For example, with a GAN approach, one might use the re-identification model as a discriminator and try to train a generator to generate custom overlapping mask patches over one’s face in order to fool the discriminator into misidentifying the person.
AlphaStar uses Impala over tree search. Comments here explain this is mainly due to action space width. But conceptually, i never grasped by one method makes better use of a given “exploration budget”.
A. Is it just tree width or also the episode length?
B. Someone (maybe Vinyals?) mentioned that “it would be hard to saturate the GPU” with tree search. So if sc2 was a light weight reversible environment, would (a narrow?) tree search become feasible?
(Lets ignore issues such as hidden information, agent league, real time. the building order assistance)
I have feature vectors with 2048 elements, also my feature may change overtime e.g. new features are added. I use Faiss as a search engine but I am not quite sure how to save these vectors. Right now I am syncing local folder with AWS s3. I do not think it is optimal because I have to sync files each time I search for similarity which takes a while.
Maybe I should use a vector database (like https://github.com/a-mma/AquilaDB or some other) or is there a more optimized method to sync my local and s3 storage?
Hey all, a team I’m part of just open-sourced our internal HPO tool called Auptimizer. Auptimizer does a couple of things. It provides a single interface to 6 different HPO algorithms including Spearmint and HyperOpt. It also makes it easy to scale your model training from CPUs and GPUs all the way to multiple instances on AWS. The repo is on Github. We have an article about it on Medium and you can find more implementation details in our 2019 IEEE Big Data paper. If you do HPO, check it out and let us know on Github how things look.
Employers are scrambling to find people with AI, machine learning and data science skills and higher education is responding. Leaders from a group of top universities gathered at GTC DC Wednesday to discuss how universities can meet this demand.
Martial Hebert, dean of the School of Computer Science at Carnegie Mellon University, was joined by Cammy Abernathy, dean and professor of materials science and engineering at the University of Florida; Kenneth Ball, dean of the Volgenau School of Engineering at George Mason University; and Joe Paris, director for research computing at Northwestern University.
GTC DC has become the premier AI conference in the nation’s capital, this year attended by more than 3,600 developers, researchers, educators and CIOs focusing on the intersection of AI, policy and industry.
Wednesday’s panel, moderated by NVIDIA’s Jonathan Bentz, a solutions architect for higher education and research, life science, and high performance computing, discussed the importance of democratizing AI and data science tools and concepts for students.
The panelists explored three ways to better democratize AI: new degree programs, new coursework and building skills.
“We are distributing important digital skills throughout every course and major — from humanities to fine arts to healthcare to genomics — and developing brand new degrees to meet the needs of the changing workforce,” Ball said.
A major challenge to building skills, however, remains access to computing resources. Hebert described computing as “one of the biggest obstacles” faced by institutions limiting the number of students who can be involved with cutting-edge work.
In addition to access to the most capable machines, students need to be equipped with the knowledge and tools to address bias in AI.
“As we head down this path, it’s not lost on us the examples where our biases as programmers are finding their way into codes that are being applied to important tasks,” Paris said.
Abernathy said she’s “amazed” to see how quickly AI and machine learning have embedded themselves in almost every discipline. As the technology spreads, she stressed the importance of reaching out to and preparing underrepresented groups.
“It’s pretty clear if you want to be employable and a leader in your profession, you need to have skills in these domains,” Abernathy said. “It’s important that we provide access to a wider range of people.”
At GTC DC, the NVIDIA Deep Learning Institute offered a bevy of sold-out courses, workshops and hands-on training in AI, accelerated computing and data science and it announced a dozen new courses on Monday.
Resources:
Get started with your own AI training today with the DLI.
I’ve found that before loading trained models, the network runs at relatively low speed. After loading the .pth file, the speed of inference boosts about 10 times faster. Does this circumstance normally come in deep learning?
I’ve tried on SSD(single shot multibox detector) on object detection task of COCO dataset, the code is written in python with pytorch 1.0.
Before loading, I got 2-3 fps on GTX1080, and after that, it reached 20 fps on the same device under same environment.
MIDAS detects microcluster anomalies from an edge stream in constant time and memory, while providing theoretical guarantees about its false positive probability. Microcluster anomalies are suddenly arriving groups of suspiciously similar edges, such as lockstep behavior and denial of service attacks in network traffic data.
If somehow you managed to get your hands on a computer with literally unlimited compute power (through an interstellar alien technology exchange or whatever), and could feed it any/all data currently available online, what would be your approach to creating a “true” AGI? Is there any approach currently out there that might realistically result in an AGI if given enough processing power/data?
I’m currently trying my hand at writing speculative fiction, and was wondering what a realistic approach in such a scenario might look like…
With lives at stake, and the clock ticking, mastering disaster may be the ultimate AI challenge.
Teams from Johns Hopkins University, Lockheed Martin, the U.S. Department of Defense’s Joint Artificial Intelligence Center and NVIDIA Wednesday outlined how they’re working to put AI to work speeding disaster relief to where it’s needed most.
The teams spoke about their work at GTC DC, the Washington edition of NVIDIA’s GPU Technology Conference, which brought together more than 3,500 registered attendees — policymakers, business leaders and researchers among them — to discuss and learn about the latest in AI and data science.
Their presentations underscored GTC DC’s role as Washington’s premier AI conference. They represent the latest efforts, detailed at the event over the past several years, to put the benefits of AI into the hands of policymakers and first-responders.
Detecting Damage with Satellite Imagery
A team from the Johns Hopkins Applied Physics Laboratory and the Joint AI Center (JAIC) spoke about how they’re using GPU-powered deep learning algorithms to track the damage caused by major storms from airborne and satellite imagery data processing.
Speakers included software engineer Beatrice Garcia and senior engineer Gordon Christie, both from the university’s Applied Physics Laboratory, and Captain Dominic Garcia, project lead at JAIC.
While their work hasn’t been deployed — yet — in disaster zones, their goal is to create AI systems that harness satellite and aerial imagery, along with other data, to point first responders and military and government decision-makers and analysts to where the need is greatest.
Such images will help first responders see, at a glance, where to deploy their resources, Christie said, as he showed an AI-enhanced map assessing the damage caused by a tornado that struck Joplin, Mississippi, in 2011.
The lab and JAIC have applied deep learning algorithms to the imagery of a number of severe storms collected from airborne platforms to accelerate detection of flooding and damaged infrastructure.
Based on the algorithms they developed and techniques they learned, the joint team is now creating a scalable environment that would provide these capabilities to any analysts. Users would have access to AI and machine learning algorithms, enabling a faster response to a variety of natural disasters.
Lockheed Prepares with Earthquake Simulation
Andrew Walsh, a senior staff systems engineer at Lockheed Martin, explained how the company is building an open dataset that can be used to train AI for better responses to earthquakes.
Lockheed Martin next explained the work that they’ve done in conjunction with a team from NVIDIA to build an open dataset for multi-platform, multi-sensor machine learning research and development.
The dataset, focused on humanitarian assistance and disaster relief, is being developed using a combination of real-world data collection events as well as simulation. The current emphasis is on earthquake scenarios.
Andrew Walsh, a senior staff systems engineer at Lockheed Martin, joined May Casterline, a senior solutions architect at NVIDIA, to explain how they choreographed a real-world collection event that included multiple sensors, aircraft, ground vehicles and teams of actors in a series of simulated earthquake scenarios. They also detailed the effort required to spatiotemporally align all the disparate data sources and described the challenges around labeling such a massive dataset.
Their dataset will be used to train AI and machine learning systems to improve responses to real earthquakes.
Disaster Planning with Data Science
Sean Griffin, president of Disaster Intelligence, spoke late Wednesday afternoon about his company’s approach to disaster prevention and response. His D.C.-based firm is working to create a common web platform that collects datasets relevant to natural and manmade disasters, which are then displayed graphically.
Users — from first responders to everyday citizens — can access the data to make more educated choices before and after a disaster.
“We used to share situational awareness by PDF or sharepoint sites,” said Griffin. But high performance computing is making it possible to update larger audiences with more relevant data.
“It’s our objective as a company to have complete saturation across the U.S. to have outage data in our platforms so that not only do we know that the power’s out, but that we can intersect that information with other key points of interest like healthcare facilities or water systems.”
Griffin presented two use cases. The first showed how Disaster Intelligence’s platform can model the consequences, cost and options for different disaster relief strategies. The second addressed how the platform improves coastal evacuations during hurricanes.
Route Planning with RAPIDS
NVIDIA is hosting a webinar on how RAPIDS, the company’s GPU-accelerated data science software stack, can help speed up route replanning for civilian and military disaster response assets. Register for the webinar, taking place Dec. 17 at 10 am PT, here.
![[P] Playing Snake with Reinforcement Learning](https://b.thumbs.redditmedia.com/quEB9iUa018-25TLmmMEgdr7uJWTo3fy065yJ5CEabM.jpg "[P] Playing Snake with Reinforcement Learning")
