Name: Insights: Who's Controlling AI?
Uploaded: 2020-02-25
Duration: 3 min 3 s
Description: Dario Gil of IBM Research, and Richard Socher of Salesforce, discuss the many concerns people have with AI. (July 2019)

[MUSIC PLAYING] RICHARD SOCHER: And AI is not dual-use but omni-use technology. In the end, we don't say, oh, what about computers? They can be used for good and bad. Yeah, computers can be used for anything. And it gets even simpler, like, a hammer. You can use a hammer to build a house. You can get a hammer to hurt people. You can use the internet to do great things. You can use cars to do-- AI will be everywhere. Every single industry will be impacted by AI. And in the end, it's up to us to define what tasks we want it to do. And then when we define those tasks, we also need to think about the data sets. AI in the end is only as good as the training data that we give it. And humans have to think carefully about ethics. And I think the field of AI has now woken up to that. DARIO GIL: So I think one of the ways that we've got to address this is to put the responsibility back to who is creating this software and what is their purpose, and what is their intent. And the accountability needs to rest with the people and the institutions that are creating this software. And we should really avoid the tendency to say as if AI as an external agency was doing this on its own. Fairness, as an example, like if we look in the context say of banking, and you're trying to build a predictive model to be able to assign credit as an example, it's based on learning from past examples. So if you just naively went and you say, well, what credit approvals have we done over the last 20, 30 years? And let me use the approvals that were generated as a means to train this model. Well, what would happen most likely is that you'll end up with a model that is likely to give more credit to men then to women. And that's not necessarily that the neural network is biased. It's the data sets that we're done were biased. So a way to compensate for that issue is you got to have robust AI engineering methodologies to be able to test for unwanted bias. And particularly, for example, if you have protected variables that are protected by law that you can not use as a means to do discrimination from the perspective of assigning classes. You just have to be able to detect that. So we believe explainability, fairness, robustness, some of these AI models are subject to adversarial attacks. And you got to have software to be able to test it. Five years ago or so on, it was just mostly the realm of a relatively few large players and of course, specialized startups that were in there. Now you go and back to the AI like software, you can go to any financial institution, insurance companies, health companies, automotive, and they want to be co-participants and co-creators of AI. So a trend that has happened has been to democratize and give tools to people, so that they can create their AI efficiently, and how you debug it, and how do you create AI engineering as a discipline just like we have done in the past with software engineering. And we will create our program, and debug it, and so on. We need all of that tooling to allow everybody to also co-create and integrate with each other. 

Insights: Who's Controlling AI?

Uncategorized

IBM's Dario Gil on the many concerns people have with AI. 

Who's Controlling AI?

Dario Gil of IBM Research, and Richard Socher of Salesforce, discuss the many concerns people have with AI. (July 2019)

I is not really but on the use technology. In the end, we don't say, Oh, what about computers? They can be used for good about your computer's confused for anything, and it gets even simp like a hammer. You can use a hammer to build a house with a hammer to hurt people. You can use the Internet to do great things. You can use cars e. I will be everywhere. Every single industry will be impacted by eye, and in the end, it's up to us to define what tax we wanted to do. And then when we define those tasks, we also need to think about the data set. A. I in the end is only as good as the training data that we give it. And humans have to think carefully about ethics. And I think the feel if the eye has now woken up to that. So I think one of the ways that we gotta address this, I put the responsibility back to who is creating the software on DDE. What is the purpose and what is their intent on day accountability needs to rest with the people on the institutions that are creating the software, and we should really avoid the tendency to say Aye, aye, as an external agency was doing this on its own fairness as an example, like we look in the context of banking on you're trying to build a predictive model to be ableto find credit as an example you know, is based on learning from past examples. You just naively went and you say, Well, what credit approvals have we done over the last 2030 years? And let me use the approvals that were generated as a means to train the model. What would happen most likely is that you'll end up with a model that is likely to give more credit to men and a woman. And that's not necessarily that the neural network in bias is the data science that were done. We're biased. So I'm way to compensate for that issue is you gotta have robust engineering methodologies to be able to test for unwanted Maya and particularly, for example, if you have protected variables that are protected by law that you can not used as a means to do discrimination. From the perspective of a signing classes, you just have to be able to detect that. So we believe explained ability, fairness, robustness These some of these models are subject with personal attacks and gonna have software to be able to test it. Five years ago or so on, it was mostly the ram off relatively few large, large players and, of course, specialized startup that were in there. Now you go on back to the eyes like software. You can go toe any financial institution insurance company's health care companies out of motive, and they want to be poor participants and co creators. Off I sew a trend that have happened has been to democratize and give tools to people so that they can create their efficiently on how you debug it. And how do you create a engineering, a discipline, just like we have done in the past with software engineering, right, and we'll create a program and debug it and so on. We need all of that tooling toe, allow everybody to also co create and integrate with each other