Mohammad is currently a Senior Data & Applied Scientist at Microsoft, and Instructor at Stanford University. He is a former Data Scientist at Apple and previously worked for Samsung, Bosch, General Electric and UCLA Research Labs. He received a PhD in Computer Science from the University of California, Riverside and B.Sc. from University of Tehran. Mohammad is the author of the book, ‘Applications of Mining Massive Time Series Data’. He has also been a keynote speaker at more than 40 Data Summits/Conferences around the globe.View the profile
About the talk
Good morning. Good afternoon. Good evening, depending on your time zone. Thanks for having me here. I'm going to talk about some of the challenges and you know, the future trends of the defending world. Going to start with this codes soon. We will be ashamed of telling your kids. We used to drive cars and we had to visit doctors to get monitored. This is indeed true. And the one of the main reasons is the topic of deep learning that is making this to become true. This is the outline of the top. I'm going to spend 5 minutes on each topic. We're going to
go through the hottest Trends in deep learning first and then we're going to see some applications and then we're going to talk about some of the tools open source packages. You can use in order to make these applications real and finally some takeaways from industrial experiences. And again, if you have any questions, I'm not sure how the format is for the webinar, but I'm okay if you just jump in and unwind yourself now. Sounds better. All right, so I'm going to I'll try to
finish hopefully in like 20 minutes so that we have enough time for Q&A. This is the Trinity of AI it's the three main components are compute algorithms and data and And you know so AI is not a new term. It's been out for a many years, but the recent advancements in computes Hardware, you know, like components such as gpus and fpgas that have been recently available and accessible and also very large datasets like image necklace, which is a baseline which is like a base reference for all deep learning applications and computer vision and image analysis. And you know,
thanks to all of these new technologies deep learning has been a very attractive and as I've been there like a new trend and a very hot topic these days we can see here the graph a guy is like the whole any intelligence system in the machine learning is a subset in a defining is also like a smaller subset of machine learning even though it's like a small area but it's a very important topic and has I need a lot of impossible tasks to become possible when we talked about data and how much data we have. This is maybe a very nice example to understand the amount of
heater that we're dealing with every autonomous car and average generates data equal to 2600 internet users. This is because of all the cameras and all the sensors inside the car and this technology that is being collected for enhancing the autonomous system. And and also there's like other than a lot of white papers out there in a lot of great people out there that coded that this technology is going to be out there in the next 45 years. So everyone is going to be or at least most of the people are using signals you soon.
So that means we're going to have so much data write one example. I mean, I'm sure everyone has seen all these autonomous Cars 2 Who transfer you know who are used for people commuting to different places, but there's also some new startup companies like neuro was founded in 2017 in San Francisco that is capable of you know, delivering groceries pizzas and other items. So it doesn't really need that. You know, so the safety of the passengers is not really critical here because no one is on the car, but
again is so it's not about only autonomous driving cars, but it's also about cars which are able to deliver food and stuff. This is there's so many applications of deep learning. This is just one example. I wanted to show you especially because we're right now in the middle of the pandemic. This is like a real-time object detection model which is capable of detecting where there's someone is wearing a mask or not. If they're bringing masks. You can see them while it is green. If you're not very good at hiding and bread is a lot of other applications have came
out with something you may have seen them. For example for social distancing. There's something funny models are able to identify if there is like enough distance between you know, the people walking around using cameras. Some of them are not like super accurate because you know, the camera is just able to is not able to like you no sense all the dimensions in like the depth necessary if we don't have all the sensors, but again, like using this video you can simply use an open source tool to label your images and training model using different. Since we're talking
about some of the hottest Trends I want to touch hands and it's been a very Hot Topic in the past few years. It stands for generative adversarial networks. And this was created by Ian Goodfellow and 2014 and it's been one of the biggest breakthrough Technologies of 2018 as primitive Technology reviews of this neural network contains two main components one is called the discriminator and one is called the generator. They show it with G&G. And so we we feed and some real examples and real images to the
discriminator and then some random or synthesized images to the generator. This is a competing now know my birthday tries to make all those synthesize images as much as possible similar to the real samples, right? So this is how is the Ghana cedi? Go to generate some a very very real images that look real but they're not and there's a synthesized and this actually created this top new topical deepfake voat deepfake actually came from again. And is he can see on the right hand side. This is image
posted by Ian Goodfellow. He's working for Apple currently, but he came up with this again 2014 and you can see like one of those examples of images that have been created. So this is like a sentence all these images on the right hand side are synthesized. None of those people exist in the real world, but you can see like the advancement of again from 2014 to 2018 from a small white and black color image all the way to like a very high pressure solution image that looks so real and and also on the left hand side, you can see some of the advancements in
the timeline which again is bald so rapidly and it's still in the Pro. Office of evolving and so this actually created this topic called deepfake. I like 10 years ago. If you wanted to create a deepfake video you had to spend hours or like hundreds of hours to produce a video that looks likes very near to reel in a studio. But today you cannot reproduce a defeat video I using open-source tools and packages in just like minutes. And this is what has made this so hard to distinguish a
debate video from a real one. I mean, there are a lot of content and white papers out there about how to authenticate real videos. There was a recent article posted on CNN. I think like 2 days ago that explained all of like how to visually be able to distinguish one from each other. But even sometimes it's not easy to do it visually or sometimes it's not even possible to distinguish it visually but there are some other different approaches for authenticating. This Microsoft has also taken some steps for a trying to you know hosts some like the
real contents on a server and then I'll send to Kate them so that whenever it gets shared it has like an authentication stamp on it so they can be detected us like the real one not the synthetic one. But anyway, it's a very severe very complex topic and and it's been very Popular these days if you're interested in generating any of these deepfake videos while you can use this open source tool call David verify that works on zoom and some other software tools that you can use to generate a deepfake videos in real
time and you will see like the performance of the algorithm is like so it's it's really fascinating. It's going to be generating these videos in real time with very low latency. This is the next next in a French drain and I we have to every AI algorithm is combined three main components data priors and task now Adidas, obviously one of the main components in any aird playing out with them and one of the main concerns and issues in the challenges is that you know, this part is like unsupervised.
So I mean, I mean ultimately we want to reach an unsupervised method but all of the deep learning algorithms currently we have our supervisor would means you need to label images and you need to you know, Define objects and images for sometimes the first like thousands of images in order to have your model trains. So this is one of the very big a challenges that exist in this world. So it's about data in the future of this part of this area is going to be like unsupportive eyes. For making the algorithms
on supervised we need to be able to develop Concepts such as concept learning and disentanglement money, which is a very very important concept and was life for just-in-time learning so I can explain it further and then on in Pryor has one of the issues that we have any planning models is that they are not robust enough for example and Brenna Thomas car if you just place a sticker like a black sticker on a on a speed limit sign the tournaments are is going to interpret that as a different signs. That's like a very classic example of thing like this means
like stop so it just in the middle of the freeway the car is going to suddenly stop if it sees that such a sign with those black stickers on it on some particular locations on the sign so we can tomatoes are not robust enough in order to make dimmer Buzz. We need recurring feedback and we need uncertainty quantification to be able to you know, so that Autonomous device device or system would be able to make make right decisions when there's some uncertainty situations and finally for the task. We are
actually going to make an Adaptive. We need to make all these all these different models adaptive and for doing that it's going to be like multitask and it's going to be working different domains and it should be like lifelong learning so that when liking you just happens the models would be able to adapt and make the right decisions. this is the ultimate of a I I mean you won't believe but we the the I would say like the ultimate destination of a is to be able to create an intelligent system that operates
such as a baby because the baby has all of those are characteristics that they we discussed in the previous slide. It's it knows exactly how to handle disentangling learning. So distant Tanglewood learning means being able to automatically on in an unsupervised fashion disentanglement and extract the features out of the images or out of the deal. That's exactly what a baby does a baby is easily able to distinguish an adult's from a baby by doing disentanglement learning by extracting the features and
understanding what features, you know, cause someone to be an adult and what features cause someone to be like a baby so they're able to make him a different reactions to those two different concepts and Also able to learn with very limited supervision, even though it takes like sometime in some year bed, but they're able to do that overtime and build that a skill set and they're able to do concept Discovery tasks and friends and learning to compose robustus the noise and all that stuff. So they all exist in a baby. So the ultimate goal of a I used to reach a baby's capabilities.
I'm going to talk about some recent applications of deep learning in addition to you know, the one like the mask protection application that I talked about at the beginning and this has been widely used in a lot of factories and a lot of manufacturing sites these days it's called automatic Optical inspection for short they call a DUI. So it means using a model to automatically detect a cosmetic or functional defects such as Medicaid effects such as scratches smudges missing screws or like you no cracks or
anything on the device to be able to detect this out of matically there so many advantages of it, you know, the the actress you can be very high compared to like a human operator the you know, the running time of solo. You can actually run this in like a point few seconds and instead of having an operator to looking at the device in From all different angles for like a few minutes to identify all these scratches and smudges. So this is actually have some cards if we worked on recently and add Microsoft which you know, the whole
goal here is to use a combination of deep learning and computer vision algorithms to be able to solve this problem. So, you know, this is also one of the distinctions between deep learning in computer vision. I'm just doing this with an example for deep learning. The model is going to be automatically extracting the features in the images. So that means all you have to do is just label the objects in your image on the left hand side. It's like for images of a surface that all of the you know, all of the Cosmetic defect have been labeled in the image then all you have to do is just
pass and a bunch of these like passing examples to of recurrent neural network. It called RNN and output is going to be like detecting these objects automatically, but in computer vision, or I would say like machine learning It involves a human telling machine what the defect or what the object should look like. So this usually includes multiple steps of an image pre-processing and processing Imaging to detect Beano Contours and in different objects that you're looking for it, but it but obviously it's not going to be doing that, you know feature extraction itself. The
developer should be aware of the features that they want to extract. So this is you know, the main difference here between these two topics for the aiy projects. We actually use this open source tool called yellow maybe some of you are familiar with this and have used this open source technology. It stands for you only look once not the only live once then this was developed by Joseph Redmon from University of Washington a few years ago that we use the third version that was developed in 2018. I think he also has like a fourth version in 2020.
That you can use there's a lot of advantages for using this object detection. I'll bring them. First of all, it's open sores. It's like the training time is like super fast comparing to other available Technologies. And then again, there has been a very very hot algorithm that these days in the Deep learning world did this algorithm has also been used and for example and military drones for detecting objects and individuals and I automatically I would say unfortunately it's been using those
cases. But anyways, like open sore so you can use it for any sort of application. This is one example. I just want to show you this model was trained using your load. This is like the first version of the model so you can see like I'm moving the camera. This is just like us very simple cheap webcam that I'm using here moving on your device and locating these, you know different teeth. Not seen any of these defects before like even the color of the effects to your I just created them myself with a pen so you can see it's like blue color. But in the data said there was no
existing blue, you know defect in the end of training sets. This is like another example, you can see like in real time. I'm not trying to locate this device in all different directions in Miles able to like an ID tag each of these objects and real time. And again the training a model using your load is like super easy. It just took just like a few hours to train on a GPU and not many, you know images were required for getting this to train. Maybe just like a few hundred thousands. Threat. So I think I have another
just five minutes to go through my last section. I just want to talk about like Python and hard because you know, these are some of the greatest tools that you can use in their open source, in order to make these deep learning applications to you know to become real. So first of all, because we need to come to get with computer. We actually need to use something open source tools. So I'm just talking about like or in Python. I actually teach a course right now at reduction to data science in or is the end of the main program language using that course. So not course I say or is the
best programming language and then I also used to teach another course introduction to programming in Python and not course. I was saying like python is the best burger in language, but I mean just to be honest, I think both are great and are very useful is more used for Data analysis for visualization examples in applications has like no more General Bowl solution and Furby language. You can use it for deployment and use it for training testing all sort of models, even though it exists in
both or an open-source very hot and IBS days ended up my kid and bisque program tools that you can use. These are some of the very useful packages available in our which you can use like you get fired for doing all the visualization and creating hotspots. You can you stream live free creative web applications give you a yr. From manipulating data carrot the carrot toolbox for training testing different models in the Heights on this deserted like some of the technologies that we actually use for that a y project
but he mentioned we use numpy weos opencv for you can use it for a traditional computer vision for example for Images for a country detection detecting line detecting pixel defects and all that stuff. Then you can use pegboard and Keras tensorflow for training in Building B funny models. I know I'm just going to wrap up in one slide and talk about some of the takeaways from these examples and some of the future trends of deep learning try even though deep learning is so cool and fascinating but try to avoid it and if if machine learning is an option if traditional
computer vision is an option avoid keeps whining because he 20 models are buried at hungry as we discussed earlier the future of this is going to be likeable timidly want to read an unsupervised approach and do plenty which doesn't exist right. Now. We spend using just like one stop sugar one image you could train, you know complex model which is not available right now. So that's why I'm currently my you know, the recommendation is to stick to an ml approach instead of a deep is it has a lot of complexity for training and needs a lot of images for training. So it's a very expensive and
time-consuming process. The future of AI is unsupervised as we discussed this entanglement. Learning is the next Trend because it's one of the you know dates the key components for making you bringing unsupervised. And then or advice on are both great tools open source tools available for free and they are developing so fast these days so they're very nice tools that he can take advantage of and also again just remember there's so many open source tools out there. You'll always just was just one example, so instead of you know, inventing the wheel again,
make sure that you do enough research on your target asking the project that you're working on because there is a very high chance you can find available open source tools for solving that problem, right? That's all they have. Thank you so much for your time and your attention. I will be available for the next 5 minutes. If there's any questions. I didn't mean to confuse you questions in Q&A. There was a good presentation a couple of minutes. And also we have an upcoming
link to join. Is it going to be after this? Can you please repeat? Where's the question where the questions going to be posted right now? Delano questions as of now No, I'm not sure whether you could recommend for learning more on. die Sigmund Elemental learning Okay, let me let me copy and paste rip in the chat. Yeah, I did. Any recommendation for learning more and disentanglement learning there is there's a so there are there are tons of Publications in 2019 and 2020. One of the main
folks that is working on this. I actually forgot her name. She's a professor in a Celtic but she's kind of like I'm leading this this this area. So again, there are a lot of papers out there. I mean, so if you're interested, I mean I can definitely send some things cuz right now I have to look them up. It's going to take some time. But feel free to reach reach out to me on LinkedIn offline and I will forward you some of the new ones some of the recent again papers. One more question. I'll post in WeChat.
Please explain the dice Dyson Daniel learning automatically striking features from the data. Okay. So this is I mean, it's like very different from feature selection method Concepts because just like she already have the features and then you just, you know, find the optimal subset of features, but in disentanglement learning, the algorithm is able to identify the you know, all the features out of the the dataset again automatically that is kind of what you planning models are doing.
For example, when you feed in, you know a bunch of images for like a car and a cat for example, if you want to classify them from each other. Algorithm is able to like find those features inside the data, but it's kind of but it's hard to You know how to translate them or export those features out of the model so she can run on any labels images and hopefully it's going to be unsupervised which means just running it on like one image and and then automatically exporting the features and telling you like what are the
features that caused this image to be a cat versus this image to become like a car? Okay. So it's going to tell you like explicitly if there is a wheel in the car, it's going to be a car say I'm sorry and damage. Can you get car if there's like, I don't know like a some animal that has like four legs. It's going to be a cat's. Okay. I mean, I'm just like making you're just making this like simple, but ultimately that's what we want. So we're not even close to having such a model and I'll grow them but the whole this whole topic is called
disentangle McLennan. And it's because it has became very popular and like a new trend recently. Especially after you planning became a Hot Topic. This is kind of look like the ongoing track, but if anyone is interested, please reach out to me on LinkedIn and I can definitely forward you some more additional resources.
Buy this talk
Buy this video
With ConferenceCast.tv, you get access to our library of the world's best conference talks.