Last month, our CEO Ahmad Wani joined three other panelists for the webinar “Artificial Intelligence for the Impact Economy” — a Q&A discussion centered around the application of artificial intelligence for social and ecological impact. Joining leaders from Descartes Labs, BreezoMeter and aWhere, Ahmad shared his expertise on applications of AI, including a bit of what One Concern has been up to in the past few months.
The following story is a condensed recap of Ahmad Wani’s speaking portion of the panel discussion. To access a full recording of the event, please visit Impact Entrepreneur’s website.
Laurie Lane-Zucker opened the event and introduced the moderator, Gideon Rosenblatt, a former Microsoft leader and consultant for mission-driven technology companies, who now writes on the intersection of human and machine intelligence. After his own introduction, he asked each panelist to share about themselves and the work that they do.
Gideon: Ahmad, would you tell us about what’s happening at One Concern? I’ve heard a little bit about it, and it sounds pretty cool.
Ahmad: One Concern is focused on building global resilience. What that really means is helping systems truly understand what their risk is, helping them avoid it during an event, or helping them recover faster after the event. We started as a research project out of Stanford, and in the initial days, we were focused primarily on earthquakes. And then we started adding other natural disasters, truly with the aim of trying to make disasters less disastrous.
One of our projects, which I thought might be interesting to share in this situation in particular, is our deployment in one of the western prefectures in Japan called Kumamoto. In a few months, in between Summer and Autumn, they will be at the peak of their cyclone season. They have a significant senior citizen population rate, and if they have to evacuate during a cyclone, we might be potentially looking at a spike in COVID-19 cases — potentially by several orders of magnitude.
This is a similar situation to several cities in the US. So we have been interested in calculating the risk of COVID-19 spread during an evacuation, first by estimating how many people would potentially be at risk of evacuation based on data from previous US hurricanes. Then we started to layer in our pandemic models, to understand what kind of shelters we should actually prioritize, what sort of mitigative actions we should take in those shelters to curb an outbreak, and what sort of people should be taken to which shelters. We were really running this as a data-driven, machine learning exercise, to see how cities can avoid a spike — that they evacuate appropriately.
You can check out our hurricane evacuation during a pandemic analysis here.
Gideon: At this point, what I’d like to do is kind of get into the heart of the conversation: to look at some of the realities of using these technologies to effect change — to create impact — around societal and ecological challenges that we’re facing.
I want this to be a little more free form, but this could be about some of the challenges you’ve run into in making this real — this could be in acquiring data, it could be in hiring, it could be in raising money, whatever it is — but I think it’d be great to just dive in.
Ahmad: One of the big challenges we, as a company, have faced is going from small pilots or small research lab implementations to country-level, or national scale or global scale implementations.
This is because there are several layers of uncertainty — both stemming from data and modeling — which come into the picture, when you’re trying to do machine learning or deep learning. And being able to understand where that uncertainty is coming from, and then being able to connect that uncertainty, and its impact, to the decision which is being made, is really hard. And sometimes, very critical for users to actually understand on the ground.
For example, If there is an earthquake right now in the San Francisco Bay Area, much of the damage would probably be caused because the Hetch Hetchy water aqueduct — a 170 mile long water pipe which supplies most of the drinking water to the Bay Area — could get disrupted, because it is crossed by multiple faults. And in that situation, you’re not just predicting the impact to a hospital, but also the ripple effects caused by a water pipe going down, or the substation 100 miles away going down — all these standard externalities.
Now think about this at a national scale: do you have all the data for all the water pipes and the gas lines and the sewage systems in the world? No. And once you start filling those gaps, using machine learning, there are uncertainties caused, because you have data gaps.
So how does that actually help a San Mateo hospital administrator make a decision? It’s a complicated process, to think: “I’m going to run out of water in the next 8 hours, so where should I evacuate, where should I set up my temporary medical shelters?” So it really requires understanding all the way from where the data gaps are, to understanding the business process, and how those uncertainties affect that.
And this problem gets magnified when you’re dealing with public safety — when you’re dealing with really time-sensitive decisions when lives are at stake. And so I think as AI makes its way into more of the day-to-day, and into other industries — these are going to be some of the challenges.
Gideon: That’s really interesting. You know, when I was working with the nonprofit sector a couple of years back, one of the toughest things that we ever tried to crack was trying to help them use technology to better understand their impact — impact assessment, evaluations, things like that. And I can’t help but feel that these tools will eventually become very powerful, and very important, in the impact economy — for doing just that.
Any evaluations on the insight side of things?
Ahmad: I think risk disclosure is an interesting example. Risk disclosure, especially for non-financial risks — whether it’s the climate, or whether it’s a pandemic — is inevitably now going to be more important than ever. Investors, and now governments, are going to be trying to figure out where they are in terms of being prepared for the next pandemic, or sea level rise, or the next hurricane. So being able to actually truly come up with a metric or a standard which can be disclosed on the financial books of companies, is going to be critical.
Now, investors like BlackRock have been publicly releasing a lot of information on this, and now with this outbreak — sadly, this is kind of going to be a forcing function, in our opinion, to force that conversation.
And technology is required, because every company is global, and supply chains are global. Understanding what is going on in Thailand is important for a chip maker in South Korea who ultimately provides some subcomponent for the iPhone, for example. So being able to really understand all these dependencies at a global scale is going to be very important.
Gideon: Thank you. I want to shift the focus a little bit to the expertise challenge: how do companies go about bringing in the people that know how to do this work? As many of you here on the panel no doubt know, hiring these folks — people with data science backgrounds — is very challenging, especially when you’re competing with the Googles and the Facebooks of the world, who can throw lots of money at a lot of kids, or PhD students, fresh out of school.
So why don’t I just turn it to the panel: what are you doing in terms of trying to attract talent with this kind of background?
Ahmad: We’ve been able to actually successfully partner with a lot of US universities, and a few European universities as well. And we’re working to help them commercialize their technology, and really bring that technology in the hands of people who can actually make decisions. And that’s a driving factor for many of those researchers and professors to engage with us.
And at the same time, that enables the students who are getting their PhDs or postdocs to get an idea of what it’s like to work at a mission-driven company. And again, we’re not looking for people who are career-focused; we’re looking for people who are mission-focused. And that is where we would like to be looking for talent.
From here, the discussion turned to an audience Q&A.
Q: “There’s been lots of great examples about using advanced analytics, but not necessarily about machine learning or wider artificial intelligence. Could the panelists speak a bit more about the specific considerations of applying machine learning for impact? Is it more challenging gathering data for the machine learning training period in impact sectors, since data may be less available and lower quality? What is the specific added value of machine learning, as opposed to simply good analytics?”
Ahmad: I think the heart of the question really is: “is machine learning always necessary for every problem?” And the answer is no.
Some of these things could just be pure logistic regression, and you essentially make a quick model or linear regression. And those are sort of simplified versions of machine learning techniques, but it’s really important to understand what kind of problem you’re trying to solve — and really, what decision the user is going to be making, or what you’re expected to deliver — and then decide.
For example, I described our project in Japan — we want to be able to predict: how many inches of water are you going to have outside your house tomorrow at 4pm, because you just saw a storm warning. Now that situation is changing continuously, and that situation has a lot of non-linearity and uncertainties that require dynamicity in your output. And so you have to be very precise, and accurate, and hyper-local, and at the same time be able to tell the user the level of uncertainty they are faced with.
And in such situations, we felt that our technology can help, as opposed to just giving them a correlation, or a risk map. Things are going to change, and these characteristics are going to be event-specific and location-specific — even to the last hour, or the last second.
And of course, there are hundreds and thousands of examples of multivariate problems where machine learning is much more suited than just regular pure analytics.
Q: “Can AI tell us what percentage of deaths in the US (due to COVID-19) are due to poor governmental action versus simply very poor generalized health (obesity, heart disease, diabetes, etc.) in our population? Can we measure those things through AI?”
Ahmad: I’m not an epidemiologist — disclaimer — but what we are learning from some of the epidemiologists we have engaged with in our pandemic project, is that there is not enough data about how the virus is spreading, as well as how the virus is actually affecting the human body.
So until we collect data, and as we come out of the COVID-19 event, and we collect more data about the transmissibility of this virus, and really understand at a human body level how it actually affects different type of people with different types of comorbidities, etc. — only then will we be able to answer these questions.
But it’s really good that we have these tools, because once the data will be available — and we are collecting a lot of it in this day and age — we will be able to answer those questions going forward.
To access a full recording of the event, please visit Impact Entrepreneur’s website.