Ovidiu Mățan: Hello, we are at the Leade.rs and we are talking with Jennifer Marsman,a Microsoft Evangelist. She just got a nice presentation about artificial intelligence. Hi Jennifer!
Jennifer Marsman: Hello!
O. M.: Our first question is: How advanced is the Microsoft Artificial Intelligence system and where is it located? In Azure? How is it working?
Jennifer Marsman: We have a number of different Microsoft Artificial Intelligence products kind of spread throughout the whole company. We kind of meet people where they are in terms of Artificial Intelligence and
Machine Learning. So for example, we have something called The Cognitive Services, and those are pre-trained models that you can just call and they do very common Artificial Intelligence tasks such as facial detection (like recognizing my face is right here) and facial verification (recognizing that I am Jennifer Marsman) and emotion recognition and text analytics (being able to detect the sentiment text, and what language it is, and automatic machine translation and the key topic extractions, if you’d like support logs, figuring out the main points in them) and all sorts of things like that. So there are these common tasks that everyone is doing over and over again. We have great pre-trained models where you can essentially just call them via the REST API call then get a bunch of data back.
O. M.: It sounds good. So, basically, a developer can just call your services and get back the analytics results?
Jennifer Marsman: Right, that's a great option if you want to solve one of those problems or something like facial recognition or something that's one of those common problems. You could just use that and then you don't have to train the model yourself. You are just using a model that's already been trained by the folks in Microsoft Research, but, if you want to build your own model for Machine Learning, there are other options you can use as well. One of them is called Azure Machine Learning. Azure Machine Learning is a browser-based system which contains a whole bunch of different modules. So, there are things for data cleaning, removing missing roads, converting variables to categorical, out-of-the-common like the data munging that
you would do in a Machine Learning task. It also contains 25 different algorithms, Machine Learning algorithms that you can utilize as well, again based on our decades of research, in this space. So, you can do something like provide your own data and then apply a neural net to it or Bayesian networks or decision trees or those sort of things, test and try different things, and then produce a model. Then, you get nice accuracy numbers on how well the models are performing, and you can use that to iterate and improve
your model. When you are happy with it, you can hit publish and that stands it up on a REST endpoint in Azure. It generates a security token for you a GUID, and you will need that API key to be able to call it. It's a kind of a nice system. In terms of the deployment story, it is really-really easy because a lot of times deployment is just like the worst part - “How do I set this up in security?” and all the all the stuff that's less fun for me
So all of that is being taken care of, and then you can just call the service So, that's Azure Machine Learning.
And then, we have another option for Machine Learning developers and that is something called The Cognitive Toolkit also called the CNTK and it is a Deep Learning a toolkit that we've opened sourced that allows you to do / build Deep Learning networks. If you've worked with a Google TensorFlow, it's very similar to that. Our performance numbers are very good actually. It exceeded
TensorFlow, at this point.
O. M.: It sounds good. This is good news.
Jennifer Marsman: Yes, it is very exciting and there are a lot of other products that are using Machine Learning inside them, like SQL server. For example, in the 2016 addition, SQL server incorporated some new Machine Learning functionality. Then, we have R server. People are programming in R which is one of the very common languages for data scientists. You can run it in R server. There is Machine Learning throughout the company, but those are some of the big ones that you should know about.
O. M.: It sounds really good. Let's say someone wants to start learning, tomorrow, Machine Learning. Where should he or she look first?
Jennifer Marsman: Great question! So there are a lot of great online
courses that are available. Microsoft runs a couple of MOOCs, massively online courses that are available for everyone. I think we have two of them, We also have a data science certification, so you could work through that course and then get a data science sort of certificate at the end, which is a great thing to add to your resume. That's one option.
Another very popular thing in for developers who are trying to get started in
Machine Learning is the Coursera course by Andrew Ng, who is a professor at Stanford. He put his course on line and that's actually when he founded Coursera. He found it with that Machine Learning course. It is this kind of thing that started Coursera. So, he does a really good job he's a great teacher and he breaks down things like gradient descent and such in a really good way so a lot of people can look at that one as well.
O. M.: Great, everyone should start learning; start to get theoretical knowledge about this. A lot of people mistake Artificial Intelligence (as we have it today) with the real intelligence of a person. So, can you comment a little bit on this?
Jennifer Marsman: Artificial Intelligence in itself is an overloaded
term. Even for us in the field … we fight over what it really means, but if you're talking about the concept of super intelligence, something that could think and act like a human, and make decisions in that sense, then there is a lot of work being done on that. We’re not quite there yet, but there's a lot of work being done in that space. So, Machine Learning is a technology that can enable that because that allows you to use historical data, and find correlations and patterns in that data, and use crazy beautiful math to then
be able to make predictions, build a model of what that looks like and then make predictions. We’re very good at being able to make predictions of based on historical data, we’re very good at being able to cluster things together and group similar things together, we're good at reinforcement learning, so like the work that was done with AlphaGo that Google did, but, yes, very impressive stuff. It uses a form of reinforcement stuff; it gives a kind of reward. It's kind of a turn-based thing and, with each turn, if you're getting closer to go, you get a reward. If not, you can get a penalty if you're not doing so well.
O. M.: It’s like the snakes that started, at some point, to fight with each other. I don't know who created this snake game with AI (link ), but, at some point they started to fight with each other for resources.
Jennifer Marsman: Yes, such conditions are always fun. Yes, but there's a lot of work being done and a lot of experimentation. So, I think we will probably get there.
O. M.: Ok, it sounds good. Where do you see we will be 10 years from now? Will we get that singularity or real intelligence?
Jennifer Marsman: Actually, at this point, I'm more scared of fake intelligence, I guess. The real intelligence I know myself is the one on which other Machine Learning specialists have spent so many panels debating and discussing - I know there are organizations that are formed to discuss this, Elon Musk has done some work in that as well. They talk about Artificial Intelligence ethics, the legal concerns and all of those, so it's something that's being fiercely debated and discussed in the field. We don't necessarily have all of the answers yet, but there are a lot of questions around, things like ethics, in certain situations. Some of the things that have made me cringe a little bit are like, for example, the ones done in Chicago, where they were doing predictive policing, so using historical data of - you know - who's arrested to find out the kind of people that were more likely to be arrested. But, if there's bias in your historical data then that's what the Machine Learning algorithms will give you, because it's not magic. It's just making predictions based on what it's all right, so “garbage in, garbage out”. You have to really think about that. There are a lot of ethical questions like that and legal questions. In the example of autonomous vehicles, self-driving cars, if one of them it gets in an accident and kills someone, who is liable? Whose fault is that? Is it the people in the car, the people in the other car? Is it that the developer who wrote the code? Is it the if they were using in algorithms. Is it the big companies’ fault? How do you figure it out? How do you unravel all that and, since a lot of the way neural networks work is a little black box, it's hard to kind of unravel who takes the blame, I guess. It's a really hard problem that a lot of us are struggling with. At the end, it is a set of data, right? Yes, at the end, it is just data … it is all data at the end.
O. M.: Thank you Jennifer for the interview.
Jennifer Marsman: Thank you very much. Thank you!
Marți, 27 Septembrie, ora 18:00
Centrul de inginerie Bosch ClujFacebook Meetup StreamEvent YouTube
de Bianca Leuca
de Ovidiu Mățan
de Vlad But
de Bálint Ákos
de Oana Călugar
de Adrian Vîntu
de Andrei Adam
de István Kiss
de Cristina Juc
de Vlad Ciurca