Ep. 252 - Alyssa Simpson Rochwerger and Wilson Pang, Authors of Real World AI: A Practical Guide for Machine Learning on Creating, Building, and Maintaining AI Projects

19:34
 
Share
 

Manage episode 293350455 series 2822865
By Brian Ardinger, Founder of NXXT, Inside Outside Innovation podcast, and The Inside Outside Innovation Summit. Discovered by Player FM and our community — copyright is owned by the publisher, not Player FM, and audio is streamed directly from their servers. Hit the Subscribe button to track updates in Player FM, or paste the feed URL into other podcast apps.

On this week's episode of Inside Outside Innovation, we sit down with Alyssa Simpson Rochwerger and Wilson Pang, authors of the new book, Real World AI: A Practical Guide for Machine Learning. We sit down and talk about some of the biggest misperceptions about AI, as well as some practical advice on how to tackle creating, building, and maintaining AI projects. Let's get started.

Inside Outside Innovation is the podcast to help you rethink, reset and remix yourself and your organization. Each week, we're bringing the latest innovators, entrepreneurs, and pioneering businesses, as well as the tools, tactics, and trends you'll need to thrive as a new innovator.

Interview Transcript with Alyssa Simpson Rochwerger and Wilson Pang, Authors of Real World AI: A Practical Guide for Machine Learning

Brian Ardinger: Welcome to another episode of Inside Outside Innovation. I'm your host, Brian Ardinger. And as always, we have another amazing set of guests. Today we have Alyssa Simpson Rochwerger and Wilson Pang, authors of the new book called Real World AI: A Practical Guide for Machine Learning. Welcome to the show.

Wilson Pang: Thank you, Brian. Really excited to be here.

Brian Ardinger: We are excited to have you both here. This is an exciting world of technology and new trends that are happening. AI is obviously on the forefront of a lot of people's minds. And I'd love to get your input on what do you really mean when you say real-world AI and how does that differ than people's perceptions out there?

Alyssa Simpson: I think one of the things that we wanted to address with this book is sort of the in-between space between, you know, the hype and maybe what you read about AI and the headlines, or what you see AI or very smart futuristic systems depicted in the movies. And then, you know, also a very academic or technical approach to machine learning that you may have come across a textbook or learned in school.

And this is kind of the middle reality, right? Is, you know, what are real companies who are using machine learning based technology? How are they using it? You know, what struggles are they having? What successes are they having? And then how does it work in the real world. In the reality that we all live in and share and products that you probably use frequently in your everyday life?

Brian Ardinger: Well, I think there are a lot of misconceptions about what AI even is. Maybe Wilson, can you tell us a little bit about some of the myths or misconceptions about AI that people commonly struggle or fumble over?

Wilson Pang: There's a few major common misconceptions. Number one, it's really, a lot of people think AI, they think is the kind of a machine can do whatever human can do, right? This is called Artificial General Intelligence. Basically, AI can repeat human. So that's happening in those small ways but it's far away from the reality. In reality, what AI, all the real-life AI, is really the application, which can be taught or learn to carry a specific task without being programmed to do so

So, the machine can learn from data. And then performing some tasks. So that's kind of a like what real world AI is. So that's number one misconception. Number two misconception is that people are thinking to build an AI, the team needs to spend a lot of their time to tune the model, work on the model, and get the best performance. It is more related to the signs or the model tuning.

In reality, science is a big part of AI, for sure. But for the AI team, they spend a majority of their time, on data. Need to collect the right data, clean up the data, make sure the data has all of the representatives and also make sure that data has quality. And then the data complete the magical part is to really improve the AI performance. So those are the two major misconceptions. I hope everyone can really learn and understand that.

Brian Ardinger: You know, you often see two sides coming out, when you talk about AI. You have one side of the spectrum where everybody talks about AI as a panacea of opportunity. And it's going to change the world for the better. On the other side, you have the folks that think AI is a massive danger and a menace and a lot of unknown repercussions. Where do you guys fall on that spectrum?

Alyssa Simpson: I'll argue all sides of that one. I think both are true and neither are true. As with anything, machine learning and AI technology is disruptive. It already has changed the world as we know it and will continue to be an incredibly powerful and disruptive technology in almost every industry. On the other side, it's just technology, right? And there's a lot of things that are powerful and it's only as powerful as what you put it towards and how you shape it.

And in other ways, it's incredibly brittle and really narrow as Wilson was referring to, this is not a magic eight ball. It's not generalized. AI often is very narrow and specific, and only is able to accomplish a fairly narrow and specific tasks that you train it for, if you have all the data available to train it really successfully to perform that task.

And so, you know, what we see in reality is companies having a lot of success. If they are able to find a use case that this technology can perform really well and do things that previously were undoable before and open up new streams of opportunity.

Brian Ardinger: Where are some of the industries that are getting the most out of AI right now. And where do you see the trends moving, when it comes to folks capturing the benefits of AI?

Wilson Pang: AI only become a buzzword in recent years. But in reality, AI has been there for a long time and high-tech industry who has big company like Google, Facebook, all those tech giants. They have been using AI for a long time. And AI has been used to optimize their price sprints, to help you find the right product, to give you a better recommendation. To show you a better content.

So those have been there for a while. The usage of those AI technologies is pretty mature, and that's also almost embedded in every part of their product. So, for high-tech industry, very mature. Meanwhile, in recent years, it's also a lot of other industries, they are catching up. Like the finance industry. The medical industry. All kinds of industries are catching up.

So, you can see the train of AI is already, the adoption of AI has become much broader. And also, the speed to adopt AI is also a leverage spot. Meanwhile, as you mentioned earlier, AI can really bring on a lot of benefits. Can also create a lot of harm. So, with this even wider adoption, we really need to make sure people understand how to use AI in the responsible way. So that way we can get the benefit part, but meanwhile not really doing a lot of damage to the society.

Brian Ardinger: Yeah, you definitely hear that. And you hear the use case scenarios when it goes bad and that's oftentimes hyped up. I actually just read this morning. I think it was a fast company article talking about how the New York City Council is considering new rules meant to curb bias in hiring when it comes to AI and putting AI in hiring practices and that. What's your take on some of those challenges that you're seeing in that particular space.

Alyssa Simpson: Yeah, I know Wilson and I both feel really passionate and strongly about the importance of responsible and ethical applications of AI. You know, one of the places where I see challenges is that people don't think about the potential harms or bias from the beginning.

And don't look at their data and training set really critically, with a critical eye to inform whether or not that training set is appropriate for the use case that that model will be applied to in the future. So, in take your hiring example, you know, there was a famous case a couple of years ago, where Amazon tried to use machine learning to look at resumes and figure out which ones are going to be the strongest candidates and put those forward.

And, you know, they found it was really biased from a gender perspective. That's because the training data set that they were using was mostly men because Amazon mostly had men working there. And so, the system wasn't trained to be biased against women. The training data had more examples of men having more success at Amazon. And so, the model ended up amplifying the biases that were there already.

So often, that's sort of a classic case of what's happening, Is that the model is supercharging, right, any existing biases in our society. And so, as we all know, are at least in the United States, right? There's a lot of biases against women. There are racial biases, you know, there's all sorts of different challenges that exist in hiring practices.

And if you are not careful, or if you are not actively sort of proactively correcting for some of these challenges in the data, you can end up with a model that reflects those back to you. And at that times makes it worse.

Brian Ardinger: With that, are there particular ways or opportunities for folks to do better job at that particular training. Is it just coming to light now? We're getting better at this, and so we're, we're learning from these mistakes. And so, they're being amplified, but we're getting better at it or walk me through the state of AI and specifically around the bias of the training of the system.

Alyssa Simpson: Yeah. I mean, I have to think, unfortunately, most people who are training models are not thinking about this at the beginning. You know, perhaps more and more, but I think it's often really overlooked at the data stuff. And they only realize it later once deployed some big into production, like, Oh woops. You know, but if you think about it from the beginning, it's a lot easier to correct for.

So, I'll take a, perhaps a different example. If you're building like a voice assistant, like a Siri or an Alexa or something. And you want it to work for people in the United States? Well, that means you need to capture voices from people in the United States, right? Not just white men. You need to also capture voices from people who speak English as a second language or from children or from elderly people or from people in the South who may have a Southern drawl.

So, you need your training data to be reflective of the population that is going to be served or be interacting with the system. So, in say New York's hiring case, if they're building an AI model and they don't want it to be biased, well, you know, look at the candidate pool, right. But you know, everyone who lives in New York state, do you have a representative sample in your training data of everyone who lives in New York state or is your training data biased already you know.

And there's a lot of different types of biases. This is like, you know, sort of a sample bias or collection bias. There's a lot of different ones, you know, but is it bias towards a particular demographic that you could look at it more critically and then include a much broader set of examples that are more reflective? That's probably the easiest thing that most people can do, but certainly not the only thing.

Brian Ardinger: That makes sense. So, let's dig into the book a little bit. Wilson in the book you outlined some of the key characteristics and the things that people should be looking at and working on when deploying a new AI system. What are the key insights in the book that people should take from it?

Wilson Pang: Our book actually walks through the whole journey, from identify the problem, to problem of the team, from defining strategy to how to get data. And also talking about a buy a model or build a model, right. There's a lot of details around how to really develop the AI journey.

If we can make a quick summary there's a few super important points there. Number one is really to identify the right problem to solve. And you always also want to start from small instead of wanting to boil the ocean, in the beginning. So, the right problem, which should be a, really small problem or a specific, and also, you know, what a success looks like if you are successful, and you knew how to measure that.

And that will give you a good start. And then after you have that right problem, then you need to assemble a right team. On the right team, you need to also have multiple different roles. Besides everybody can think of I need a data scientist. I need a machine learning engineer, right?

You also need the people who really need to know the business, either product manager or business eminence to help you to understand the problem. And then a data scientist can connect the department to a science problem. I also needed engineers to have you deploy this tool production.

And after you deploy to production, you also need to monitor and the operation, all of those AI model, right? So, assemble a right team is super important. So that's number two. After that, I think another super important part comes is really to get the data. Get the right data.

So, the right data also has a lot for meanings on there, that can mean you need to have the data quality, right. You need to have the data to be representative of different categories or groups, right? Alyssa mentioned earlier, if you don't have the right data to represent the different group of users, you can create a lot of bias. So, data cleanup, data quality. That's another big piece. After that, you probably get the right data to run properly. You will have a successful pilot and they get some good results.

Or the company will help you to celebrate the success. And they'll also want to what's next. So that comes with data. You need to scale from a pilot to production, where you also need to invest a lot more to consider how you come to the production. How you can access the data regularly. You need to refresh the model. You need to refresh the data. And also, you probably also need to define some common process to help you manage the data, manage the model.

And that you also need to continue there, like at a production scale. Now we are not really running a pilot that you, maybe your model is serving meetings or hundreds of meetings for people. You have to make sure the model is available. And you can deploy the increase the demand. So, there's a lot of things to consider, once it comes to the production.

And also, once you start to have more and more productive use case, you know, AI is impacting the business in a huge way. So that is the part, consideration like a lot of ethical consideration or responsive governing, a lot of those kicks in. So you need to define not just the product, but a program, also a product to make sure you can get the benefit of from AI across the whole organization instead of just the one product.

So those are the key points to consider when you really want to deploy successful AI. I know that's a lot. And I think I only covered a small piece of it. And if you might want to learn more check out our book.

Brian Ardinger: That's an interesting point because I think a lot of companies want to dip their toes into AI and get better at it. But there's so many different moving parts. So, what recommendations would you have for ordinary folks, ordinary companies trying to get up to speed on this topic and prepare.

How should they look to find an expert on the outside to lead them through this process? Should they start building a team? What are some steps that people can do if they want to get more into this AI world?

Alyssa Simpson: Well, there's a book that we wrote, that might provide step one. Step one, read the book. You know, but in all honesty, you know, I think that's part of the reason we wrote the book is because we got asked this all the time, and this was a way for us to kind of brain dump on paper in a little bit more scalable way of like, Hey, here's a quick guide of how to get started.

There is no one size fits all approach. It depends on your company size. It depends on your budget. It depends on your risk appetite and a lot of other things. But I think generally speaking, my guidance would be, find a problem that is well-suited for AI. We talk about this in our Goldilocks chapter, which is pick the right problem that you want to solve. Is really important to solve for your business.

It's not a good approach to say, Hey, I want to do AI. Let me like sprinkle this AI magic on top of my company. And that's not how it works. It's Hey, I have a really hard problem to solve. And, you know, I think it's really well suited to be solved by machine learning technology. And that's a much better way to approach it, right.

It needs to be a problem that is really important to your business. That's critical to get done, right. That has high value that you have a lot of data in order to solve that problem, that you can train a model to help you solve.

Brian Ardinger: You've been around in the AI space for quite a long time and seen a lot of changes. What are you most looking forward to in the years ahead?

Wilson Pang: I think there's a lot of evolvement in this AI data field. There are a few areas, particularly I'm super excited about. Really the kind of this big language model are, you probably are really seeing like what kind of a magic AI can help you to play. It is those kind of big language models, they leverage almost every piece of information on the internet, seeing all like terabytes of data and also almost capture all the knowledge from all over the word and with all of those data, they can perform a lot of amazing tasks, like write an article for you, and also maybe even write a program for you. Right.

I know it's still early. They're, the people, they discover ability, but they don't really know how to really use this. And also, there's still a lot of use cases that even the big model cannot really handle. Right. But I do think they are missing. So now it’s you GPT-3. When it comes to GPT-14, GPT-15, there might be a lot of magic it can do to this society.

So that's the one part I'm super excited. On the other hand, also say AI now is not only a research problem, or a problem where only a high-tech company can tackle. Now I see a lot of different tools, to have a developer to build AI. Or maybe have a data scientist to deploy AI in production.

So, with all of those tools, I feel the adoption for AI will be much broader. And the speed to adopt AI will be much faster. So, obviously a lot of different use case from any kinds of industry can be supported by AI. Some of those use cases, maybe we can, we cannot even imagine today. So those are the two areas I feel super excited

Alyssa Simpson: For me. It's more personal. I'm excited to like, not have to repeat myself to Alexa. Or you can turn the lights off. Or I'm excited that, for new drug discoveries. You know, I suffer from migraines and, you know, neurology and the best medicine in the Bay Area, which the fanciest doctors have not been able to, you know, help me that much.

And you know, I'm excited for new drug discoveries and things that are going to impact my everyday life for myself and those that I love. Whether or not it's a customer experience improvement or real big medical breakthrough.

For More Information

Brian Ardinger: Absolutely. It's going to be an amazing world. We're, we're looking forward to it as well. We'd love to have you back on the show in the future to continue this conversation. But in the meantime, if people want to find out more about you Alyssa or Wilson or the book, what's the best way to do that?

Alyssa Simpson: You can find me on LinkedIn or AlyssaSimpsonRochwerger.com. Our book is available on Amazon. We'd look forward to hearing your feedback and thanks so much for having us, Brian.

Wilson Pang: Yeah. And you can always find me through LinkedIn. And Brian thanks again for having us here.

Brian Ardinger: Well, thank you both for being on Inside Outside Innovation. I'm very excited to see where this world is going, and I appreciate you all sharing your insights.

That's it for another episode of Inside Outside Innovation. If you want to learn more about our team, our content, our services, check out InsideOutside.io or follow us on Twitter @theIOpodcast or @Ardinger. Until next time, go out and innovate.

FREE INNOVATION NEWSLETTER & TOOLS

Get the latest episodes of the Inside Outside Innovation podcast, in addition to thought leadership in the form of blogs, innovation resources, videos, and invitations to exclusive events. SUBSCRIBE HERE

You can also search every Inside Outside Innovation Podcast by Topic and Company.

For more innovations resources, check out IO's Innovation Article Database, Innovation Tools Database, Innovation Book Database, and Innovation Video Database.

212 episodes