字幕表 動画を再生する 英語字幕をプリント [MUSIC PLAYING] LAURENCE MORONEY: Hi, everybody, and welcome to "TensorFlow Meets." In this episode, I'm chatting with Haohan Wang-- HAOHAN WANG: Thank you for having me here. LAURENCE MORONEY: --who's been doing some really, really cool work with neuroscience, and TensorFlow, and machine learning, and all kinds of cool stuff, and even a new course. So you've been working on Dyad X Machina, which is a beautiful name for what looks like a beautiful site. Could you tell us all about it? HAOHAN WANG: Sure. So I will start from the name Dyad X Machina. Dyad means two as one, which is me and my partner, Christian Fanli. LAURENCE MORONEY: And Christian is going to be on a future show, which is great. HAOHAN WANG: Yes. So machine-- things we constantly work with machine. And we're at the intersection of deep learning and affective computing. So as for why we started Dyad X Machina, I think the story really starts four to five years ago when we first met. So I was a student in finance and he was working in human behavior area. And we started to collaborate because we both shared the same strong interest in machine learning. LAURENCE MORONEY: Right. HAOHAN WANG: So one of the first machine learning projects that we worked on together is to create a trading algorithms. LAURENCE MORONEY: OK. HAOHAN WANG: Well, so Christian challenged us to think outside the box instead of using the traditional technical or fundamental analysis into your algorithm, why can't we add some human behavioral elements, like emotions, into your algorithm? LAURENCE MORONEY: OK. HAOHAN WANG: Well, so we end up including sentiment into our algorithm. And surprisingly, sentiment is a great training-- a very indicative training signals. LAURENCE MORONEY: I can see that because I'm sure when some people are buying, then everybody jumps on board. And when some people are selling, everybody-- you know, that kind of thing. There's a lot of sentiment. HAOHAN WANG: Yeah, early signal. LAURENCE MORONEY: Yeah, definitely. HAOHAN WANG: Yeah. Well, so after that, I think it really brought emotion into our attention. And then we started to study a lot more into this field, affective neuroscience especially. And we read numerous book in the field. And I think we started to notice a problem, which is how society at large tends to ignore how important emotion is, especially like in our daily decision-making process because people tend to overvalue cognition and rationality over emotion. LAURENCE MORONEY: Interesting. HAOHAN WANG: Yeah. So I think, well, that's really the starting point that we think probably it's our job-- it's Dyad X Machina's job to bring this affective layer back into people's daily life. LAURENCE MORONEY: So it's you saw an opportunity to bring emotion into what are traditionally logical decisions? HAOHAN WANG: Yep. There is where Dyad X Machina started, yeah. LAURENCE MORONEY: Interesting. And now, a lot of the work that you've been doing and a lot of the learning that you've been doing, you're turning into a course now, right? HAOHAN WANG: Yes. So start from there actually-- well, here is where deep learning and TensorFlow came in. So we were studying machine learning, but then we discovered deep learning. So I think so here is also the point that we hit a point that we cannot make any more progress because we're both working full time and we'd dedicate all our free time on reading books-- LAURENCE MORONEY: Those day jobs just get in the way, don't they? HAOHAN WANG: Yep-- trying to understand machine learning and affective computing. So I think we had a discussion and made a hard decision-- I took one year off, focused fully on deep learning and affective computing. So here, you can think of this course-- and we end up making this course because this course is really the synthesis of our learning because we really dedicated a lot of time and effort on learning deep learning. And we think this course will be a great start for people who are new to deep learning to get started. But more importantly, I think our main focus is to help people who want to use deep learning to the field that they are passionate about to be able to get started. LAURENCE MORONEY: So your course is about applied deep learning with TensorFlow and Cloud AI? HAOHAN WANG: Yes. LAURENCE MORONEY: There's a lot in there. So what kind of content do you have? HAOHAN WANG: Well, so this is our very first course. And we're a little bit ambitious trying to put everything in there. But the course is really meant to help people who are new to machine learning to be able to build their first deep learning model and to take it all the way to deploy their model as production-level API. LAURENCE MORONEY: OK. HAOHAN WANG: Then we move on to talk about the basics of deep learning and how to design an experiment with some typical neural networks using Keras. LAURENCE MORONEY: Keras, yep. HAOHAN WANG: Yep. And then we move on-- we dove deep into TensorFlow. We start from low-level TensorFlow. We introduce the concept like dataflow graph and a TensorBoard, then we move on to high-level TensorFlow. Then help people to build a model in the cloud, train the model, evaluate the model, and eventually deploy their model as a production-level API. LAURENCE MORONEY: So the deployment part is really fascinating to me because there's lots of great material out there about training models and maybe doing a little bit of a test, but making it real world, making it applied is really cool. HAOHAN WANG: Yeah, that's our intention-- help people to apply it to the field they're interested in. LAURENCE MORONEY: Nice. And we'll put a link to the course in the comments below. So you've started with neuroscience, and then you've moved into taking a year off and creating a course. And you've obviously gotten very deep into machine learning and you've gotten very deep into neuroscience and the intersection between the two of them. What advice would you give to people who are just starting out? HAOHAN WANG: Well, yes. So I'd like to talk a little bit about-- connect it back to our learning experience of a journey of deep learning and possible a little bit about affective neuroscience. So we also summarize this four P's of learning that I'd like to share here. LAURENCE MORONEY: The four P's of learning? HAOHAN WANG: Well, yeah. LAURENCE MORONEY: OK, I'll try to remember them myself. HAOHAN WANG: We published an article on our website so people can check it out. LAURENCE MORONEY: Ah, so we've got a link for that, so I don't need to remember it. HAOHAN WANG: Yeah, and you don't have to remember it. LAURENCE MORONEY: Great. HAOHAN WANG: So the first P-- papers and books. LAURENCE MORONEY: Papers, OK. HAOHAN WANG: Well, I think it's really the foundation because we make sure we read papers and books every single day. And we got up 4:30 AM, the first task is to read paper. LAURENCE MORONEY: Wow. HAOHAN WANG: Yeah, deep learning papers. So at the beginning, we read some more like basic deep learning paper to know what is trending in the field or how people solve problems with deep learning. Then we move on to be more specific about our domain-- how people use deep learning to solve like neuroscience problems-- affective neuroscience problems? LAURENCE MORONEY: So 4:30 AM for the first P? HAOHAN WANG: Yeah. LAURENCE MORONEY: I think you've lost me already. HAOHAN WANG: Yeah, first day. LAURENCE MORONEY: Just kidding. So papers. And what was the second? HAOHAN WANG: Paper and books. LAURENCE MORONEY: And books, yeah. HAOHAN WANG: Well, paper and books. Second P is practice. LAURENCE MORONEY: Practice, OK. HAOHAN WANG: Yeah, so I think practice is very important. Especially like I think I am very passionate about like every day, I will use TensorFlow to build some-- start from the basic neural networks and understand like dataflow graph. And like, I'm very interested in using TensorBoard to visualize my training, and evaluating resource, and how to tune my model's hyperparameters. So I think then practice. And you can find those problems very easily on Kaggle and Crowd AI. LAURENCE MORONEY: Kaggle, right. Yep. HAOHAN WANG: So the third P, which I think personally is very helpful for my learning, which is preach. LAURENCE MORONEY: Preach, OK. HAOHAN WANG: Yeah, you can think of lecturing or teaching. So every day, I will give Christian a mini-lecture about what I learned during the day about deep learning, some theories, or even neural network I built that day. So I think it's mutually beneficial. It helped me to consolidate my knowledge and helped him to have an overview about what I learned and also making some progress from his side. LAURENCE MORONEY: Excellent. HAOHAN WANG: Well, last P, we think is really helping us to get where we are right now is passion. LAURENCE MORONEY: Yeah. They definitely need passion. HAOHAN WANG: Yeah. So well, first like I mentioned, we saw this great potential of deep learning and TensorFlow that you can leverage to really bring the transformative change into the domain because this is new technology. And regularly in the past, at least based on our study about the effect of neuroscience, no one has ever leveraged this technology to build any algorithm yet. So we set out to make that happen. LAURENCE MORONEY: So bringing emotion into algorithms. HAOHAN WANG: Yes. LAURENCE MORONEY: It sounds fascinating. So you started this journey as you were doing financial trading stuff-- HAOHAN WANG: Yep. LAURENCE MORONEY: --and then you just took a look at emotions in that. You know, from out of that has grown Dyad X Machina. HAOHAN WANG: Yeah. LAURENCE MORONEY: And now, you have a course. You have your daily four P's that you're doing. It's like there's a lot of amazing stuff there. So what's next for Haohan? HAOHAN WANG: Yeah, so next for me and Dyad X Machina is we're right now, still working on running experiments every day and doing research in affective computing area. We're trying to put things together to be able to eventually present to the-- share with the community. So we will be working on this part of the study. What's more is another thing about the course. LAURENCE MORONEY: OK. HAOHAN WANG: We are planning to make a new course. LAURENCE MORONEY: Oh, wow. HAOHAN WANG: And we plan to deliver the course by the end of this year or the beginning of next year. The course will be called "Affective Computing and Deep Learning using TensorFlow." LAURENCE MORONEY: "Affective Computing and Deep Learning using TensorFlow?" HAOHAN WANG: Yes. LAURENCE MORONEY: I'll look forward to it. HAOHAN WANG: Yeah. So in this course, we will be teaching people how to use TensorFlow and the Google Cloud machine learning engine to stream in your psychophysiological signals over time and to process your signals. And eventually, the algorithm should be able to make a real-time prediction about your affective state. LAURENCE MORONEY: OK, excellent. So your course on affective computing, it's going to be coming out pretty soon-- maybe the end of this year or sometime early next year. And maybe after the course comes out, we'll have you on again to talk about it. HAOHAN WANG: Yes, I would love to. LAURENCE MORONEY: So we'll look forward to it. So thanks so much, Haohan. And thanks, everybody, for watching this episode of "TensorFlow Meets." If you have any questions for me or if you have any questions for Haohan, please just leave them in the comments below. And whatever you do, don't forget to hit that Subscribe button. Thank you. [MUSIC PLAYING]
A2 初級 Dyad X Machina: 機械学習に感情を持ち込む (TensorFlow Meets) (Dyad X Machina: bringing emotion into machine learning (TensorFlow Meets)) 4 0 林宜悉 に公開 2021 年 01 月 14 日 シェア シェア 保存 報告 動画の中の単語