Causality 101 with Robert Ness - #342

[Music] welcome to the 12

castes I'm your host Sam Cherrington hey everyone I am on the line with Robert Isuzu ness Robert is a machine learning

research engineer at gamilon and an instructor at Northeastern University Robert and I met at the last NURBS conference where he had an accepted

poster session around his paper integrating Markov processes with structural causal modeling enables counterfactual inference and complex

systems which he also presented at the black and AI workshop this kicked off a bunch of great conversations between the two of us leading ultimately to

collaboration that we'll talk a little bit about in this conversation Robert thanks so much for joining me on the tour mo AI podcast thanks for having me Sam your introduction makes me think I

should have came up with more clever name for that paper you know what a lot of the papers we talk about on this show are quite the mouthful so yours is no exception

maybe someone will build a model you know that seeks to determine a inverse correlation or correlation between the length enos of a potato and its number

of citations or something like that but let's set that aside for now and have you spend a few minutes introducing yourself how did you get started in

machine learning what picture interests you know ultimately we'll be spending a lot of time here talking about causality how did you come to become interested in that you know my path to machine

learning was a bit I'd say unconventional I started off working in in Asia and China

specifically I was did a degree at Hopkins and International Studies and was planning to pursue a degree in economic in economics focusing on

economic development I got involved with some Internet companies out and Bay in that's that got me into coding and

databases and data in general and I decided I was interested in and that and that and went to so I applied to four

programs in statistics particularly with a focus on computational statistics I went back to the States came back to the States went to Purdue

University to do my PhD in stats my PhD work was on causal inference graphical models basically how to learn causal

models from data particularly in a context of systems biology and from then after I graduated I went straight into

industry got it now we hear very frequently folks refer to their path into machine learning as unconventional or indirect in your case you came into

an interest in gaming and that lets you to apply for or I'll go into grad school for statistic what was that particular connection really it's when you're on the back end of an app and you're

looking at the data and you're realizing that there's a lot of insights to be had if only we could model this data and turn it into some service on the front

end mm-hmm I realized I mean this was you know people had we're just kind of starting to talk about data science and and and you know how very and had just recently came out and said that

statistics is the new sexiest well I came here what the exact quote was and take your metaphor yeah pick a metaphor orange is the new black statistics is

the new I don't know rock star and so and yeah that's that's kind of why I pivoted to to the two stats in machine

learning yes I guess through stats I mean if you people might argue whether or not stats and machine learning are the same thing might the problems that were I was working on my PhD we're using

probabilistic graphical models so you know which has strong roots in artificial intelligence so that was my introduction there is she learning yeah one of the things that's come up in our

car sessions about causality and the work that you're doing with your courses is

the idea that historically talking about causality has been the you know the domain of statisticians and you know

folks like economists and that a lot of that conversation is inaccessible or isn't really tailored to the needs of developers and data scientists machine

learning engineers I didn't realize all the time we were talking about that that your background was in economics you have some of the exposure to the way

that causality has been traditionally kind of used and talked about you know maybe I guess I'll just use this as a segue to kind of opening up the floor to

ask you you know what how do you define causality the interesting thing about

causality and maybe part of why maybe it is a challenging thing to deal with particularly for statisticians I would say is that it's very difficult to talk

about it without finding yourself having a philosophical conversation and you know so going you know and this is something that Phil that you know what is the causality you know these are

something that philosophers have been wrestling with through the ages right Hume has has a counterfactual definition of causality that's you know a follows

from B and had a not happened B would not have happened but you know philosophers going back to the Buddha all kind of take their stab at what is

causality and so there's a there's different philosophical arguments for causality and what it means I think from a practical standpoint what most people

mean when they say Staz causal inference is they mean the estimation of causal effects so if you're say for example at a tech company

and you want to run some kind of experiment about the about whether a feature will drive a click or some other

key performance indicator or metric your you're asking you your experiment is essentially trying to get at the question of what is the causal effect of

this feature on this outcome and you'll be using assumptions and methods from statistics to estimate assuming your

assumptions are valid those causal effects but when we talk in machine learning we're now hearing you know so I had nerves like you said

yoshua bengio gave this talk about having agents that can understand the causal structure of the world and and that causality is essential from moving

from system one system to cognition as you daya pearl was a very preeminent causal inference researcher talks about causal reasoning and in terms of free

will and you know the ability to understand intention and and so there are definitely definitely a lot of angles to tackle this question from the

perspective arbiter of artificial intelligence that you know people who are running experiments in Facebook and Netflix are not really thinking about

yeah it's interesting that you you kind of you know where you started your response and ended your response when I

talk to people about causality and agreed you know at nerves and even before you know causality over the past

year has really been one of those topics I think maybe peaking at nerve so it was definitely a hot topic at North's I don't think I think it's too early to say that it's peaking and I don't think

that that's the case but it was definitely a hot topic at Naropa yeah but even before there was kind of this growing crescendo of interest and enthusiasm around this idea of causality

when I talk to folks about it the conversations tend to be you know either the the philosophical conversations kind

of very abstract conversations that you mentioned that you started your explanation with or you know talking about it like it's you know another

Algar or tool or approach on the on the shelf and that hey you know we've been doing

machine learning and now we you know you know we realize we just need to pick up this you know this causality tool and kind of sprinkle it on what we've been doing all along and you know we've got a

better approach to doing things do you get that sense as well in the industry and what do you think's driving that yeah and I I sense a lot of cultural

clash in terms of trying to articulate the problems in one field and what the ideal solutions are coming from one camp and you know what people are already doing who are working on causal

inference and what they think the smart ways of what the next steps are in that line of research I think you know from the algorithmic standpoint clearly yeah that's where we want to go we want to be

able to algorithmic eyes causal reasoning right now and there are ways of doing that and they're connected of course so for example if we're talking

about like what I said the tradition no question of of causal inferences you know can we estimate a causal effect if the causal effect is again the effect of

something that you care about like some kind of utility you're not really that far away from say a Markov decision process or reinforcement learning especially if for example you're you know the thing that you're the

experiment that you're running you're running in sequence so so that's you know you you run the experiment you get some kind of results you you update some kind of belief structure about your

model and then running it run a new experiment that's getting very very close to the language of reinforcement learning but if you're in if you're working reinforcement learning and you try and dive into that literature you're

going to see a bunch of papers on Public Health experimental designs and you know and really kind of very domain-specific assumptions built up around linear

models none of the stuff that you're working on if you're a typical reinforcer learning or a researcher another issue would be for example if

you are working on trying to if you're working on deep learning for example and you're thinking okay well how can I just apply

deep learning to causal inference well one of the chief problems in causal inference is the question of whether or not something is identifiable right so this is a say that I wanna I have some I

have a model and I have a causal question that I want to ask my model and I have some data know given my data is this question even is it even possible

for me to answer it and the answer might be no this is something that I found that people with a background in machine learning particularly a deep background image and deep learning have trouble

thinking about because they're they tend to think that if I get the right architecture and that gets enough data and I label the data in the correct way

and I get the right loss function then I can solve this and so that that's a huge class right there right that's a it's a fundamental class and so you know

getting a lot getting over some of those I consider them cultural barriers is is I think what we need to be striving for

as a community the example that you brought up of you know you're an engineer working at a you know ad tech

startup or you know any kind of web startup and you want to estimate the probability that the change that you're

making you know has a result that you're predicting you kind of use that broadly as an example of where you might apply causality but we've been trying to

estimate these kinds of things for a long time what's different about the causal approach to this well and if you had to define what if what a causal

model can do that a non causal model can't do at a most basic level it's that it can predict the effect of an intervention and so what an intervention

is is some kind of action that's that you take or a modeler takes or people who are making decisions based on a model on a model are taking that affects the data generating process and so

if you are affecting the dealer generating process then your energy in your changing it well that means that's the data that you're now getting from that new process is different from the

data that you trained your model on now at the end of causal model your causal model can account for that change and its predictions will still be valid but as we know with even some of the more

powerful models in machine learning if we when if they're just pure predictive models they don't have a causal structure to them then the second you change the data draining process such

that it differs from the training data you're going to have it you're going to have issues and it's not going to be very robust to new data that your D'Amato hasn't seen before and so that's that's kind of when I what I try and

pitch people on like why should I care about causal inference that's really the main thing is that so in the example I think they were referring to I it's a simple example but I say suppose that

you were you know you had some kind of sensors that were collecting data about the climate the weather in the local the microclimate where you live in the

morning and using that and then in the evening you put you you record whether or not it rained okay and so after a while you click you collect enough data and then you create a predictive model

that predicts whether or not it'll rain and then based on this model you make a decision based on this prediction you make a decision and that decision is whether or not to carry an umbrella with

you in the morning now that's fine because what happens there is carrying the umbrella is going to have no effect on the weather but we can look at a comparative example say for example in

business where you look at some business indicators and then you you and and then at the end of whatever period it is a quarter you know a cycle you record you

know the revenue for your company and then you do that for a while and then you train a model that predicts future revenue given these businesses indicators and then if your future revenue looks like it's going to be low

you run some advertising well that advertising is going to now start generating data that is a did buy that advertising and so when you retrain your

model on that new data this kind of feedback loop that it where where your actions are affecting the data generating process and feeding back

into the data this this is an issue for example in fraud detection so that's you know what happens once you start you're able to detect fraud then you start rejecting the fraudulent transactions

and now all the only fraudulent transactions that make it into your future trading data are these edge cases of fraudulent transactions and so when you retrain our motto is now has a strange bias towards those edge cases

and so that can confuse things and so yeah at a very basic very basic use case for causal modeling in a in machine learning is to be able to predict the

outcomes of interventions and so and once we can bridge interventions we can do a lot of clever things on top of that so we can for example we could have models or we can have modules of our

models that are more robust and can transfer across domains so that we don't have to retrain them once we get to a new data set or a new domain and considering the cost of training a model these days that can be that could lead

to some significant cost savings well can we can we can estimate causal effects which it's very relevant and feels like reinforcement learning as I

said and you know and then there's the question of counterfactual reasoning so just to recap that before we move on what you're saying is that the

traditional machine learning models that are kind of sitting on the shelves for us to use you know tend to make this assumption this IID assumption that the you know the data that it's trained on

are all independently and identically distributed right and in many use cases the decisions that we're making based on the models that we're using actually

changes the distribution and the independence of the data in future time steps and it sounds like we just live with

that application but it causes problems for us and one of the reasons why people are starting to get excited about causality is because it holds some

promise for fixing this and you know making our application of machine learning more kind of applicable in the real world that we live in yeah you know so speaking of real world

applications people would be interested I think you've had speakers like Tim knit on who are talking about algorithmic fairness and so causal inference would have a lot to say about

that for example right so like is hisae policies effect on some kind of outcome like say an adverse outcome for a specific group of people is it a direct

cause of that policy is it's kind of being mediated through indirect causes or is it just correlated to that to that initial cause because of some kind of

confounding factors in the in the outside environment that generated the data you know so that is one reason why when we were worried about like you know

algorithmic fairness and an ethical way I clearly you know cause what inference has a lot to say about that the example that I just gave of course and just making sure that our models if we're

making decisions that affect the trading data that goes into our models and making sure that's is somehow accounted for and you know those are those are some very basic applications of it

although I think that when when machine learning people at machine learning conferences get excited about causality I don't think it's so much because of

estimating causal effects or ethical AI or explainable AI I think it really has to do with things like modeling like disentangle representations and and in

variational auto-encoders for example like if for these if we're trying to parse apart these latent variables that are that have some kind of meaning and transfer across data sets can we bring

can we bring some causal reasoning to bear on that problem I think people are interested in people in a probabilistic programming community like myself are

very interested in the ability to build causal models as programs and use those to reason counterfactually that that's the kind of cure that's the kind of

bread-and-butter AI capla keishon of causal inference although these other these other that I mentioned are clearly very relevant you mentioned disentangle

representation and variational auto-encoders when I think of that the kind of analogy that comes to mind for me is like embeddings and you know a

transformation from you know the feature domain to some embedding space and typically that's a kind of opaque representation is the idea that with

causality that embedding space or representation space is more meaningful causal understandable explainable or is it something different than that yeah so

generally the idea is that this embedding space is somehow capturing latent causes and so there's some very there's a very important question about

ok statistically is it even possible to identify latent causes even in a multivariate setting which some people think is possible like there's a lot of you know you can show mathematically

that you get a lot of problems if you try doing that so people are worrying or thinking about that and they're thinking about you know there are when you have a

a a probability distribution say of a of a effect given a cause a conditional probability distribution of an effect

given a cause is driven by some natural causal mechanism that drives that effect from the cause right and we assume that that's causal mechanism since it's a

part of nature is going to be invariant across datasets right so if we can look at that type of intuition and you know what it means for these mechanisms or

what what the nature of these mechanisms and how they might for example be independent of one another what that might say about the distributions that they entail you know can we now build

those as assumptions inside of a say a deep model like this right build it into the first say for example the the the loss function or the structure of the

model and come up with latent representations that we can have stronger believe that these are actually good representations of you know talking about causality we've

talked about some of the you know the kind of philosophical origins of just defining what causality means and its popularity within the research community

how ready are the tools around causality that they can be practically used by you know folks in industry are we there yet or is there are we not quite there yet

there are some excellent packages both in R and in Python that will do useful things when it comes to causal effect

estimation so you know it's I there's a fellow at Microsoft Research named Amit Sharma who has a library called do Y which is I think when it comes to just

that problem of estimating causal effects I think is has it has one of the best interfaces that I've seen for that problem it's very intuitive because you lots of options and it's separate and it

separates the problem of identification and estimation very well I am somebody who comes from the deep probabilistic programming community I like to think

about a data generating function as a causal process and I like to think about

how I can model that in in in a program

and and use tools like tensor flow or PI torch to capture the complexities and

nonlinearities in that system so so for example with the workshops and the courses that I teach I use pyro which is

a generative modeling framework from uber AI labs that extends pi torch now that said some of the challenges they're

in there some challenges there when it comes to inference on these models it can be I think that's an active area of research and so if you're not you know

you kind of have to turn away at the inference problem when you're working on these models but it's certainly doable well maybe

fine or restate the question to you know where our people successfully using this kind of kind of body of knowledge around

causality in machine learning not in stats or economics but in machine learning you've mentioned some examples of representative problems but are these

problems that are you know solvable today that people are you know using causal approaches in or are these things that we would like to apply causality to

once the tools are more mature I would you know the best applications within the AI spaces that I've seen and I very much hope if your listeners take issue of what I'm about to say that they email

me and prove me wrong but you know most of the the the cutting-edge stuff is happening within a risk research

community within industry there are I would say that's I've seen people in industry working on say an analysis by

synthesis approaches which is where say you have some kind of latent grammar of abstractions within a domain that you

use to generate say images or synthetic environments for say for example training self-driving cars for example and when they designed these grammars

they had a kind of causal semantics to them I've seen some early success with those types of approaches the biggest application industry that I've seen be successful that's kind of not surprising

is when it comes to say actual machine learning models in production would be people who are applying causal inference

techniques in experiment and experimentation to the optimizing of hyper parameters for machine learning models so it's just running experiments

on models to make them to tune their parameters so I hinted at this earlier you've got a course that you are teaching at Northeastern University as

well as through your own upstart education company all deep AI and I'll have you talk a little bit about and we're launching a collaboration

around bringing a study group that you'll be leading for this course to the twimble community and anyone who's

listening to this who's interested in learning more can join us for an overview session that we'll be doing on February 1st but so let's have you share

a little bit about your thinking around this course how it's different from other educational resources around

causality and in particular who it's who it's for given kind of where you think the you know the market is and where

causality is from a utility perspective okay yeah so it's deep school of a IOT PAI and it helps people break into new

opportunities surrounding cutting-edge AI and so it assumes that you already have a background in data science and machine learning and that you can code although not your preliminary

backgrounds fine you know I started it because as a as a research engineer working in the field I was seeing that you never have a case where somebody would identify some kind of pivot they

wouldn't they want to make like say for example I'm a data scientist and I want to transition to an engineering role where I'm working on machine learning algorithms or that I'm a software engineer and I want to take a lead on

building our company's new experimentation platform or even kind of off-the-wall things like I want to implement and deploy an agents that plays online poker right and so to get

these types of goals to really achieve that me you need to combine learning about new research in an area an AI with actually finding tools and libraries and

tutorials that help you do the implementation and you know so the process of going through a bunch of research papers and medium blog posts and Jubran notebooks it's slow it's

arduous it's there's a lot of hype and noise that you have to churn through in order to get a little bit of signal so all deep solving this by bundling the science and the implementation and

presenting a way that delivers the transformation that people are looking for by the end of the workshop so the first sequence of

courses are having to do with causal modeling and machine learning and so if you've been to you know if you'd like you were like you were saying before you

know we've gone to these conferences we know that's the gets to common-sense reasoning and AI we need to look at causal inference but like I said if you

look at the causal inference literature you're gonna you're gonna find weird toy problems language that's focused on econometrics and Public Health you know

a lot of talk about problems specific variations of linear models and stats that's are probably a lot different from the kinds of problems that you're working on it's very hard to penetrate

so the course does what the course does is explain or the course is rather they they explain at a high level in the language and in the context of machine

learning what these well what what causal inference is about what causal modeling is about what counterfactual inference is all about and how to Al

Gore ties it and if you're like say for example you're familiar with things like topic models other latent variable models Bayesian nonparametric models and

you're using you're comfortable working in a deep probabilistic programming language like pyro it's it's I think it's a great choice are those things

prerequisites because I imagine there are a lot of folks out there that are interested in causality want to be able to engage in conversations about it or understand the conversations that are

happening about it but don't know anything about probabilistic programming or any of the other things that you mentioned I would say that don't or not prerequisites if you have a background

there you're going to be able to take off right away putting probability programming aside it's it's more focused on the idea of

model-based machine learning which is that which is saying that to tackle machine learning problem you should be thinking very very precisely about your

data generating says and then write a create a bespoke model tailored to that process and you know and since I would argue that for a

human being it's impossible to think of a process that generated the data without you thinking about it in causal terms it really just that way of

thinking models sorry Maps directly to the causal modeling and so probabilistic programming is just as a really way really easy way of going about that if

you don't have experience probabilistic programming so that so none of the examples are the types of large deep models that take all data train and are

very difficult to debug they're all very easy to sink your teeth into while at the same time and they're all very kind of contextualized within the kinds of machine learning problems that's somebody with a mission learning

background what might be used to but at the same time they're there it's clear how they can be generalized to more deeper and more nuanced more complex

models you made a point in there about the causality I believe or I don't recall if you were precisely referring to causality or probability

probabilistic programming but the idea that in thinking about a model you're

thinking about the the way that the data is generated can you elaborate on that point a little bit yeah so it's it's a

philosophy that is commonly articulated as model-based machine learning and so yeah the typical machine learning playbook is that you you get the data sets you look at the structure the

features you look at the structure the race of the target response variable you start transforming things to fit whatever modeling framework you are familiar with or whatever happens to be

fashionable at the time maybe you go on to you go and pick the most fashionable model and you kind of switch out the eminence tailor for whatever your data

is and just assume it's going to work and then you kind of iterate on it you tweak things until you can ramp up predictive accuracy this the model-based machine learning

approach is is saying that it's taking a different philosophy for building a model you're saying that rather than being a hammer in search of a nail you should say all right let me let me think

about the process that generated this data let me build a bespoke model with respect to this data and and and try it

and then let me try to you know evaluate and critique this model in terms of how well how faithful I think it is to the original data generating process of course you can still have a loss

function that optimizes for prediction now this relates to causality because if you're thinking about a process that generated data typically the way you

think about it is as a causal process so first this happened and that that happened and then maybe some noise was added and then it was translated all of these things are you're thinking about a

kind of causal narrative for from a kind of causal origin story for your data and so so in this course we build on that model-based machine learning intuition

that generative modeling intuition what is it that excites you about causality is it the you know the use cases is it

about the kind of the elegance of the proofs like what is it that really gets you going about this topic you know I think for me the cool thing is to be

able to write really just write programs correct explicitly in terms of like a program what you think the actual processes that

generate the data and we have a lot of cool background research from all kinds of different domains that do that so we have say for example simulations in like climatology or multi-agent systems we

have you know differential equations you know stochastic differential equations or we have stochastic processes that we use to kind of describe how something came to be right

and so it turns out that we can actually implement these as causal models and then reason about them counterfactually so what a counterfactual means is that you know you know specifically a kind of

counterfactually called twin road counterfactual we were thinking about parallel universes you're saying that like I did this thing and here's where I am so I ate this hamburger and now I have a stomachache

had I not eaten this hamburger I would not I would not have a stomachache and that's something that's been very challenging for machine learning because that's that's example where that data

points where you did not eat this the the hamburger and didn't get the stomachache that doesn't exist in the training data right so you need some kind of way of simulating the alternate

reality for doing that and so for me personally just kind of writing algorithms that that that that does that that's super fun so like I you know I

was I was saying that you know it's very normal for a human being to say like oh man like had I not data that's woman in

college I'd be a much happier man today that's that kind of reasoning is really easy for us think about how cool it is that you can actually write a an

algorithm that can reason the same way right now now I'll be honest with you I don't I've seen some some I've seen a few actual practical examples of this

like in terms of with in machine learning I've seen something things like raising about occlusion in an image you know some of the things that we see like

with deep fakes for example you can you can actually consider that a kind of counterfactual problem like what would this video look like if it were Obama talking instead of somebody else hey you know those are those are fun examples I

don't know if they if they're really kind of the killer app I think we're still looking for that yeah it was really fun to play with I was just getting that I think the you know what

I'm hearing you say is that you know with regards to the the field in general your course you're not necessarily promoting this as something that someone is going to go home and apply in their

you know day job but rather it's going to position them to start to reason about the application of machine

learning in some different ways that are starting to get I would say that's to be clear the stuff that you need to get a

job on the causal inference team at Netflix or Facebook is definitely in there that's definitely covered deeply in this course and you need to understand that stuff before you can move into some of

the more cooler AI applications it's just for me it's to the those cooler AI applications are kind of what drives my

motivation but you know I I definitely I definitely have respect for all the people who are interested in running good experiments yeah yeah there's still a little bit of like a disconnect

between you know if you're if there is a causal inference job to be had at Netflix or Facebook there is some practical application or else those teams wouldn't I guess right yeah yeah

and those and those teams are like I said those are typically experimentation teams kind of working on you know serving ads and trying to price correct ask correctly for an oxygen trying to

figure out what's trailer to show you when you log into Netflix that's not it's and and and it gets kind of closer to AI when they're thinking about what

they call time burying treatments what you might what machine learning person might call mark off the decision process just trying to think of if I do these things in sequence and have been having

kind of run automatically how do I do that yeah but I think there's cooler stuff than that so for those that are

listening to this and aren't aware of what it means when I say a study group one of the things we do as part of our broader twimble

meetup and community is go through online courses together we've done that for a ton of the fast AI courses as well

as courses like the Stanford CS 224 NLP course the and ruins deep learning that AI course and we're excited to kick off

a collaboration with Robert around his causal modeling and machine learning sequence and this is particularly special because this is the first time we have an

instructor of a course leading the study group with us for more information about the study group you can visit to Malaya

comm slash community and join the community join the slack channel for the course and we'll post information there about the information session that again

will hold on Saturday February 1st at 8 a.m.

Robert any party thoughts or words on we didn't get to your paper no we'll have to we'll have to talk about that at some other time but I think you know this you

know this is certainly not the the first conversation that we'll need to have on the show about causality because again it is coming up over and over and over again and the few people that I've

mentioned as I've started talking about the this collaboration you know I've gotten a lot of enthusiasm because it's one of these things that a lot of people are talking about but you know very few

people really understand what it is and and how to apply it so I'm definitely looking forward to getting this kicked off as our you know many other people that I've talked to

but again sorry any parting thoughts from your end yeah and I'd say to all of the research engineers of the research

scientists data scientists out there who are maybe feeling a bit stuck on the on trying to say pursue you know certain career goals and and and and hearing

about all of that of the hype and the trends that if you really go deep on the problems that you think are cool you really can't go wrong I love this causal

modeling stuff I love how it relates to symbolic AI I love how I love the idea of creating an algorithm that's can mimic how I regret dating the woman I

did I dated in college I love all that stuff I love I love all that stuff about this field and you know even before it was sexy so I think um you know if

you're interested in if you're interested in talking about it more I I'd love to love for you to reach out to me awesome well Robert thanks so much for taking the time to chat with us and

by all means reach out to Robert RI about the course or Robert about his work and hopefully we'll catch you in the study group thanks Robert thanks Sam

alright everyone that's our show for today for more information on today's show visit 1200 Icom slash shows as always thanks so much for listening and

catch you next time

[Music]