NVIDIA Developer News

Title is misleading. Nvidia have created a network which can slow down (i.e. interpolate) videos by a factor of four.

Actual link: https://news.developer.nvidia.com/transforming-standard-video-into-slow-motion-with-ai/

Most importantly, the footage shot in the video was at 2500 fps native [1] - at 300k FPS you have to use a significantly reduced image size. FPS is bottlenecked by the time it takes to read out the sensor, which is linear in the number of rows in the image. So if you read out half the sensor, you can double your frame rate. You can exploit this to get thousands of frames per second on a cheap machine vision camera.

The camera framerate is largely irrelevant. While you can slow down footage by four times (neat), that doesn't mean you can get 300k fps footage out of a compact camera. The great benefit of this technique is really that you can convincingly fake super-slow motion without having to sacrifice sensor resolution. I can easily see this as a cool iOS or Android feature.

Maybe if they'd actually used footage that was at the limit of camera technology this would be more impressive (I don't see a technical reason why not?). Why not combine it with super resolution too? Interesting security implications, as cameras with frame rates higher than (I believe) 1M fps are ITAR controlled.

[1] https://www.youtube.com/watch?v=5mZovjRlkWs

Surely, that is a solved problem now that Google has "advance[d] the state of the art in natural language technologies and buil[t] systems that learn to understand and use language in context"? [1]

Or, I guess, one can just buy more NVIDIA GPUs, and simply solve the problem using Torch, following their online guide on "understanding Natural Language with Deep Neural Networks Using Torch" [2].

In general, I hear that many language understanding tasks are finally conclusively conquered, as shown by OpenAI with a combination of supervised and unsupervised learning that has led to an ImageNet-class increase in the state-of-the art results in language understanding datasets [3].

Why, these days, we can even build an "AI System [that] Understands Music Like Humans Do"! (NVIDIA, again; [4])

The inhabitants of Scunthorpe and all the Dicks, Pennys and Dikshits of the world can finally rejoice! Machine learning is on the job!

______________

[1] https://ai.google/research/teams/language/

[2] https://devblogs.nvidia.com/understanding-natural-language-deep-neural-networks-using-torch/

[3] https://blog.openai.com/language-unsupervised/

[4] https://news.developer.nvidia.com/ai-system-understands-music-like-humans-do/

In my line of work we say:
"It ain't cancer until the pathologist says so."

Basically it is a joke amongst the pathologists that no matter what MD does the diagnosis, s/he will always send a biopt to the pathologist and ask if it is cancer. Well, in simple terms of course.

I guess the correct moment to say you have been diagnosed with cancer is when the pathologist tells the physician it's cancer.

Beautiful thing is neural networks / deep learning can really outperform the pathologist now that we have introduced digital pathology. Sure they will for some time give the final verdict but before too long AI will do the actual diagnose with better accuracy than a human possibly could in a fraction of the time. Reports started coming in the last year. E.g. [0][1][2]. Last one is directly mentioning years.

0) https://www.extremetech.com/extreme/233746-ai-beats-doctors-at-visual-diagnosis-observes-many-times-more-lung-cancer-signals

1) https://news.developer.nvidia.com/diagnose-heart-disease-with-ai/

2) http://bigthink.com/paul-ratner/heres-when-machines-will-take-your-job-predict-ai-gurus

> ML/AI has no conception of

Time: False. Time can be included as an input. Previous system state can also be communicated forward in various ways (ex the LSTM architecture).

Conceptual context: I'll agree, since that seems like it would require general abstract reasoning ability (ie strong AI). Similarly, robust treatment in the general case of physical continuity and cause and effect both seem like special cases of conceptual context to me.

That being said, bear in mind that we already have examples of not so robust treatment for special cases of physical continuity as well as cause and effect. (Example: https://news.developer.nvidia.com/transforming-standard-video-into-slow-motion-with-ai/) Who knows how close (or far) we might be from a general solution?

Here are a few more who have tried it or use deep learning for SAR:
https://www.planckaero.com/maritimedrone/
https://flytbase.com/flytstories/search-and-rescue-drone/
https://news.developer.nvidia.com/autonomous-search-and-rescue-drones-outperform-humans-at-navigating-forest-trails/

Data for training such models for practical use are difficult to get. Often, one has to train a model on a different but related dataset that contains your object of interest and ensure that your model generalizes well. These models are highly sensitive to the data that they are trained on, so it is best if you can gather data which is closer to the actual environment (in which you want to use the model).

It's a widely-used and widely-studied method, so the popular implementations are pretty optimized and the toolkits extensive. Unless you have some fundamental insight/optimization specific to your problem, I don't think it's actually worth it to roll your own. Easier to just start with something open sourced and build a small module on top.

Nvidia is rolling FEM into Physx[1] - exciting stuff! GPU-accelerated FEM is going to be a game changer if they didn't cut too many corners in the pursuit of speed.

I wonder if it can be coupled with FleX...

[1]: https://news.developer.nvidia.com/announcing-nvidia-physx-sdk-5-0/