 Hello thank you very much everyone for joining in today. I'm very glad to be here. I'm Tanmay Bakshi and I'm from Toronto Canada and I'd like to say a big thank you to the team at the Linux Foundation and the open source summit for actually inviting me to have this keynote at this great conference today. But I'd like to start off by telling you all a little bit about myself and what I like to do. As mentioned I am a software and cognitive developer as well as an algorithmist at Darwin Ecosystem. I'm also an author, TEDx and keynote speaker, YouTuber, chorus instructor for Udemy. I'm an IBM champion for cloud, honorary cloud advisor and the host of an IBM Facebook live series called Watson Made Symbol with Tanmay. But going into a bit more detail, as I mentioned I'm a YouTuber and I have a YouTube channel called Tanmay Teaches with over 80,000 subscribers in over 145 videos. On this channel I absolutely love to teach about topics like computing, programming, algorithms, machine learning, Watson and artificial intelligence, math, science, and of course neural networks which I'll be getting a lot deeper into in just a little bit. But as I also mentioned I'm an author and I've written a book called Hello Swift iOS app programming for kids and other beginners. And I'm also currently working on writing another book about IBM Watson which will give you a simple beginning to the deep learning as a service API that IBM Watson provides. But as you can tell so far I am really passionate about services like IBM Watson and especially how you can integrate custom machine learning capabilities into your applications. And to demonstrate for you how much I really love machine learning and what I've done with it, I've handpicked three of the open source projects that I've worked on that I think you're going to be very interested in. And out of all these projects it starts off with Ask Tanmay. Ask Tanmay was actually my very first artificial intelligence project that I created in October of 2015. It's an open source NLQA or natural language question answering system that can answer your person, organization or location natural language questions. It's built off of IBM Watson's natural language understanding and natural language classifier services and uses custom machine learning built in by a bi-directional attention flow by Allen AI in TensorFlow. In fact not only is this just any NLQA system, it's the world's first web-based NLQA system to be powered by IBM Watson. In fact this over here is a chart of how the Ask Tanmay system works. This is the Ask Tanmay algorithm, the new version that I've created. And while I won't go into too much detail of how it works just yet, I don't have enough time for that unfortunately. I will tell you a little bit of how it works. It starts off of course with the question and the data that Ask Tanmay has to actually try and find the answer from. And in this case I'm using a search engine, Google. But then Ask Tanmay actually has to try and find out what your question is actually looking for. And that's done via the ATD or answer type detection stage. Ask Tanmay then has to use deep learning through DLARFS and DLNAM steps to actually try and find answers that could be the correct answer to your question. And from there it has to try and score and merge those answers via the SCM or similar candidate merging step which will merge different answers that mean the same thing semantically, but they may not mean the same thing syntactically. But instead of me standing here telling you how Ask Tanmay works, I'm sure that you're going to be very interested to see a demo of Ask Tanmay itself working. Thank you. Now let's switch over to my Mac here where I'll show you a demo of the Ask Tanmay system in action. As you can see over here I've got a terminal window open and if I'd like to ask Ask Tanmay a question all I need to do is run a simple shell script. And I'm going to be timing this shell script as it runs. And so I'll just run Ask.sh and I'll give it my question. In this case let's give it a question that we all know the answer to so we can make sure it's right. In this case where is the open source summit North America being held? All right. So now of course it asks me for my password. And from here what's going to happen is Ask Tanmay will start off by initializing BIDAF in TensorFlow. What this is doing right now is it's opening up the saved and pre-trained model for BIDAF. Then it instantly searches two pages of Google results, runs ATD, realizes we're looking for a location, runs the entire natural language understanding step and as you can see it returns Los Angeles as the correct answer with 54% confidence. And in fact you can see here out of all of the candidate answers that Ask Tanmay had at the end all three are technically correct. Los Angeles, LA and California are all correct in this case. But another question that I'm sure of course a lot of us here I mean practically everyone knows the answer to. Let's ask it something like who invented Linux and the Git. All right. So now of course what's going to happen Ask Tanmay will again go over to Google versus going to start off by initializing TensorFlow. It's going to load the pre-trained model. It's going to run ATD and SERSE or search engine results summary extraction. It'll run the natural language understanding step. It'll do everything in parallel so this is very scalable with more cores and as you can see it returns Linus Tervalds as the correct answer with 75% confidence. And that was a demo of the Ask Tanmay system in action. I'll make this a little bit bigger so you can see that in the back. All right. So that was a demo of Ask Tanmay in action. I do hope you enjoyed that demo. Now we can get back over to my slides here. And now thank you. Thank you. Thank you. Also that was how Ask Tanmay works. Now though I'd like to take you to a separate field in AI. Utility and security. With an application of mine that I'm very excited about and it's called Deep Spade. Deep Spade stands for deep spam detection. And it aims to build off of the work of a group of programmers called Charcoal SE. Charcoal Stack Exchange are a group of programmers dedicated to making Stack Exchange a better place. And they actually have a tool that allows you to actually detect, report, flag spam on the Stack Exchange network instantly. It's called Smoke Detector. And instead of me telling you more about Charcoal and Smoke Detector, I thought it would be better if we heard from Charcoal themselves. And so I've invited Felix from the Charcoal team to tell you a little bit more about Charcoal and what they're doing. But unfortunately, he wasn't able to join us today. Physically, he's in Germany right now. However, we can definitely get on a Skype call with him and actually hear a little bit more about Charcoal and what they're doing. So we can switch over to my Mac here. And I'll start off by of course going over to Skype here. So I'm going to call Felix. Hi, Felix. Hello. Hi. How are you? I'm good. Thank you. Great. So would you like to tell the audience a little bit more about yourself and where you are right now? My name is Felix and I'm calling from Germany. My username on StackExchange is Magish. All right. Great. So now tell us a little bit more about what Charcoal is and what you're doing there. It's an open source project by members of StackExchange. It started off to make moderator tooling, but right now our main focus is to detect and deal with spam. Okay. So tell me a little bit more about Smoke Detector. I know this is the project that you've been working on. What is Smoke Detector? It's a Python script that scans all new posts on StackExchange and runs them through a set of filters to see whether there's spam or not. Okay. So tell me a little bit more about how the actual filters behind Smoke Detector work. So Smoke Detector has these filters to try and find spam. What are these filters? They are mostly reg X patterns that members have identified over the years. Okay. So I see it uses regular expressions. And so tell me a little bit more about the auto-flagging feature. I know Smoke Detector can actually take accounts and actually automatically flag. Tell me a little bit more about that. Based on a database of member feedback on reported posts, we can judge how accurate our filters are. And based on combinations of that, we're using volunteer accounts to cast spam flags automatically on posts we're very sure are spam. I see. Now one more interesting thing, Felix, is that so far deep spain, it's open source, and it's in its source code was completely closed source. But just like Ask Tanmay, I made that open source in June, another project of mine during my keynote at Developer Connect. I'd like to open source deep spate to the entire community and of course the charcoal on stage here right now. So I really do hope I've been able to contribute to the Smoke Detector project, Felix. And I really do hope that of course in the future I'm able to continue to improve the system and help you improve the accuracy of Smoke Detector. Yeah. I look forward to see how you can help us improve the Smoke Detector with machine learning. Thank you very much for joining in today. I know it's around 2 a.m. for you there. But again, thank you very much for joining. Have a great day. You too. Thank you. Goodbye. Bye. That was Felix. Thank you. Thank you. And so now deep spade, which runs deep spam detection tasks via deep learning is now completely open source on my GitHub page. And of course I do welcome everyone in the community to contribute to it, learn from it, and of course apply it to other websites that you believe may benefit from it. I know YouTube, email, Quora, they can definitely use systems like these. Now we can switch back over to my slides and that was Team of Charcoal. But now that you know a little bit more about Smoke Detector itself, let's talk more about deep spade. Deep spade itself is powered by a natural language classification model built with convolutional neural networks and gated recurrent units. And I'm going to be talking a lot about why exactly this model actually works and why it's so great. In fact, I actually already talked about it yesterday during my BOF talk and of course my blog on the linux.com website. But the back end technology of this entire system is actually powered by Quora's with a TensorFlow back end and Python 2.7 code. And in fact, Charcoal actually gave me a day of dump of all the spam and non-spam posts collected from Smoke Detector. And using that, I actually got 98% accuracy on around 15,000 testing rows. So I'm very satisfied with how this works. But again, instead of me standing here telling you how this works and telling you the accuracy count, I think it would be better if you could actually take a look at deep spade itself and how it works. All right. So now what I'm going to do here is I'm going to switch over to a PHP interface that I've got running for deep spade and how it works. Now this PHP interface allows me to enter in posts and it should allow me to actually classify a spam or non-spam. And I've also got this text file over here. This text file contains two stack overflow posts. And these are real posts copied and pasted from Stack Exchange. So what I'll do is I'll go ahead and take the very first post, which seems like it may be a spam post because they give out loans to any country all over the world. And so theoretically, this LSTM, remember LSTM, GRU and CNN combination on some of these two models will run on an i7 CPU without any GPUs and TensorFlow returns spam as the correct answer within around four seconds. And remember, it's still running all of the initialization code. It's still running all the pre-processing steps. And then it's returning the answer. The actual performance of the system is much, much greater. Now we can actually go back and take the non-spam post. And now I should be able to post it. I should be able to paste it into this text box. And if I click on submit, this interface should tell me within four seconds that this is in fact not a spam post. And that was a demo of the deep spade system in action. Really do hope you enjoyed that. And of course, this is just one example of how it can be applied to Stack Overflow and Stack Exchange, but it can be applied to a wide variety of many, many more websites and practically any online community forum. We can get back to the slides now. And next, I'd like to take you to an entirely separate field where I believe that artificial intelligence itself is making the most impact. And this field is the healthcare field. And I'd now like to talk about a project that's very close to my heart called the Cognitive Story. The basic point of the Cognitive Story is to augment people's lives and to do so through the power of cognitive computing and artificial intelligence. In fact, this project is actually a collaboration effort between Darwin ecosystem, IBM, not rocket science, and me. The first chapter in this Cognitive Story is all about assisting those with special needs to allow them to live better, easier, and really fuller lives. And I'd now like to play a quick video clip describing Boo's story, the original person that we're trying to help with the Cognitive Story and, in fact, the origin of the Cognitive Story itself. The pair was joined by 13-year-old Tanmay Bakshi, who was recruited to join the team because of his skills with neural network systems. The road to Boo's hometown was sometimes desolate, sometimes blustery, but by the time the team reached the small town, the sun was bright and the sky was blue. It's not home to many, a hamlet like many others, but in this borough, the developers were coming to see just one resident. They arrived at Boo's ready to learn and understand a woman who can only communicate the broadest of concepts and only to those most intimate with her condition. Her name, Boo, comes from a sound she could make as a baby but can no longer make. Boo is quadriplegic, suffering from Rett syndrome, a rare and debilitating neurological disorder, but her parents know when she's happy, when she's irritated, or when she wants to watch TV, the simplest of desires interpreted by the most intimate knowledge. Now, the hope is that artificial intelligence can match EEG signals from Boo's brain to what her parents say she's thinking, and a unique language born from cognitive computing can act as Boo's translator to others and relieve her isolation. So now that you know a little bit more about the story behind the cognitive story, I'd now like to take you a little bit deeper into the technical part itself. Of course, it starts off with the data that we're able to gather from Boo, and in this case, we've decided to use Electroencephalogram or EEG, brainwaves for this task. But once we've gathered the data from Boo, it's then my job and my role in the cognitive story to actually use deep learning algorithms to try and understand those brainwaves, to give Boo this artificial communication ability. And out of all the deep learning algorithms I could have used, I've decided to use LSTM or Long Short-term Memory Recurrent Neural Networks for this task. And the reason that I've went ahead and actually decided to use these LSTM neural networks is that these actually showed the best performance when I just tried them out or tested them on some EMG or Electromiogram data that we've already gathered. And now the next step is that the team and I will actually be traveling back up to Boo and we're going to be gathering the real EEG data within around two to three weeks. And with that EEG data, I'm going to train these LSTM neural networks to give Boo the sort of artificial communication ability. But one more misconception that a lot of people have is, well, how are we going to train a supervised learning algorithm on unsupervised data that we ourselves don't understand? And well, it's quite simple. This is supervised data. And that's because there is someone who can understand Boo and that is Boo's mom, which is why we've given her the title of Intimate Interpreter. Boo's mom can understand the very broad concepts that Boo tries to convey. And with her help, we're going to try and live-label the training data as it's gathered from the headset so I can train those neural network systems. If you'd like to find out more about the Cognitive Story Initiative and get in touch, we'd be glad to collaborate at the CognitiveStory.com. And that sums up what I had to say about the Cognitive Story and how we're using the power of cognitive to live better lives. But next, though, to sum up, I'd like to tell you all that all three of these projects, as Tanmay, Deep Spade, and the Cognitive Story are completely and 100% open source. As Tanmay, version 3.1, the new version that I just talked about today, will be open source within the next two to three weeks. Deep Spade, as you just saw, is now completely open source and the Cognitive Story is available on Darwin Ecosystems GitHub page. In fact, one more thing that I find really interesting is that every single one of these projects, as Tanmay, Deep Spade, and the Cognitive Story, were all developed in a Linux environment and all use Git version control. And so, as you can tell, these tools that Linus, Linus Tervalds, has actually blessed us with are really, really useful for not just developers at this point, but really everybody across the world when they're using a phone, when they're using Google, or any other online software, it's all powered by Linux, powered by one of the most popular version control softwares out there, Git. And so, of course, thank you very much to Linus for actually making all of these tools available to us developers. And that sums up what I had to say today. Thank you very much, everyone, for joining in. That's going to be all that I've got. I really wanted to do a live Q&A session, though, because, of course, I mean, we've got two more minutes. But unfortunately, I won't be able to do that right now, but you can ask me questions after my keynote, and I'd be glad to answer them apart from that. If you'd like to reach out to me on any of the following social media, I'd be glad to get in touch. And that sums up what I had to say today. Thank you very much.