been trained to answer, such as ‘Is there a cat in this image?’ An untrained neural network begins
with random values for the weights and biasing factors attributed to each neuron, and is trained to give the right answer by being shown many images in which, for
8 Scientific Computing World October/November 2018 Figure 1: Intel PSG FPGAs offer fine-grained and low latency between compute and memory
example, cats have been tagged. Each neuron ’decides’ on its output values based on the input image, passing those values on as inputs to neurons lower down the network’s layers. The final layer of the network delivers the network’s verdict. Then, crucially, the error between the network’s decision and the right answer is calculated and used to propagate a series of adjustments to the weighting and bias factors for each neuron back through each layer of the network, from the output to the input. Repeat this process often enough, with a large enough dataset of accurately tagged images, and the network will adapt its weights and biases so that it can ‘recognise’, in this example, cats within an untagged image. This training process is computationally
intensive – think days or weeks of dedicated computational effort in large data centres. However, once the right set of weights and biases have emerged from the training process, applying them, a process known as inferencing, can be
done with much less processing power and, often, using much less precise number representations. Object detection and classification is such a useful skill that there’s an annual competition, the ImageNet Large Scale Visual Recognition Challenge, to test the accuracy of new algorithms using a standard dataset of tagged images. This closely fought contest has seen rapid improvements in accuracy using new algorithms or fresh adaptations of
g
“One approach to machine learning that is currently getting a lot of attention is the convolutional neural network (CNN), widely used in image- processing tasks such as object detection and classification”
@scwmagazine |
www.scientific-computing.com
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36