Name: Insights: Confronting Deep Fakes, Job Displacement, Bias, And Other Worries Of AI
Uploaded: 2020-02-25
Duration: 4 min 49 s
Description: AI can be marred by bias that adversely affects women and minorities. (November 2019)

[MUSIC PLAYING] JEREMY KAHN: One of the more frightening use cases of AI right now is the advent of deep fakes. These are videos that have either been manipulated or created in their entirety through machine learning. They've been used in some advertising campaigns. Some charities have used them for public awareness campaigns. But unfortunately, they've also been used already in cases of revenge porn. And there is a real fear that they're going to be used for political misinformation campaigns in the future. Obviously, another big concern with AI is about job displacement and the fact that a lot of skills right now that are performed by humans can be performed by machines. I know Francesca you've been doing some work on which task can be automated, you think, and which can't. And you want to tell us a little bit about that research. FRANCESCA ROSSI: Sure. So that's the work that IBM did to the context of the partnership that we have with MIT. And that work is really interesting because it uses AI itself to analyze the impact of AI on jobs. So it's another example where AI itself can be used. So we use machine learning techniques to analyze about 170 million job postings from 2010 to 2017 to see the trends on jobs. But it's important to make the distinction between jobs and tasks within jobs. Because all jobs will be transformed and some tasks will be automated, will pass from human beings to machines. So not the complete job itself, but the specific task. And of course, some tasks are easier to be tackled by machine learning approaches. So those that require the analysis of big amounts of data, the prediction of what happens in the future, and so on. Other tasks are less easy for AI and machine learning to be taken over. But some of the findings of these interesting work is that these passing of tasks from human beings to machines is happening and we see the trend. But it's not happening as fast as we thought. So this is very reassuring because it gives time to train and retrain the workforce. And we have at IBM many different programs to do that. So that the workforce is ready for the transformation. And the people can actually be ready to work with their new job, which where they have mostly to team together with AI algorithms. JEREMY KAHN: Francesca, you mentioned biases as one of the issues facing AI at the moment. And it seems like every time there's a new development, a new model people are really excited about, it doesn't take very long for someone to come along and find out that actually the data set from which the model was trained had a lot of hidden bias. There was just an example of a natural language model that Google came out with in the last 18 months called BERT. It's showing great promise in a lot of things that Google's just change its search engine to be based in part, at least in English, on BERT. But they've found that because BERT is trained from a vast amount of human generated text that's available on the internet, it exhibits gender bias, in particular because it's been trained on all of this text which has been generated over decades when people were maybe less sensitive to some of these gender issues. How do we solve this problem? I'll go to Anton first. How do you deal with this? ANTOINE BORDES: Yeah, I mean, the first step of dealing with this is recognizing that the bias is there in the data and that all data set are biased, which is something that is top of mind for most of the developers now in AI, I hope. And actually, in a company right now something that's very important is that your data set will be biased, so you need to have tools internally to check the bias to see and to try to counterbalance it. So that's the first part. FRANCESCA ROSSI: It's very important to understand also that any one of these issues, whether its bias, and we have another open source tool kits for explainability, another one for robustness. So whatever the issue that you've focused on is not something that can be even identified and resolved by AI experts alone. So it has to be addressed in a very multidisciplinary context and multi-stakeholder context-- those that develop AI, those that deploy, those that use it, and those that are impacted by the decision of the AI system-- just to understand the right notion of fairness to be used in a certain context. So we hope that these open source tool kits are raising the awareness and advancing the understanding of what needs to be done. [MUSIC PLAYING]

Insights: Confronting Deep Fakes, Job Displacement, Bias, And Other Worries Of AI

Uncategorized

AI can be marred by bias that affects women and minorities.

Confronting Worries Of AI

AI can be marred by bias that adversely affects women and minorities. (November 2019)

way, one of the more frightening use cases of a I right now with the advent of deep fakes. These are videos that have either been manipulated or created in their entirety through machine learning. They've been using some advertising campaigns. Charities have used them for public awareness campaigns, but unfortunately they have also been used already in cases of a revenge porn. And there is a real fear that they're going to be used for political misinformation. Campaigns in the future are another big concern with the eye is about job displacement. Um, and the fact that a lot of skills right now they're performed by humans can be performed by machines. I know friends. You've been doing some work on which task could be automated, you think in which can't. And I know you want to talk a little bit about that research so that the work that IBM did the context of the, uh, partnership that we have with M. I t. And that work is really interesting because of uses a eye itself to analyze the impactor over Italian job. So it's another example where a guy can you so we use machine learning techniques that one alive. About 170 million job posting from 12,017 toe. See the trend own jobs. But it's important that you make the decision between jobs and passed within jobs because all jobs will be transformed. But on some tax will be automated will park from human beings to machines. So not the complete job itself, but the specific talk and, of course, some tax. He's here, Toby, tackled by a machine learning approaches. So those that require their knowledge is a big amounts of data, the prediction of what happens in the future. It's on other, less easy for a guy machine learning to be taken over. But some of the findings of the interesting work is that these passing off track from human beings to machines is happening. Oh, and we see the trend, but it's not happening positively fought. So these very we're shooting because it gives time, toe train and retrain the workforce. And we haven't idea many different programs to do that, so that the work even ready for the transformation and the people can actually be ready to work in with the new job, which is where they have mostly team together with a eye on Good Francesca. You mentioned bias, a CZ, one of the issues facing a eye at the moment. And it seems like every time there's a new development a new model people are really excited about. It doesn't take very long for someone to come along and find out that actually, the data set from which the model was trained had a lot of hidden bias. There was just an example of AH natural language model that Google came out with in the last 18 months, called Bert. It's showing great promise and a lot of things that Google's just changed its search engine, you know, to be based in part at least in English, on on Burt. But they found that because Bird is trained from, ah, a vast amount of human generated text that's available on the Internet, it exhibits gender bias in particular because it's been trained on all this text that has been generated over decades when people were maybe less sensitive to some of these gender issues. What? How do we solve this problem? I'll go Antoine first. How do you deal with this? Yeah, I mean, I felt the first step of dealing with it, being recognizing that the bio data and that's all that it or buy it, which is something that that stop of mind for for most of the developer 90 I hope and actually no company right night something that's very important that you know that the city via So you need to have tools internally to check the by acid to see and to try to counter balances. That's the spot, very important toe. Understand also that any one of these issues, whether it's by and we have another open source, took it for explain ability, another one for a box. So whatever the the issue that you focus on is not something that can be even identified and resolved by AI ai people, I expect alone. So it will be address in a very multi disciplinary context, a multi stakeholder context, those that developer guy does the deployed on the use it and those that are impacted by the decision of the system. Just just to understand the right notion of fairness, Toby used in a certain context. So we hope that these open source to kids are, you know, raising your Web nets and making you know, advancing the understanding of what needs to be done