julialogo_thumb

Mocha.jl: Deep Learning for Julia

Deep learning is becoming extremely popular due to several breakthroughs in various well-known tasks in artificial intelligence. For example, at the ImageNet Large Scale Visual Recognition Challenge, the introduction of deep learning algorithms into the challenge reduced the top-5 error by 10% in 2012. Every year since then, deep learning models have dominated the challenges, significantly reducing the top-5 error rate every year (see Figure 1). In 2015, researchers have trained very deep networks (for example, the Google “inception” model has 27 layers) that surpass human performance.

Figure 1: The top-5 error rate in the ImageNet Large Scale Visual Recognition Challenge has been rapidly reducing since the introduction of deep neural networks in 2012.
Figure 1: The top-5 error rate in the ImageNet Large Scale Visual Recognition Challenge has been rapidly reducing since the introduction of deep neural networks in 2012.

Moreover, at this year’s Computer Vision and Pattern Recognition (CVPR) conference, deep neural networks (DNNs) were being adapted to increasingly more complicated tasks. For example, in semantic segmentation, instead of predicting a single category for a whole image, a DNN is trained to classify each pixel in the image, essentially producing a semantic map indicating every object and its shape and location in the given image (see Figure 2).

Figure 2: An example of generating a semantic map by classifying each pixel in a source image. Source: Shuai Zheng et al. Conditional Random Fields as Recurrent Neural Networks.
Figure 2: An example of generating a semantic map by classifying each pixel in a source image. Source: Shuai Zheng et al. Conditional Random Fields as Recurrent Neural Networks.

Continue reading

atcom_logo

ATCOM: Real-Time Enhancement of Long-Range Imagery

Imaging over long distances is important for many defense and commercial applications. High-end ground-to-ground, air-to-ground, and ground-to-air systems now routinely image objects several kilometers to several dozen kilometers away; however, this increased range comes at a price. In many scenarios, the limiting factor becomes not the quality of your camera but the atmosphere through which the light travels to the camera. Dynamic changes in the atmospheric density between the camera and object impart time-variant distortions resulting in loss of contrast and detail in the collected imagery (see Figure 1 and Figure 2).

Figure 1: Wavefront distortion by the atmosphere causes the image captured to be blurred. The effect becomes more pronounced when imaging over long distances or through more turbulent atmospheric conditions.
Figure 1: Wavefront distortion by the atmosphere causes the image captured to be blurred. The effect becomes more pronounced when imaging over long distances or through more turbulent atmospheric conditions.

Several approaches have been developed to combat this effect that can be roughly divided into two categories: hardware-based and signal processing approaches. The primary hardware technique is adaptive optics (AO), an approach favored by the astronomical community to observe stellar phenomena. AO techniques generally employ a deformable mirror to correct the incoming wavefront before it is captured on the sensor. While this provides the ability to improve imagery to account for distortions, the equipment required is fragile and expensive and is therefore not suitable for many applications. In contrast, signal processing techniques are limited only by the computational hardware they run on. In our case, we have leveraged the processing power of GPUs to achieve the performance necessary for real-time processing of high-definition video. Thanks to modern GPUs, we are now able to process live 720p video streams at over 30 fps, as the video below shows.

Continue reading

gpu_computing_spotlight_358x230

Harnessing the Caffe Framework for Deep Visualization

Jeff Clune, Lab Director, Evolving Intelligence Laboratory at The University of Wyoming.
Jeff Clune, Lab Director, Evolving Intelligence Laboratory at The University of Wyoming.

The need to train their deep neural networks as fast as possible led the Evolving Artificial Intelligence Laboratory at the University of Wyoming to harness the power of NVIDIA Tesla GPUs starting in 2012 to accelerate their research.

“The speedups GPUs provide for training deep neural networks are well-documented and allow us to train models in a week that would otherwise take months,” said Jeff Clune, Assistant Professor, Computer Science Department and Director of the Evolving Artificial Intelligence Laboratory. “And algorithms continuously improve. Recently, NVIDIA’s cuDNN technology allowed us to speed up our training time by an extra 20% or so.”

Clune’s Lab, which focuses on evolving artificial intelligence with a major focus on large-scale, structurally organized neural networks, has garnered press from some of the largest media outlets, including BBC, National Geographic, NBC News, The Atlantic and featured on the cover of Nature in May 2015.

[The following video shows off work from the Evolving AI Lab on visualizing deep neural networks. Keep reading to learn more about this work!]

For this Spotlight interview, I had the opportunity to talk with Jeff Clune and two of his collaborators, Anh Nguyen, a Ph.D. student at the Evolving AI Lab and Jason Yosinski, a Ph.D. candidate at Cornell University.

Brad: How are you using deep neural networks (DNNs)?

We have many research projects involving deep neural networks. Our Deep Learning publications to date involve better understanding DNNs. Our lab’s research covers: Continue reading

matlab_logo

High-Performance MATLAB with GPU Acceleration

In this post, I will discuss techniques you can use to maximize the performance of your GPU-accelerated MATLAB® code. First I explain how to write MATLAB code which is inherently parallelizable. This technique, known as vectorization, benefits all your code whether or not it uses the GPU. Then I present a family of function wrappers—bsxfunpagefun, and arrayfun—that take advantage of GPU hardware, yet require no specialist parallel programming skills. The most advanced function, arrayfun, allows you to write your own custom kernels in the MATLAB language.

If these techniques do not provide the performance or flexibility you were after, you can still write custom CUDA code in C or C++ that you can run from MATLAB, as discussed in our earlier Parallel Forall posts on MATLAB CUDA Kernels and MEX functions.

All of the features described here are available out of the box with MATLAB and Parallel Computing Toolbox™.

Mobile phone signal strength example

Throughout this post, I will use an example to help illustrate the techniques. A cellular phone network wants to map its coverage to help plan for new antenna installations. We imagine an idealized scenario with M = 25 cellphone masts, each H = 20 meters in height, evenly spaced on an undulating 10km x 10km terrain. Figure 1 shows what the map looks like.

Figure 1: 3D plot of the map data and antenna positions
Figure 1: 3D plot of the map data and antenna positions

On the GPU, in the following listing we define a number of variables including:

  • map: An N x 3 height field in a 10km x 10km grid (N = 10,000);
  • masts: An M x 3 array of antenna positions, at height H;
  • AntennaDirection: A 3 x M array of vectors representing the orientation of each antenna.

Continue reading

ImageNetLogo

NVIDIA and IBM Cloud Support ImageNet Large Scale Visual Recognition Challenge

This year’s ImageNet Large Scale Visual Recognition Challenge (ILSVRC) is about to begin. Every year, organizers from the University of North Carolina at Chapel Hill, Stanford University, and the University of Michigan host the ILSVRC, an object detection and image classification competition, to advance the fields of machine learning and pattern recognition. Competitors are given more than 1.2 million images from 1,000 different object categories, with the objective of developing the most accurate object detection and classification approach. After the data set is released, teams have roughly three months to create, refine and test their approaches. This year, the data will be released on August 14, and the final result submission deadline is November 13. Read on to learn how competing teams can get free access to the latest GPU-accelerated cloud servers from NVIDIA and IBM Cloud.

Figure 1. Examples of ImageNet images demonstrating classification with localization.
Figure 1. Examples of ImageNet images demonstrating classification with localization.

About the Challenge

Teams compete annually to develop the most accurate recognition systems, and each year the sub-tasks are more complex and challenging. Since none of the algorithms developed to date are able to classify all images correctly 100% of the time, classification accuracy is measured in terms of error rates. In this way, the winning submission will exhibit the lowest overall percentage of incorrectly classified images. When ILSVRC started in 2010, teams had a single assignment: classify the contents of an image (identify objects) and provide a list of the top five classifications with their respective probabilities. Since then, the complexity of assignments has grown as the organizers have added sub-tasks for classification with localization and with detection (identifying objects and specifying their locations with bounding boxes). Specifying bounding boxes is important because most pictures contain multiple objects. To illustrate the requirements of this type of task, Figure 1 shows two sample ImageNet training images annotated with their object categories and bounding boxes. Continue reading

labellio_logo_thumb

Labellio: Scalable Cloud Architecture for Efficient Multi-GPU Deep Learning

Labellio is the world’s easiest deep learning web service for computer vision. It aims to provide a deep learning environment for image data where non-experts in deep learning can experiment with their ideas for image classification applications. Watch our video embedded here to see how easy it is.

The challenges in deep learning today are not just in configuring hyperparameters or designing a suitable neural network structure; they are also in knowing how to prepare good training data to fine-­tune a well­-known working classification model, and in knowing how to set up a computing environment without hassle. Labellio aims to address these problems by supporting the deep learning workflow from beginning to end.

The Labellio training data collector lets users upload images, or downloads images from the internet from specified URL lists or keyword searches on external services like Flickr and Bing. While some users have their own domain­-specific training data in hand already, others don’t. Labellio’s data ingestion capabilities help both types of users start building their own deep learning models immediately. It is extremely important to assign the correct class to each training image to build a more accurate classification model, so Labellio’s labelling capability helps users cleanse input training data.

Labellio helps users build and manage multiple classification models. Creating a classification model is typically not just a one-time operation; you need trial and error to develop the most accurate model by replacing some of the training data. Continue reading

GPUProTip_179x115

Voting and Shuffling to Optimize Atomic Operations

2iSome years ago I started work on my first CUDA implementation of the Multiparticle Collision Dynamics (MPC) algorithm, a particle-in-cell code used to simulate hydrodynamic interactions between solvents and solutes. As part of this algorithm, a number of particle parameters are summed to calculate certain cell parameters. This was in the days of the Tesla GPU architecture (such as GT200 GPUs, Compute Capability 1.x), which had poor atomic operation performance. A linked list approach I developed worked well on Tesla and Fermi as an alternative to atomic adds but performed poorly on Kepler GPUs. However, atomic operations are much faster on the Kepler and Maxwell architectures, so it makes sense to use atomic adds.

These types of summations are not limited to MPC or particle-in-cell codes, but, to some extent, occur whenever data elements are aggregated by key. For data elements sorted and combined by key with a large number of possible values, pre-combining elements with the same key at warp level can lead to a significant speed-up. In this post, I will describe algorithms for speeding up your summations (or similar aggregations) for problems with a large number of keys where there is a reasonable correlation between the thread index and the key. This is usually the case for elements that are at least partially sorted. Unfortunately, this argument works in both directions: these algorithms are not for you if your number of keys is small or your distribution of keys is random.  To clarify: by a “large” number of keys I mean more than could be handled if all bins were put into shared memory.

Note that this technique is related to a previously posted technique called warp-aggregated atomics by Andrey Adinetz, and also to the post Fast Histograms Using Shared Atomics on Maxwell by Nikolay Sakharnykh. The main difference here is that we are aggregating many groups, each designated by a key (to compute a histogram, for example). So you could consider this technique “warp-aggregated atomic reduction by key”. Continue reading

quantalea_logo

.NET Cloud Computing with Alea GPU

Cloud computing is all about making resources available on demand, and its availability, flexibility, and lower cost has helped it take commercial computing by storm. At the Microsoft Build 2015 conference in San Francisco Microsoft revealed that its Azure cloud computing platform is averaging over 90 thousand new customers per month; contains more than 1.4 million SQL databases being used by hosted applications, and over 50 trillion storage objects; and has 425 million users in the Active Directory system. Microsoft also said that the number of registered developers for Visual Studio Online services increased from 2 to 3 million in less than half a year.

Clearly there is increasing demand for GPU cloud computing. The number of use cases for GPUs in the cloud is growing, and there are a number of ways that GPU cloud computing may be useful to you, including

  • If you need GPUs but don’t have access to them locally;
  • If you are a data scientist and need to scale your machine learning code with multiple GPUs;
  • If you are a developer and want to benchmark your applications on a more recent GPU with the latest CUDA technology; or
  • If you want to use GPUs for video-creation services, 3D visualizations or game streaming.

Cloud computing services are provided at different levels: IaaS (Infrastructure as a Service), PaaS (Platform as a Service) and SaaS (Software as a Service). In this post I will consider two GPU cloud computing use cases, and walk you through setting up and running a sample cloud application using Alea GPU. The first one uses GPUs through IaaS to acclerate .NET applications with Alea GPU on Linux or Windows. The second use case shows how to build a PaaS for GPU computing with Alea GPU and MBrace.  To learn more about Alea GPU, please read my earlier Parallel Forall post. Continue reading

gpu_computing_spotlight_358x230

Leveraging OpenACC to Compute High-Accuracy Molecular Energies

Janus Juul EriksenFor this interview, I reached out to Janus Juul Eriksen, a Ph.D. fellow at Aarhus University in Denmark. Janus is a chemist by trade without any formal education in computer science; but he is getting up to 12x speed-up compared to his CPU-only code after modifying less than 100 lines of code with one week of programming effort.

How did he do this? He used OpenACC.

OpenACC is a simple, powerful and portable approach for researchers and scientists who need to rapidly boost application performance for faster science while minimizing programming. With OpenACC, the original source code is kept intact, making the implementation intuitively transparent and leaving most of the hard work to the compiler.

NVIDIA recently announced the new OpenACC Toolkit, an all-in-one suite of parallel programming tools, that helps researchers and scientists quickly accelerate applications.

“OpenACC is much easier to learn than OpenMP or MPI. It makes GPU computing approachable for domain scientists,” says Janus. “Our initial OpenACC implementation required only minor efforts, and more importantly, no modifications of our existing CPU implementation.”

Figure 1: The active site of the Bacterial Leucine Transporter (LeuT) protein, on which the LSDalton team has been doing a number of high-accuracy calculations. The LeuT protein belongs to a family of transporters that are linked to neurological diseases.
Figure 1: The active site of the Bacterial Leucine Transporter (LeuT) protein, on which the LSDalton team has been doing a number of high-accuracy calculations. The LeuT protein belongs to a family of transporters that are linked to neurological diseases.

Janus is part of the research team developing the quantum chemistry code LSDalton, a massively parallel and linear-scaling program for the accurate determination of energies and other molecular properties for large molecular systems.

In need of speed, the LSDalton team was awarded an INCITE allocation which gave them access to Oak Ridge National Laboratory’s Titan supercomputer. With this, they needed to find a way to use the power of the supercomputer: enter OpenACC. Demonstrating success on Titan with their GPU-accelerated code, they were recently one of 13 application code projects selected to join the Center for Accelerated Application Readiness (CAAR) program. This means they will be among the first applications to run on Summit, the new supercomputer debuting in 2018 which will deliver more than five times the computational performance of Titan’s 18,688 nodes.

This access will enable the research team to simulate larger molecular structures at higher accuracy, ultimately accelerating discoveries in materials and quantum chemistry.

Continue reading

Theano Logo

Introduction to Neural Machine Translation with GPUs (part 3)

Note: This is the final part of a detailed three-part series on machine translation with neural networks by Kyunghyun Cho. You may enjoy part 1 and part 2.

In the previous post in this series, I introduced a simple encoder-decoder model for machine translation. This simple encoder-decoder model is excellent at English-French translation. However, in this post I will briefly discuss the weakness of this simple approach, and describe a recently proposed way of incorporating a soft attention mechanism to overcome the weakness and significantly improve the translation quality.

Furthermore, I will present some more recent works that utilize this neural machine translation approach to go beyond machine translation of text, such as image caption generation and video description generation. I’ll finish the blog series with a brief discussion of future research directions and a pointer to the open source code implementing these neural machine translation models.

The Trouble with Simple Encoder-Decoder Architectures

In the encoder-decoder architecture, the encoder compresses the input sequence as a fixed-size vector from which the decoder needs to generate a full translation. In other words, the fixed-size vector, which I’ll call a context vector, must contain every single detail of the source sentence. Intuitively, this means that the true function approximated by the encoder has to be extremely nonlinear and complicated. Furthermore, the dimensionality of the context vector must be large enough that a sentence of any length can be compressed.

In my paper “On the Properties of Neural Machine Translation: Encoder-Decoder Approaches” presented at SSST-8, my coauthors and I empirically confirmed that translation quality dramatically degrades as the length of the source sentence increases when the encoder-decoder model size is small. Together with a much better result from Sutskever et al. (2014), using the same type of encoder-decoder architecture, this suggests that the representational power of the encoder needed to be large, which often means that the model must be large, in order to cope with long sentences (see Figure 1).

 Figure 1: Dramatic drop of performance w.r.t. the length of sentence with a small encoder-decoder model.
Figure 1: Dramatic drop of performance w.r.t. the length of sentence with a small encoder-decoder model.

Of course, a larger model implies higher computation and memory requirements. The use of advanced GPUs, such as NVIDIA Titan X, indeed helps with computation, but not with memory (at least not yet). The size of onboard memory is often limited to several Gigabytes, and this imposes a serious limitation on the size of the model. (Note: it’s possible to overcome this issue by using multiple GPUs while distributing a single model across those GPUs, as shown by Sutskever et al. (2014). However, let’s assume for now that we have access to a single machine with a single GPU due to space, power and other physical constraints.)

Then, the question is “can we do better than the simple encoder-decoder based model?” Continue reading