Why Deep Learning Now? | AI Revolution Documentary

hi welcome to another ColdFusion video in recent years you've probably been hearing a lot of hype about deep learning but what is it really all about and here's another curious question why has deep learning only now just come into the spotlight seemingly from nowhere in this video we'll answer these questions and learn the impacts that deep learning has on our world they'll also be a guest appearance from Justin shank who has worked as a contractor finto finally we'll take a look at some cutting edge deep learning applications that are happening right now let's get into it for most people the terms a I machine learning and deep learning are interchangeable before continuing it's important that we make some concrete definitions artificial intelligence or AI is an umbrella term for a branch of computer science its aim is for machines to mimic human cognition with a focus on complex problem-solving machine learning is a subset of AI it focuses on how to make computers learn on their own without the need for hand coded instructions machine learning systems analyze vast amounts of data and learn from previous mistakes the result is an algorithm that completes its tasks effectively deep learning which is what we're going to be focusing on is a subset within machine learning this technology attempts to mimic the activity of neurons in our brains using matrix mathematics this arrangement is called a neural network and dates all the way back to 1957 the modern and more complex networks are known as deep networks but weren't proven practical until around 2012 so to answer the second question why has deep learning only just come about now deep networks have only become feasible because of two reasons an increase in computing power and vast amounts of data since the introduction of the first Intel ship in 1971 the number of transistors have roughly doubled every two years this doubling of transistors on the chip essentially means an exponential increase in computing power this idea was first proposed by Gordon Moore co-founder of Intel and is commonly referred to as Moore's law from the mid-2000s computers became powerful enough to make neural networks perform but the lack of training data meant that these nets couldn't be used for anything actually useful and this brings us to the second point the explosion of the internet eventually led to vast amounts of data that could be used for training purposes so why do our networks need so much data essentially the more data there is the more robustly the network can be trained put yourself in the shoes of a deep learning network trying to recognize a cat for example if you've only seen three cats would only have a few camera angles and lighting conditions to work with something as simple as seeing a cat from a different angle or in a different light would throw you off but on the other hand if you've seen thousands of different cats you'd have a much easier time recognizing one this is the importance of data computers aren't as intuitive as humans and need more examples in the field of deep learning data is the essence that allows machines to learn you could look at it this way if data is a new form of oil the Internet is a huge oil reserve and deep learning systems of the machines that run on this data so the true field of deep learning all began in 2012 in this year deep learning exploded into the spotlight when an artificial neural network was used in the competition to recognize the world's biggest data set of images it was the first time a neural network was used in this competition and it blew all previous types of algorithms out of the water at this moment the world realized that deep nets were useful after all this was the birth of the modern AI movement thanks to the exponential growth of computing power and the vast amounts of data deep learning networks have gained the ability to recognize objects and translate speech and real-time they're everywhere from our phones to smart home systems that's all the stuff that you've heard before the thing that many people may not know is that deep learnings impact is far wider than most people think vast amounts of data can be found in all corners of Industry deep learning affects anything from oil exploration to energy grids social media information medical records code compiling server farm power management border policing and so on the organization and interpretation of this data is very useful to businesses at all levels the problem though is that this data is unlabeled and unstructured this means that it can't be used to train machine learning programs that depend on supervised instructions deep learning networks can avoid this drawback because they excel at label Asst unsupervised learning especially when it comes to prediction and pattern recognition to explain further let's hear from Justin shank hi my name is Justin shank I'm an AI researcher at Pelt Ariana and an Intel software innovator so deep learning it's quite inspired by an understanding of the brain the neural network idea we can compare it also in terms of its ability to learn certain representations and the brain for example is at a young age a child can learn a word after hearing at only one or two times and that's considered amazing for machine learning standards to be able to generalize from very few instances of data so deep learning allows us to find patterns and also relatively limited data making generalizations that's a that's one of the amazing things about deep learning ok so let's take a look at some of the emerging hardware for deep learning the Intel Murphy Dias neural computing is Intel's effort to provide D playing on the go simply put it's a standard USB Drive that plugs into a computer that enables rapid prototyping of deep neural network interface applications locally that is it doesn't require users to connect to the cloud so if you have a drone for example that is flying over the water and trying to see if anyone is drowning to drop them the life preserver you want that to be able to compute what it sees in real-time without having to send that data back to some server Intel has also purchased navona systems for 400 million dollars the varna chips are said to be able to transfer data in and out at about 2.4 terabytes per second at a very low latency that's supposed to be about 10 times faster than traditional chips if we leave the rest of Apple aside for a moment the neural engine in the iPhone tends a 12 Bionic chip is impressive it signifies a shift towards deep learning being local and integral to mobile capabilities uber an uber eats deep learning system called Michelangelo enables the cars to arrive on time by analyzing data from millions of previous trips and applying it to your specific situation it also knows where all the drivers are and what trips they're completing at any particular time it uses this information to calculate which optimum driver should pick you up how else do you think the cars arrive so quickly the technology is also making personal health care easier data from wearables provide patient specific information directly to their healthcare professionals deep learning can even help predict a range of diseases by analyzing eyeball retinas or it can diagnose eye or other diseases as accurately as doctors marking a milestone inside this hospital in Shenzhen in southern China doctors here getting a hand from artificial intelligence for example we have vivid images the AI system will sort them and select the pictures showing a high possibility of cancer we only need to check five selected images for a diagnosis a eye is able to learn from numerous amounts of data that's what humans can't do now the accuracy of early detection of esophageal cancer has reached 90% roughly the same level of diagnosis made by human doctors since the launch it served 400,000 patients right now it detects breast cancer and prostate cancer but we could also work on any other cancer type that we train models for a team of pathologists that went through thousands of images teaching the computer what cancer cells look like it's going to make the doctors much quicker they're going to make them better which is going to overall benefit the patient the mission is to increase the accuracy and availability of cancer diagnosis and save people's lives I can hardly think of anything that's more motivating MRI CT and EEG scan data is used to learn patterns and diagnose diseases such as cancer and heart disease it seems like there's almost no limit to what deep networks can achieve in the healthcare industry okay so let's look at some more fascinating new possibilities from just the past six months or so that have just been made possible through deep neural networks so style transfer is something already known to most of you taking the style of one artistic image and imposing on something completely different but what about motion transfer this is now recently possible a recent neural network can take a professional dancer and transpose their motion onto people who well can't dance take a look a similar thing has been done with animals for potential use in video games this neural network is controlling this arm learns to pick up objects in a very real way it can't just use its hand to pick up the objects because the object is too wide and grasp can't fit in order to complete the task it actually had to demonstrate learning and figure out that it had to break up the object before it could pick it up another reason neural net uses the data it receives from Wi-Fi signals bouncing off humans to see their motion in the dark how about generating an animated face just from an outline deep neural Nets also help in denoising usually cleaning up an image like this is a very labor-intensive or impossible it can now be done in an instant when neural networks this has obvious applications for cleaning up low-light images and smartphone cameras but can also remove other noise including randomly generated text it's important to note that while doing all of this the network have no access to the original photos it doesn't stop there though a pair of neural networks working together in the same system managed to give super resolution to pixelated images it's not hard to see how this network and the previous noise reduction Network could be used to bring perfect photos to anyone just look at the detail extracted in these shots imagine having a choppy video at a few frames per second and then creating super slow motion just from that this is now possible thanks to some neural networks playing video the network was trained on thousands of videos so it learned what made a good slow-motion video and now it can apply it to learned knowledge to any video the problem of deep fakes has now been tackled by neural nets I mentioned this as a solution in my deep fakes video and a lot of people argued that it couldn't be done but a few months later here we are enter round off we have some new artistic neural net loosen ations pretty interesting to look at so what can we say in conclusion deep Nantz as a technology has the ability to complete highly-skilled work that was traditionally expensive it could lead to new scientific breakthroughs and a drastic fall in the price of goods and services but there remains a question if these systems are so good and become better than us at our jobs what are we going to do this is probably one of the hardest questions to answer the solution could involve a natural creation of whole new industries that don't exist yet or an enabler for us to do even more intellectual work than ever before that's a whole discussion for another day regardless the discovery and recognition of patterns and regularities in the world around us lies at the heart of scientific and technological progress it's how we advance and innovate it's also an area where deep learning excels the question isn't whether or not deep learning is useful but what is the limit to creativity in its applications so that just about rounds out at the end of the video I want to thank Intel for helping out with this video and if you want to check out the Intel compute stick you can click the link below I'd be interested to know if you guys have come across any interesting applications of deep learning lately I think this is one of the most interesting fields to watch and there's been advancing faster than anything I've ever seen alright thanks for watching this has been – go go you've been watching cold fusion if you just stumbled across this channel and you're here for the first time feel free to subscribe cheers guys haven't you

48 thoughts on “Why Deep Learning Now? | AI Revolution Documentary”

  1. Anybody noticed that the most smart people on this technology are Asians? Tells a lot how this guys on top of the line in technology.

  2. I hope new techs centered on AI and policies to utilize them will achieve the sustainability of human civilization.

    They will do that by being friendly to the natural ( including body-internal ) & social environment, taking the balance between production & distribution in economic & social activities, uplifting ourselves in a humane way 💖

    https://www.pinterest.jp/pin/596445544381001862/ / https://ameblo.jp/lovelucifer/entry-12266020322.html

  3. Can I download Chess algorithm AlphaZero on my iphone ? I manage to win once against the algorithm Stockfisch and many draws !

  4. Deep Learning in R

    >Perform deep learning with R and neural networks with R.

    > Gain mastery on the theory and concepts of Artificial Neural Networks (ANN) and understand deep learning neural networks (DNN).

    > Execute Deep Learning with R and understand various deep learning algorithms.

    > Master the execution of both Artificial Neural Networks (ANN) & Deep Learning Neural Networks (DNN) using the H2o deep learning R package .

    > Execute and master Deep Learning Neural Networks (DNN) Machine Learning classification & regression problems in R with a live deep learning example.


  5. I'm an optimist about AI. I think it will just destroy our species in the long run, instead of doing something far worse.

  6. I think the idea is for us to create A.I. that can do work for us so we don't have to. We could lead the lives we want and answer real questions. But as usual someone will try to profit off this and perpetual inequality.

  7. Im an English learner, and l love your videos very much. I'd appreciated that if you could put your subtitles on the video, i would understand the meaning better and exactly. Keep going Dagogo! Love you!

  8. I just want to see a game with proper ai for enemies/NPCs/companions.
    Imagine having an antagonist that learns how you play and how to counteract it, so you then have to change your own strategies.
    Total war perhaps? Or a 4x game. Deep learning in StarCraft. Fun!

  9. If you use your phone to assist another person, then it will need to be told not to bias the answers that it fits to your character.

  10. You have a great taste in music. I noticed some Mogwai playing at about 7:30 :). Great videos too. Keep up the good work.

  11. Deep learning with AI is going to change the world in a way that only humans will exist for the distruction of themselves…

  12. I'd trade about 7 billion people and about 100 years of technology to be able to have a decent place to hunt fish and farm. Fuck why did I have to be in this generation.

  13. I wonder why you didnt mention the most important thing that makes deep neural networks work today? Maybe to technical i suppose.
    Mores law is and was just a small time window, he even corrected it himself by making the time steps for transistor size longer since it was really easy to begin with and now we are facing quantum effects which limits the transistor size to shrink much more. My phones octa core CPU/GPU at 10 nm is pushing the limits and its not possible to go much smaller and its also becoming much more expensive. Just look at the CPU market today where you have to pay 2000$ for a high end intel CPU. That's why GPUs with thousands of cores or ASICS designed for tensor calculations are beginning to be a big breakthrough when it comes to computation.
    The real answer to why the giant leeps has come over the last decade can be found in detail watching computerphiles channel. The advancements has been possible because of new algorithms to train multi layer deep neural networks much more efficiently. The computer science to have many layers of neurons each "assigned" or trained correctly didnt exist before.
    A good example is Alpha zero which only trained against itself and beat Alpha go 100 out of 100 games when Alpha go was already superior to any human player.

  14. I thought amd and Nvidia where more involved in ai and deap learning due to GPUs being better at neural networking than traditional CPUs like Intel's. from what I understand that why intel is investing as much in there up coming entry to GPU space with the hiring of Raja Koduri and several other key players in GPU space

  15. AI will initially do good and will be well received. Later on, it may prove fatal when it detects humans themselves as errors as humans do so many mistake. I hope the technology will be well controlled and risks associated with will be strictly dealt on.

  16. Here's hoping it doesn't lead to skynet. But deep learning does look really promising. I'd love to be able to use its labor saving attributes for creative pursuits such as making it feasible for me to complete a graphic novel project I've been working on for years

  17. It all looks amaging and hopeful for mediocre minds, while undermining the complexity of real world and real situations. Is it all force fitting some 'science' and 'technology' for profits???

Leave a Reply

Your email address will not be published. Required fields are marked *