CUDA GPU Support for Enterprise Deep Learning

[fa icon="calendar"] Jun 7, 2016 5:33:07 PM / by Edward Junprung

Edward Junprung


Deep learning, an advanced subset of machine learning, is probably the hottest trend in AI, and for good reason. Deep nets are flexible, powerful learning algorithms that can solve some of the world’s hardest problems.


In March 2016, AlphaGo (an AI developed by Google DeepMind) defeated grandmaster Lee Sedol in Go, the most complex board game that AI has solved. This demonstrated how far computers had come, since they now surpass human experts in their ability to find solutions to very difficult problems. 
Deep learning’s record-breaking accuracy has important implications.
  1. Enterprises that apply deep learning will outperform their competitors in areas such as fraud detection and e-commerce recommendations. Fraud costs companies hundreds of billions of dollars annually, so being able to adapt and stop fraud (without human intervention) means enormous cost savings
  2. The Internet age has produced more data than humans can cope with. For example, the popularity of cloud computing makes it difficult for sysadmins to effectively manage tens of thousands of servers using existing tools. Deep neural networks can automatically learn from signals and predict the likelihood of server failure before it happens, enabling sysadmins to rebalance their workloads, maintain uptime and save money.
  3. Researchers are using deep learning to tackle problems that were once deemed impossible, with new breakthroughs published every month. Deep learning is the catalyst for new insights in diverse fields such as neural science, medicine, physics, genetics and augmented reality. 


Neural networks were invented more than 50 years ago, so they’re hardly a new field. But several important developments in the last 10 years have made deep learning capable of new feats.
  1. With the growth of the internet, there is lot more data available. With more data, you can train neural networks to produce more accurate models of the world. That data is being streamed in time series from IoT devices, published as text on social media, and created as logs on cloud computing instances. Big data changed everything. 
  2. Advances in algorithms such as deep convolutional neural nets mean we can perform tasks like image recognition with record-breaking accuracy. These new powers also extend to automatic machine translation, predictive analytics and sound recognition.
  3. Hardware advances in GPUs and distributed cloud computing have put incredible computing power at our disposal. This level of computing power is necessary to train deep neural nets.
The combination of the three factors above has made deep learning practical to businesses that wish to leverage its power. 


GPUs are massively parallel calculators, which makes them well adapted for vector and matrix operations: the foundation of neural networks. While originally designed to process graphics, GPUs ability to perform mathematical operations at scale makes them ideally suited to neural network training, far outstripping many CPUs. 
This speed is crucial because large companies, which are swimming in data, need to experiment with deep learning architectures, adjusting them until they are properly tuned. This tuning and training stage can be time-consuming without GPUs. Deep nets can take a lot of tuning. To avoid waiting weeks for nets to train, you need to run on fast, distributed hardware.

Deeplearning4j is the only commercial-grade, open-source, distributed deep-learning library written for Java and Scala. Integrated with Hadoop and Spark, DL4J is designed to be used in business environments on distributed GPUs and CPUs.
This is important because most large companies have bet their stacks on the JVM. The toughest part about implementing deep learning is integrating it with the web of business applications and libraries that companies already use. Since DL4J builds on and works with existing JVM frameworks, we make it easy to incorporate deep learning into enterprise-scale applications. If they’re going to apply deep learning, it’s going to be with Deeplearning4j.


Fraud Prevention

Fraud is one of the hardest problems to solve in finance and online payment processing. Scammers constantly adapt and exploit holes in online fraud detection, and fraud detection algorithms need to be updated constantly. But hard-coded, rules-based fraud detection can’t keep up: there are not enough data scientists in the world to anticipate every hack. Deep learning can identify anomalies automatically and evolve in response to rapidly changing online behavior — it doesn’t need hard-coded rules to know how to identify fraud. It does so without human intervention, stopping fraud sooner and more frequently, because of its record-breaking accuracy. 
Server Management

The explosion of cloud computing has made server management an increasingly important, and difficult, task. Modern data centers can contain hundreds of thousands of servers, making management formidable for sysadmins using existing tools. Deep learning allows sysadmins to predict server failure and prevent downtime before it happens, by redistributing workloads pre-emptively.
Network Intrusion Detection
By monitoring the data packets that enter and exit a data center, deep learning can identify packets that exhibit unusual traits to flag abnormal activity and possible cybersecurity breaches. Similarly to fraud, online criminals constantly adapt and exploit holes in network intrusion detection software — cumulatively costing companies over $400 billion a year. 
E-Commerce Recommendation System
E-commerce giants like Amazon use deep learning to create personalized recommendations for each buyer based on browsing and buying behavior. Research firm MarketingSherpa determined that these recommendations were responsible for 11.5% of all revenue. 


Deep learning is giving enterprises the opportunity to transform their businesses with AI. 
  • Boost existing models using cutting edge tools that has been proven elsewhere to be radically better.
  • “Ready to go” deep learning that fits within the enterprise stack.  

Topics: Non-technical

Edward Junprung

Written by Edward Junprung