As reported by Nature, a new AI competition will be occurring soon, the MineRL competition, which will encourage AI engineers and coders to create programs capable of learning through observation and example. The test case for these AI systems will be the highly popular crafting and survival video game Minecraft.
Artificial intelligence systems are have seen some recent impressive accomplishments when it comes to video games. Just recently an AI beat out the best human players in the world at the strategy game StarCraft II. However, StarCraft II has definable goals that are easier to break down into coherent steps that an AI can use to train. A much more difficult task is for an AI to learn how to navigate a large, open-world sandbox game like Minecraft. Researchers are aiming to help AI programs learn through observation and example, and if they are successful they could substantially reduce the amount of processing power needed to train an artificial intelligence program.
The participants in the competition will have four days to create an AI that will be tested with Minecraft, taking up to eight million steps to train their AI. The goal of the AI is to find a diamond within the game by digging. Eight million steps of training is a much shorter time span than the amount of time needed to train powerful AI models these days, so the participants in the competition need to engineer methods that drastically improve over current training methods.
The approaches being used by the participants are based on a type of learning called imitation learning. Imitation learning stands in contrast with reinforcement learning, which is a popular method for training sophisticated systems like robotic arms in factories or the AIs capable of beating human players at StarCraft II. The primary drawback to reinforcement learning algorithms is the fact that they require immense computer processing power to train, relying on hundreds or even thousands of computers linked together to learn. By contrast, imitation learning is a much more efficient and less computationally expensive method of training. Imitation learning algorithms endeavor to mimic how humans learn by observation.
William Guss, a PhD candidate in deep-learning theory at Carnegie Mellon University explained to Nature that getting an AI to explore and learn patterns in an environment is a tremendously difficult task, but imitation learning provides the AI with a baseline of knowledge, or good prior assumptions, about the environment. This can make training an AI much quicker in comparison to reinforcement learning.
Minecraft serves as a particularly useful training environment for multiple reasons. One reason is that Minecraft allows players to use simple building blocks to create complex structures and items, and the many steps needed to create these structures serve as tangible markers of progress that researchers can use as metrics. Minecraft is also extremely popular, and because of this, it is comparatively easy to gather training data. The organizers of the MineRL competition recruited many Minecraft players to demonstrate a variety of tasks like creating tools and braking apart blocks. By crowdsourcing the generation of data, researchers were able to capture 60 million examples of actions that could be taken in the game. The researchers gave approximately 1000 hours of video to the competition teams.
Use the knowledge that humans have built up, says Rohin Shah, Ph.D. candidate in computer science at the University of California, Berkeley explained to Nature that this competition is likely the first to focus on using the knowledge that humans have already generated to expedite the training of AI.
Guss and the other researchers are hopeful that the contest could have results with implications beyond Minecraft, giving rise to better imitation learning algorithms and inspiring more people to consider imitation learning as a viable form of training an AI. The research could potentially help create AIs that are better capable of interacting with people in complex, changing environments.
Amazon Announces DeepComposer and Other AI Technology
Amazon’s annual re:Invent conference in Las Vegas began this week with three major AI announcements. The company presented the public with Transcribe Medical, SageMaker Operators for Kubernetes, and DeepComposer.
What is being called the biggest announcement of the three, Transcribe Medical is the newest edition to the company’s transcribe speech recognition service. It will transcribe medical speech for primary care. The program is capable of operating in medical speech as well as standard conversational diction.
According to the company, Transcribe Medical can be used across thousands of healthcare facilities, and it will help aid medical professionals in taking notes and other important information. It offers an API and will be able to be used with most smart devices containing a microphone. When the program reads and processes the information, it returns text in real-time.
Transcribe Medical is currently being used by SoundLines and Amgen.
Vadim Khazan is the president of technology at SoundLines.
“For the 3,500 health care partners relying on our care team optimisation strategies for the past 15 years, we’ve significantly decreased the time and effort required to get ton insightful data,” he said in a statement.
DeepComposer is an AI-enabled piano keyboard that will allow AWS customers to use AI and a MIDI controller to compose music. Amazon is calling the new technology the “world’s first” machine learning-enabled musical keyboard. It has 32 keys, and it is a two-octave keyboard.
Composers who use the program can choose whether to record a short musical tune or use a prerecorded one. They will then select a model for their desired genre and the model’s architecture parameters. They can also set the loss function, a feature used to measure the difference between the algorithm’s output and expected value. The composer can also choose hyperparameters and a validation sample. DeepComposer then creates a composition which can either be played in the AWS console or exported or shared on SoundCloud.
DeepComposer uses a generative adversarial network (GAN) to fill in compositional gaps in songs. Random data is taken by a generator component and used to create samples which are forwarded to a discriminator bit. The discriminator bit then separates the real samples from the fake ones, and the generator improves along with the discriminator. The generator progressively gets better at learning how to create samples as close to the genuine ones as possible.
SageMaker Operators for Kubernetes
AWS also launched Amazon SageMaker Operators for Kubernetes, which allows data scientists to train, tune, and deploy AI models in Amazon’s SageMaker machine learning development platform. AWS customers are able to install SageMaker Operators on Kubernetes clusters, and this can create Amazon SageMaker jobs natively using the Kubernetes API and command-line Kubernetes tools.
Aditya Bindal is the AWS Deep Learning senior product manager.
“Now with Amazon SageMaker Operators for Kubernetes, customers can continue to enjoy the portability and standardization benefits of Kubernetes … along with integrating the many additional benefits that come out-of-the-box with Amazon SageMaker, no custom code required,” she wrote in a press release.
Kubernetes is an open-source general-purpose container orchestration system that is used to deploy and manage containerized applications. This is often done via a managed service like Amazon Elastic Kubernetes Service (EKS). Scientists and developers are able to gain greater control over their training and interface workloads with the program.
A New AI System Could Create More Hope For People With Epilepsy
As Endgadget reports, two AI researchers may have created a system that creates new hope for people suffering from epilepsy – a system “that can predict epileptic seizures with 99.6-percent accuracy,” and do it up to an hour before seizures occur.
This would not be the first new advancement, since previously researchers at Technical University (TU) in Eindhoven, Netherlands developed a smart arm bracelet that can predict epileptic seizures during nighttime. But the accuracy and short time-frame the new AI system can work on as IEEE Spectrum notes, gives more hope to around 50 million people around the world who suffer from epilepsy (based on the data from World Health Organization). Out of this number of patients, 70 percent of them can control their seizures with medication if taken on time.
The new AI system was created by Hisham Daoud and Magdy Bayoumi of the University of Louisiana at Lafayette, and is lauded as “a major leap forward from existing prediction methods.” As Hisham Daoud, one of the two researchers that developed the system explains, “Due to unexpected seizure times, epilepsy has a strong psychological and social effect on patients.”
As is explained, “each person exhibits unique brain patterns, which makes it hard to accurately predict seizures.” So far, the previously existing models predicted seizures “ in a two-stage process, where the brain patterns must be extracted manually and then a classification system is applied,” which, as Daoud explains, added to the time needed to make a seizure prediction.
In their approach explained in a study published on 24 July in IEEE Transactions on Biomedical Circuits and Systems, “the features extraction and classification processes are combined into a single automated system, which enables earlier and more accurate seizure prediction.”
To further boost the accuracy of their system Daoud and Bayoumi “incorporated another classification approach whereby a deep learning algorithm extracts and analyzes the spatial-temporal features of the patient’s brain activity from different electrode locations, boosting the accuracy of their model.” Since “EEG readings can involve multiple ‘channels’ of electrical activity,” to speed up the prediction process, even more, the two researchers “applied an additional algorithm to identify the most appropriate predictive channels of electrical activity.”
The complete system was then tested on 22 patients at the Boston Children’s Hospital. While the sample size was small, the system proved to be very accurate (99.6%), and had “a low tendency for false positives, at 0.004 false alarms per hour.”
As Daoud explained the next step would be the development of a customized computer chip to process the algorithms. “We are currently working on the design of efficient hardware [device] that deploys this algorithm, considering many issues like system size, power consumption, and latency to be suitable for practical application in a comfortable way to the patient.”
New Tool Can Show Researchers What GANs Leave Out Of An Image
Recently, a team of researchers from the MIT-IBM Watson AI Lab created a method of displaying what a Generative Adversarial Network leaves out of an image when asked to generate images. The study was dubbed Seeing What a GAN Cannot Generate, and it was recently presented at the International Conference on Computer Vision.
Generative Adversarial Networks have become more robust, sophisticated, and widely used in the past few years. They’ve become quite good at rendering images full of detail, as long as that image is confined to a relatively small area. However, when GANs are used to generate images of larger scenes and environments, they tend not to perform as well. In scenarios where GANs are asked to render scenes full of many objects and items, like a busy street, GANs often leave many important aspects of the image out.
According to MIT News, the research was developed in part by David Bau, a graduate student at the Department of Electrical Engineering and Computer Science at MIT. Bau explained that researchers usually concentrate on refining what machine learning systems pay attention to and discerning how certain inputs can be mapped to certain outputs. However, Bau also explained that understanding what data is ignored by machine learning models if often just as important and that the research team hopes their tools will inspire researchers to pay attention to the ignored data.
Bau’s interest in GANs was spurred by the fact that they could be used to investigate the black-box nature of neural nets and to gain an intuition of how the networks might be reasoning. Bau previously worked on a tool that could identify specific clusters of artificial neurons, labeling them as being responsible for the representation of real-world objects such as books, clouds, and trees. Bau also had experience with a tool dubbed GANPaint, which enables artists to remove and add specific features from photos by using GANs. According to Bau, the GANPaint application revealed a potential problem with the GANs, a problem that became apparent when Bau analyzed the images. As Bau told MIT News:
“My advisor has always encouraged us to look beyond the numbers and scrutinize the actual images. When we looked, the phenomenon jumped right out: People were getting dropped out selectively.”
While machine learning systems are designed to extract patterns from images, they can also end up ignoring relevant patterns. Bau and other researchers experimented with training GANs on various indoor and outdoor scenes, but in all of the different types of scenes the GANs left out important details in the scenes like cars, road signs, people, bicycles, etc. This was true even when the objects left out were important to the scene in question.
The research team hypothesized that when the GAN is trained on images, the GAN may find it easier to capture the patterns of the image that are easier to represent, such as large stationary objects like landscapes and buildings. It learns these patterns over other, more difficult to interpret patterns, such as cars and people. It has been common knowledge that GANs often omit important, meaningful details when generating images, but the study from the MIT team may be the first time that GANs have been demonstrated omitting entire object classes within an image.
The research team notes that it is possible for GANs to achieve their numerical goals even when leaving out objects that humans care about when looking at images. If images generated by GANS are going to be used to train complex systems like autonomous vehicles, the image data should be closely scrutinized because there’s a real concern that critical objects like signs, people, and other cars could be left out of the images. Bau explained that their research shows why the performance of a model shouldn’t be based only on accuracy:
“We need to understand what the networks are and aren’t doing to make sure they are making the choices we want them to make.”