Neo4j’s Amy Hodler on Machine Meets World

Amy Hodler, Director, Graph Analytics and AI Programs at Neo4j, joins Infinia ML’s James Kotecki on Machine Meets World, the interview show about artificial intelligence.

The video show streams live on YouTube weekly. It’s also available in audio form on Apple Podcasts, Google Podcasts, Spotify, and Stitcher.

Recorded June 16, 2020.

Key Quote

“I definitely think we need a public conversation about what our expectations are for AI. So it’s not just ‘I want more accurate AI.’ It’s do we want AI doing particular things? We need to bring together, I would say, people from different walks of life.

“So you’ve got your scientists that we all think about, but also your ethicists and your business leaders, your citizens, your policymakers, and have more of a joint dialogue about what we want from AI systems. They are different than software we’ve had in the past because of the reach and the power that they really have the ability to amplify the best and the worst of what we can do.” [20:40]

Other Episode Highlights

Striking Examples of AI Failure [5:10]

Why Talking to Data Scientists Gives Her Hope  [18:45]

Bias Isn’t Always Bad [24:35]

See Also

Amy Hodler’s Speech on Responsible AI (November 11, 2019)

Show Suggestions?

Email mmw@infiniaml.com.

Full Transcript

James Kotecki:
And we are live from Infinia ML. This is Machine Meets World. I am James Kotecki talking artificial intelligence with my guest today Director of Graph Analytics and AI Programs at Neo4j, Amy Hodler. Welcome.

Amy Hodler:
Thank you. Nice to be here. Thank you, James.

James Kotecki:
So let’s start with some context before we dive into all these media ethical issues that I’m really excited about. Can you just quickly explain Neo4j and what your title and role is there to someone who maybe doesn’t even have that much familiarity with the industry?

Amy Hodler:
Yeah, absolutely. So Neo4j is the leader in graph databases and technology and to simply put it, graphs are a representation of complex networks of things. And that just means things like a social network, an economic network, a biological network, how are things connected? And Neo4j basically codifies and persists these things in a database. And then they add on capabilities like graph analytics so that people can make more sense of things that are non-intuitive.

Amy Hodler:
So a lot of times there’s connections in our data that you can’t even see if you’re looking at it in tables and rows. So graphs and graph databases save those connections and persists those as more natural, like you would on a white board with like circles and lines of how things go together. The reason why we do that is that people can kind of have those inferred understanding and it would help people like journalists. So investigative journalists in the Panama Papers find things. It might help a bank find fraud in transactions.

Amy Hodler:
So those are the kinds of things that people use graphs and graph databases for. My role as the Director of Graph Analytics and AI Programs, it means I do a lot of marketing, a lot of talking, a lot of research to people about what they can do with these, what you can use advanced analytics graph analytics for, and also how they fit with AI.

James Kotecki:
So let’s talk more about the AI fit. Is what you’re talking about because people have different definitions of AI. So one of the funny things about AI is you always have to continually redefine what it is you specifically mean by AI and in what context. So is AI and inherit part of graph analytics or is it something for lack of a better term extra that you’re kind of adding onto it or are they part of the same thing as you conceive of it?

Amy Hodler:
Well, I think of AI systems, because it’s really one thing as anything that’s trying to make a probabilistic decision similar to the way a human might. So with context, because humans use context to make decisions. Graph because it is all focused on relationships and how things are connected. Graphs can provide context for AI. So I wouldn’t call graph itself AI, but what it can do is give an AI system a better way to make decision. An example might be something like a chat bot. And somebody says, “Hey, I’m going camping in Tahoe next weekend with my husband.” And he might say, “Well and I need a tent.”

Amy Hodler:
What kind of tent do you send them to on your chat bot? You probably want to look at the weather in Tahoe next week. Definitely has to be big enough for two people. And maybe because it’s Tahoe, you are providing a little more of a high-end tent because you’re making some kind of judgment there. So providing context for AIs to make better decisions and also machine learning predictions as well. You can use relationships to improve your machine learning predictions.

James Kotecki:
Well, providing context for AI and ML to make better decisions is right at the core of what we’re talking about when we’re talking about AI accountability and ethics. So obviously these things are related, but as we dive into that topic, how did you first kind of come to that topic? You have a talk that I saw on YouTube. I know you’ve given it a couple of times, really fascinating talk about AI ethics that I encourage people to check out. How did you come to this topic?

Amy Hodler:
Well, it’s kind of interesting. I’ll say begrudgingly actually at the beginning. One of my colleagues, Lisa [Bigoni 00:04:08] suggested that we look into what grass could do to help with accountability and ethics when NIST of the US National Institute of Standards and Technology put out a request for information of basically what kind of standards and measures should we have for AI responsibility. And at first I didn’t quite see the connection of graphs and AI responsibility.

Amy Hodler:
And then as I started to look more into it, I realized that many of the failings that if you look up AI fails on Google, you will find many things that are related to not being flexible, for not knowing what your dad is trained on, for not having context to make decisions. And so it became, as I just did a little bit of research, it really opened my eyes as to the way companies that do work with graphs and how companies should use graphs and networks to really improve what they’re doing. So an unusual source that got me started here.

James Kotecki:
And let’s talk about some of your … Favorites not the right word, but what are some of your go-to AI fails that illustrate the kind of problems that you’re talking about?

Amy Hodler:
Yeah, there’s a couple that just always stick with you and one of them is related to recruiting and bias in recruiting systems that provide a quick pass of candidates that should be offered for jobs. So there’s several examples of this and companies always are it seems it’s always unintentional and they seem to react pretty quickly when they realize it’s there. But there was some examples of recruiting in tech companies that basically would be bias against women resumes. And what they found is that they were trying to base their learning of who would be a good candidate on current employees. And the current employees were on average, more likely to be men than women.

Amy Hodler:
And looking at how they describe themselves on LinkedIn was very different the way women described their accomplishments. So it turns out that women use different adjectives. We have different hobbies. We were in a different projects or groups in college. And so it was automatically discriminating against that unintentionally. So that’s one example that it’s really about by example. It was totally unintentional, but those things happen.

Amy Hodler:
Another one is the Gender Shades project does quite a bit on AI and facial recognition. And they looked at both genders and different shades of people’s skins. And at the extreme ends, if you were white male, they looked at three different AI facial recognition technologies. And if you’re a white male, it would be accurate those on average between 99 and a 100% of the time, which is just impressive to me. And if you were a black female, it was correct 65 to 79% of the time. I’m still impressed with something being 80% correct, but you can see how the inaccuracy may cause problems especially because some facial recognition has been used for policing.

Amy Hodler:
And even if it wasn’t policing, if it was just getting into buildings. So instead of your key card, I would hate to be late for meetings because 30% of the time I couldn’t get into the building without an extra pass or something like that. So those stick in my head. And then the third one that I really, really will never forget relates to COMPAS software, which is used to basically do recidivism rates when people get booked. So it’s the policing kind of software where they’re trying to estimate whether someone will have repeat offenses.

Amy Hodler:
There was one gentleman, Victor Rodriguez who had committed a crime, had been in prison for a while and basically had trouble with parole because he was being marked as potentially a high risk. But all of his 10 years of good behavior, everybody thought he should be let out, but the parole board was not comfortable because the software had predicted that he would be a high risk. And he wasn’t allowed to actually look at what the logic was behind that. So those three I think really, really just stick in my head quite a bit of examples of how I think we can do better.

James Kotecki:
Right. In the parole example, he wasn’t able to crack open the black box so to speak and look inside and see how the algorithm was making its decisions. And the same black box is almost the wrong phrase there, because sometimes we use the phrase black box to mean algorithms that we just as humans cannot understand because it’s just beyond our grasp. But in this case, it’s unclear to me whether that was that kind of black box or whether the company was just closing it off to outside observers. If you could have looked into it, you could have understood it, but they just didn’t let people do that.

Amy Hodler:
Yeah. It feels like it was a little more on the ladder, but because we don’t know. His lawyers were not able to look into that black box that we really don’t know if it was something that could have been understood. There’s some suspicion on a few questions that they believe gave him a higher ranking, a higher risk ranking, but they’re really not sure. And I think the reason why I like that example is because it does highlight just as you said, that there are things that are unknowable and we should be careful with those and careful in how we apply things that aren’t interpretable. But there are things that you can know, but for some reason you’re not allowed to know. And I think that’s very dangerous when we’re protecting commercial rights in very sensitive decision-making.

James Kotecki:
So we’re talking about AI faults and ethical challenges in hiring in facial recognition, which is related obviously, especially in people’s minds right now to policing. And now in the idea of parole, is there some kind of unified thing that went wrong in all of those cases? Was there something that a person should have known or should have done differently or really talking about a combination of a whole bunch of different factors in these cases?

Amy Hodler:
Well, I’m going to say yes to both those questions or both of those possibilities. So I think it was many things that go wrong, a combination of things. However, I think the overarching of why we allow all those many things to go wrong is just this lack of kind of a checklist or a lack of a framework, I will say for understanding is, am I building this AI responsibly?

Amy Hodler:
And there’s many, many things that can go into that, but any company could start, which is asking those questions like, how do I know this is responsibly built? How do I know the data was responsibly collected? How do I know the impact of if something does go terribly wrong? Have I looked into the actual different impacts when things go wrong, as opposed to just trying to say it’s 99% accurate? That’s good enough. So I think there’s many little things, but the overarching is just to develop a framework for responsible AI.

James Kotecki:
If those three known examples of AI bias are out there, what’s your sense of how many more subtle forms of bias are out there that we don’t know about because they didn’t make the news? And how big of a problem is this? Really?

Amy Hodler:
I think it’s probably quite large and I think subtle bias, subtle flaws, even if it’s not bias per se are more pervasive than we probably understand, and possibly bigger than the big examples. So when you have something really heinous that sticks out, it’s easy for people to get emotional and react and try to make a change of that particular item. But if you’re not dealing with the underlying problems, again, not having that framework, there’s all these minor things that, to me, equate to nudges. So you can nudge people or processes or decisions in different directions.

Amy Hodler:
So the COMPAS software example that one also the COMPAS risk score is also something that prosecutors are able to see ahead of time. A judge is able to see that ahead of time and in some cases not set bail or set bail at a higher level, depending on that score. And so this is a subtle nudge, but it doesn’t say they were put in prison because of this. And in some cases, it’s not that they were actually ever convicted, but because of a score could nudge someone to not be able to have bail, which might also keep them from being able to work, which might cause other problems.

Amy Hodler:
The facial recognition one, if there are companies using facial recognition for building admittance, and I mentioned, well, I’d hate to be late. Well, can you imagine if you were late for meetings? I don’t know, let’s say once a month and the perception was you’re a person who’s late for meetings and you don’t get a promotion. I mean, there’s all these little nudges I think that can happen that aren’t apparent on the surface.

James Kotecki:
And I guess if you were to put that in absolute terms too like you can imagine a situation where for one year a system is effectively discriminating against female candidates, and it’s very obvious and it’s found out and it stopped. And let’s say a 100 female candidates were explicitly discriminated against versus a system that’s in place for 10 years or longer running with more subtle biases in the background that people don’t catch. And maybe over that course of that time, maybe even more women actually ended up not getting jobs that they should’ve gotten, but it’s just not caught in that scene.

Amy Hodler:
Yeah. Yeah. And there’s one example that is not necessarily super subtle, but it has to do with how data’s collected and how you train your information. So you can train AI systems based on really good data, but if the data is not, or the data itself, the quality is good, but the data is not collected appropriately if it’s collected in certain timeframes. So a lot of the predictions of how much police and what resources you should put in an area is based on historical data. And the belief is some of the historical data is biased. And so there was bias reporting.

Amy Hodler:
And now you’re recording on what looks like on the surface good data, but the collection either wasn’t appropriate or it’s just historical. And so you’re training on historical data versus today’s data. There’s a really great example of collecting data on which I think it was strains of wheat should be grown in certain areas. And this was done in an economic area where women … In a cultural area, women did most of the farming, but it wasn’t appropriate for them to talk to men in public. And so their husbands were explaining what wheat they thought should be grown. The women hated that wheat variety, because it was harder to pick. And so in that case, yeah, so some interesting things like even how you collect the data can impact what your predictions are going to be.

James Kotecki:
We mentioned this earlier the idea of the black box. One proposed fix to this at least conceptually is the idea of interpretable AI, understandable AI. There’s different terms that people throw around about it. And people mean different things about it, right? Like someone might say, “You should be able to interpret it, but you can never fully understand it.” Different things like this. Where do you stand on that? I know you have a phrase in your talk that you don’t necessarily have to trade accuracy for interpretability. So let’s just dive into that question of interpretability a little bit more,

Amy Hodler:
Yeah. When people talk about explainability, they usually mean they can in general explain how a result came out from, and that’s usually they’ll look at a prediction and then they go backwards and say, “Well, it must have been based on these factors.” Interpretability means how the decisions were made and each step along the way for this predictive model. And that’s very difficult in a lot of in particular deep learning, but that can be very difficult even in machine learning, especially if you have a lot of data. But the belief is that in order to be very accurate, you can’t have a lot of interpretability. It’s just somehow impossible.

Amy Hodler:
And one, I don’t believe that. And there’s some excellent, excellent work done by Cynthia Rudin out of Duke University. She has a lab, she has code she can share. So if anybody’s interested in that, look up Cynthia Rudin Duke, you will find really some good shareable code. But what she does and I think we should do more of is look at interpretable models, statistical methods, some boring old decision trees. Try to look at making those more accurate. And a lot of times in her work, she has actually shown that she can build a model that’s just as accurate as these black box models using methods that are just, I don’t want to say, boring statistics and decision trees so that you can actually find each step of the way of the decision. And I think it’s important.

Amy Hodler:
You’re not always going to be high or accurate there. But if you are doing decisions on things that are really critical, you better be sure that you know why those decisions were made, so you’re not training on something arbitrary like accidental like a zip code. There was an example of training based on cancer visualizations trying to train a model to predict. And they were actually has had some good accuracy, but they turned out later that it was cuing of the model and the kind of a mobile model of the machinery. So not based on something that you would actually want it to train on. So yeah, I think there’s a lot being done on interpretability. And I think we’re going to see more of that as people start to apply AI in a more broad sense, that just having that is going to be important.

James Kotecki:
I mean, beyond the technological, though, there are certain baseline things that we as people need to do, right? So how much of this is a problem of using the right technology, using the right system, the right model, the right level of interpretability? And how much of it is just a fundamental, ethical grounding that we need to have as people? Because if it’s the latter, that makes me be me a little bit more pessimistic. Because, I mean, people have been struggling with ethical questions for thousands and thousands of years.

Amy Hodler:
Yeah. I think it’s both. I think that AI problems are often human problems and I don’t mean that as we just are unethical beings, but that we need to put in place frameworks, guidelines that allow us to be our better selves. What I really find actually gives me hope -- so hopefully it’ll make you feel better -- is when I talk to data scientists . . . When I talk about this topic to businesses, you get the head nods. When I talk to data scientists, they are so hungry for a framework that lets them do better. They usually got into the business -- coders and data scientists -- to do things that are powerful, that can help society. They don’t want to misuse this, but it’s such a powerful tool.

Amy Hodler:
They need the frameworks to both have the air cover to do the right things, but also just makes it easy for them to do the right things, to prioritize the right things. If you’re not prioritizing, understanding whether there’s bias in your data, tracking where your data come from. If that’s not given, if the data scientists aren’t given the time and the focus for those things, they’re not going to do them. It’s not going to happen as much. So I think there are human feelings, but there are things that we can do to do better. And the people with the fingers on the keyboards really want to do better.

James Kotecki:
And let’s zoom out further to the people that are in society that actually aren’t working in these businesses at all. They’re not technologists. They’re not the business people in AI. These are just people in society. How much of the demand for some of this needs to come from external, I guess, market forces, so to speak? And then how much of a responsibility does that give folks in our industry to educate the public very broadly about what’s going on?

Amy Hodler:

Well, I definitely think we need a public conversation about what our expectations are for AI. So it’s not just ‘I want more accurate AI.’ It’s do we want AI doing particular things? We need to bring together, I would say, people from different walks of life. So you’ve got your scientists that we all think about, but also your ethicists and your business leaders, your citizens, your policymakers, and have more of a joint dialogue about what we want from AI systems. They are different than software we’ve had in the past because of the reach and the power that they really have the ability to amplify the best and the worst of what we can do.

Amy Hodler:
And so I think it just needs to be a dialogue of what our expectations and what we want from it. And an example of that is if you look at the Chinese credit system where billions of people will have a social credit rating already have, and even more will have soon. Where you have a social credit score and the system works fairly well as designed, but it also is designed to give you a social credit score lower if you do something like jaywalk. And with facial recognition, they can see you jaywalk or smoking in a non-smoking area. That system is working as designed, but do we want it to do that? And so I think those are the conversations that everybody needs to have is what do we really want and what are our expectations there?

James Kotecki:
And you talk to people about this a lot. Do you think that beyond the data scientists, beyond even business leaders, do you think society is ready for this kind of conversation? Do you think in the next few years we might see, I guess Andrew Yang was maybe the one politician who was really talking about AI on the campaign trail this year. But do you see in the coming years a larger societal push for regulation and policy change in this area beyond just what individual business leaders might decide to do?

Amy Hodler:
Definitely. So I think first of all I want to say, I think journalists play a crucial role in this and being the interpreters for some of the tech speak to what it has real impact. So I think that’s going to be important. And I do think you’re going to see people join the conversation and we can kind of see that already in some of the calls for AI regulations that are already coming. So you have the EU ethics guidelines that are out. They came out early this year.

Amy Hodler:
You have corporations as well like Amazon and Microsoft calling for rules on facial recognition. You have NIST, which is the government organization looking at what should AI standards be? How do we even develop AI standards? You have researchers that are also an analyst that are also trying to promote this. I think this is already starting to happen. It’s just not necessarily front and center yet and I think that’s just going to increase over time.

James Kotecki:
Do we have positive examples? Maybe they’re not as readily available just because of the nature of how this works and is portrayed. But do we have positive examples of non-biased AI, maybe doing a better job of helping people to overcome their ethical challenges versus what we were talking about at the beginning of this conversation where a lot of ethical problems and bias challenges and flaws with AI systems? Do we have examples of that on the other side?

Amy Hodler:
Well, I think we have AI systems and machine learning in general, examples of how to improve what’s already there. And so there are projects like the Algorithmic Justice League that helps data scientists actually look at what they have and make improvements. There are things like the AI Fairness 360 Toolkit, which can help you de-bias your data. So it helps you find data and then actually take bias out of them. And then as far as examples like real world, like hands-on examples, there’s also been AI that originally was biasing results of looking for breast cancer.

Amy Hodler:
And it turns out women of different shades have different commonalities in the MRIs and the visuals that you use to take a look at that. And so they’ve actually used AI to bias it depending on the ethnic background so they can be more sensitive to certain people of different backgrounds and pick up breast cancer easier. So bias sounds like a bad thing, but sometimes you actually want to include bias for certain reasons because different groups have different sensitivities or different markers for, in this case, health issues.

James Kotecki:
Yeah, I think this is an interesting point that maybe semantically would be harder to convey too in a headline, in a newspaper, right? Biased AI in a newspaper headline is always bad. But I think what we’re saying is that there’s always some kind of bias. Either that bias is not being acknowledged or we’re not aware of what it might be, or we’re able to kind of deliberately get a handle on it and then shift that bias in a direction that benefits us or does what we want it to do from a policy or societal perspective.

Amy Hodler:
That’s a great point because and part of putting in that framework and having that conversation of what are the outcomes we want to see? We can then say, “Well, there’s all these levers and knobs you can turn with training models and how you implement them.” And we can either go straight on what the existing information is in general on average or we can say, “Hey, for this group, we want to turn this lever. For this outcome, may be we want to increase diversity in our workforce.”

Amy Hodler:
So we actually turn the AI lever to be a little biased towards candidates of more diversity. So there are things you can do to guide AI to better outcomes. And that’s a great point is that if we can have a conversation about the outcomes, then we can guide models to where we need them to be.

James Kotecki:
And then it will come down to what we think is fair, right? AI fairness will ultimately come down to what we as people and as society think is fair. Which of course is extremely challenging. Because as I mentioned with Ethical, we’ve been debating these things for a long time and there’s different political and social and philosophical perspectives on that. But this is a tool that helps you either achieve better things or kind of further worse things.

Amy Hodler:
Mm-hmm (affirmative). Yeah. So yes, I think the having that conversation, you can have that conversation at a very high level and then do your implementation at the model or the business or the organizational level. One of the things I do try to tell people if there’s nothing else. So if you guys want the biggest practical tip here if there’s nothing else you do, know and track your data. That is a starting point for any machine learning and any AI is the data itself.

Amy Hodler:
That is one of the reasons why I’m pretty bullish on graphs being able to help, because it is very good for data lineage. Where was it? Where was the data tracked? What’s in your data? When was it changed? How was it collected? Do you have duplicate data that you’re double-dipping on so you’re amplifying the results? So there’s things like that you can do, that is the core foundation to getting all of the other questions to understanding what your system is doing is just knowing and tracking your data, so step one.

James Kotecki:
Well, bringing it back to graphs like a true interview professional, just where we started in this conversation. So Amy Hodler, Neo4j I really appreciate you being here on Machine Meets World today.

Amy Hodler:
Okay. Thank you, James. It was a pleasure.

James Kotecki:
And thank you very much. Yes, you, the person watching, thank you for watching and or listening to this conversation. Whether you’re listening or watching, you can do the opposite. We have this as a podcast. We have this as a video. Infinia ML produces Machine Meets World and you can email the show MMW, Machine Meets World mmw@infiniaml.com. Give me guests suggestions, feedback, comments, whatever you want. Thank you so much for joining us today. I am James Kotecki and that is what happens when Machine Meets World.

Share this post