Fold, tape, flip; fill, fold, tape. Fold, tape, flip; fill, fold, tape.
When Hurricane Maria ravaged the island of Puerto Rico in September 2017, Pedro Rodriguez and his wife, Laura, nervously anticipated word from their loved ones. For two weeks, they waited in the dark. “Like, literally, the entire island disappeared electronically,” Pedro says. “For days.”
While their minds raced, and their nerves frayed, they became cardboard-box-making specialists.
Fold, tape, flip; fill, fold, tape.
They tried to avoid wringing their hands, so instead busied them by making care packages. “We just bought as many supplies as we could. We prepared boxes, and as soon as we heard ‘Post Office open,’ down they went,” Laura recalls.
Canned chicken and vegetables. Batteries. Handheld fans. If they thought it might provide some help, comfort or assistance as their families — and their homeland — recovered from the monstrous Category 5 storm, into the boxes it went.
“Our way of coping,” Laura says, “was to try to help.”
Perhaps, then, it’s not surprising that more than three years later, Pedro Rodriguez finds his professional expertise has enabled him to expand that ethos to a grander scale.
As the senior technical leader of multiple deep learning artificial intelligence projects at the Johns Hopkins Applied Physics Laboratory in Maryland, Pedro has a vast portfolio. One highlight is APL’s Humanitarian Assistance for Disaster Relief program, a machine learning and artificial intelligence-enabled program to collect and process overhead imagery into categories for analysis. It allows disaster recovery teams to assess in hours what has previously taken days, or sometimes weeks.
HADR, as the program is known, processes flood segmentation (locating and marking areas of flooding), road analysis (identifying blocked and unblocked roads), and building damage assessment. That last one, which classifies buildings based on the Federal Emergency Management Agency’s (FEMA) protocol, categorizes structures into no damage, minor damage, major damage and completely destroyed.
In partnership with the Department of Defense’s Joint Artificial Intelligence Center, HADR was pressed into service to aid FEMA’s recovery efforts after 2018’s Hurricane Florence and 2019’s Hurricane Dorian devastated parts of the U.S. and the Caribbean islands. It was called to action in 2020 after Hurricane Laura, a deadly and destructive Category 4 storm, battered Louisiana.
“It was coincidence,” Pedro explains of the link between his professional project and his personal hurricane history, noting APL was able to quickly leverage technology developed for another project when the light bulb went off that it could dramatically assist in these situations.
“ Experiencing hurricanes, knowing especially what my family went through most recently, has played a huge role in my enthusiasm for this project. Once we started using this technology for disaster relief applications, I just knew it was an area we could profoundly impact.”
“But now, we are all in.”
The silence that turned Pedro and Laura into mailing mavens finally broke after a couple of weeks. One of Pedro’s aunts was able to drive to a specific spot on the island where she’d search for cell phone service and send word that, fortunately, all were OK. After two weeks, the landline in Laura’s mom’s house started working again, and neighbors she hadn’t seen in years lined up at her door to ask for its use.
Pedro’s lasting impression of his first visit back after Maria was visual. “There was no green,” he says, all the trees and bushes left bare from the storm.
Studies show that occurrences and strengths of major hurricanes are expected to increase as the planet warms. In 2020, there were a record 30 named storms during the Atlantic hurricane season. Personal experiences aside, the world’s ability for people, states and countries to recover from these dangerous events is, and will continue to be, paramount. Pedro knows that.
“Experiencing hurricanes, knowing especially what my family went through most recently, has played a huge role in my enthusiasm for this project,” he says. “Once we started using this technology for disaster relief applications, I just knew it was an area we could profoundly impact.
“I want this to be a strength of our group for a long, long time.”
It was the search for precision that led Pedro Rodriguez, an electrical engineer in a family of civil engineers (and one accountant), to machine learning and image recognition. Once, in an electromagnetic class in college at the University of Puerto Rico, he was designing an antenna for radar that wasn’t working quite right. His professor took a look and devised a quick fix.
“He pulled out a knife and cut a piece of the antenna off,” Pedro recalls. “And suddenly, it worked. I was like, ‘Are you kidding me? Is this really engineering?’ The fact that it was not perfect, it freaked me out. To me, pixels in digital images are perfection.”
“ I’ve heard people say he is the OG — the original gangster — of transfer learning here at the Lab. He was doing deep learning before it was cool.”
About 10 years ago, before deep learning, machine learning and artificial intelligence were as commonly interspersed into the public lexicon, Pedro was a young(er) engineer focused on transfer learning — training a neural network on a problem similar to the one you’re trying to solve, then applying part of it in a new model on the actual problem.
“I’ve heard people say he is the OG — the original gangster — of transfer learning here at the Lab,” Ryan Amundsen, a signal processing engineer at APL and a mentee of Pedro’s, says with a chuckle. “He was doing deep learning before it was cool.”
“Because of how smart he is and his outgoing nature, I see Pedro as the person who kind of brought computer vision to the Lab for AI purposes,” added Dean Fisher, who manages the Tactical Intelligence Systems Group at APL, putting Amundsen’s observation in, perhaps, more formal terms.
“When deep learning came about, I recognized that I could do with deep learning neural networks what I was doing with this other thing, called HMAX [Hierarchical Model and X],” Pedro explained. “The cool part was that it was like a one-week fix, and the results were immediately 20-40% better.”
PedroNet — a grayscale, deep learning neural network that Pedro trained himself in APL’s maker space, Central Spark — was born. And his profile within the Laboratory skyrocketed.
“Word got around really fast,” Pedro recalls of the interest in his work. “I was literally walking around with PowerPoints, showing it to people.”
Pedro had wanted to be an astronaut and applied to NASA’s program (before completing his Ph.D.). He was introduced to the Laboratory (and his wife, who works as a systems accountant at APL) through an undergraduate internship at NASA’s Goddard Space Flight Center, a short distance from the Laboratory. His career, which he put on a beeline for the Lab after that initial taste from Goddard, began with work on classified projects in APL’s Force Projection Sector, where PedroNet came about. For now, Pedro is no closer to his goal of being an astronaut, and he’s still never worked in the Lab’s Space Exploration Sector.
In many ways, the successes that led him here — now a prominent voice in the Laboratory’s AI and machine learning efforts — have come because of paths that diverged from their originally planned end point.
That was the antithesis of what he sought as a young professional. Pixels with a set value, processes that return specific — expected — results; these unshakeable constants lured him. “Even though machine learning has some ambiguity to it, I remember feeling a sense of calm when I gravitated toward this field and computer vision.”
PedroNet, a microexample of his technical acumen, was part of what launched him on his current trajectory. His gregarious personality, introspective nature and unapologetic ambition helped too.
“ I think much of it comes back to the question of ‘What do you have to do to be world-class?’ You identify what you need to do and you do it. That’s Pedro in a nutshell.”
“He was so open about sharing his knowledge and insight [that] people sought him out as the subject-matter expert in this emerging area,” Fisher said. “But you see the impact now with the many premier programs the Lab is working on. That’s all part of [APL Director] Ralph Semmel’s strategic vision to make APL an AI center of excellence. Pedro’s contributions there have been tremendous, and it started with collaboration and his willingness to share.”
“I think much of it comes back to the question of ‘What do you have to do to be world-class?’” added John Piorkowski, the chief AI architect in APL’s Asymmetric Operations Sector and head of its Applied Information Sciences Branch, of which Pedro’s group is a part. “You identify what you need to do and you do it. That’s Pedro in a nutshell. He will put his mind to something and work to achieve it.”
On a frigid afternoon in mid-January 2020, Pedro pulls out his smartphone and drops it on the table in his office at APL’s suburban Maryland campus. To his right, a whiteboard displays work-related scribbles on its upper half and a DO NOT ERASE warning for the content below it — self-portraits of sorts drawn by his daughters, Natalia and Elena. Tiny reminders of his world outside these walls.
Open on the screen of his phone is QuakeFeed Earthquake Alerts, an app monitoring earthquake activity around the world. A larger reminder of that same world. Pedro’s filtering it to only show quakes registering 4.0 or higher on the Richter scale since Dec. 28.
He turns the phone.
“This is where all my family is,” he says, zooming in on southwest Puerto Rico and the area where his hometown, Guanica, sits facing the Caribbean Sea. Red pins flood the region. His fingers tap the screen, removing the 4.0 or higher qualification, and the pins, denoting earthquakes, proliferate.
“Look at this,” he says, his tone escalating with each finger stroke as he moves the map around. “It is crazy. What the heck is this? When is it going to end? It feels like, every time it shakes, the damage caused takes the next step to being worse than Hurricane Maria.”
That is not a comparison Pedro makes lightly, particularly not after visiting Puerto Rico himself later that January and experiencing the aftershocks and tremors the locals understood could go on for a year or more. “The randomness of destruction was shocking,” he said. “It was weird to see a house completely destroyed and the house next to it perfectly fine. That just did not compute in my mind. What is the difference between this house and that house?”
This destruction was loud; it was obvious. Disasters are like that, Pedro knows. He may not have been on the island during Maria, but he’s certainly experienced his share of hurricanes. It is, oddly, the quiet that Pedro remembers most about them.
Even now, more than two decades after Hurricane Georges ripped through his hometown, it’s not the destructive force of the 140-mph wind gusts, the driving rains and flooding, or the catastrophic damage that spring to his mind first. It is serenity. The tranquility of a sky suddenly so clear he and his brother walked outside to stare up, through Georges’ eye as it passed over their house, and gaze at the moon in the silence.
“It was like the most beautiful night,” he says. “If it had been daytime, the sun would’ve shone on us.”
The peacefulness didn’t last — in part because the second wall of the storm was about to shatter its illusion, but, more to the point, because his father came running out of the house to corral his children back into the safety of its concrete walls. Minutes later, the storm’s pounding resumed. Their home was spared, but the family suffered through six months without power or water.
There was a lesson there for Pedro, though perhaps not one he realized. When it comes to disasters, every second counts. That urgency has guided Pedro over the course of his career. How can we do this better and faster?
That’s grown into its own legend in a way, as he’s taken on the moniker of a “BS detector.” In more elegant terms, Pedro is consistent in that he can quickly detect when an approach is, well, hogwash.
“You can trace that back to 2012 or 2013,” Pedro says with a laugh when confronted about this nickname. “At APL, one thing we take seriously is our role as a trusted agent. I did a high-level briefing on deep learning, and I just brought it back to Earth. ‘It works for this; it doesn’t work for this.’ For a long time, there’s been so much hype about deep learning, and while today there’s probably hundreds of people at APL who could tell you the truth and not overhype it, at the time it was refreshing to hear someone not trying to sell it.
“A couple weeks later, I was getting invited to a lot more high-level briefings so I could help evaluate the ideas that were being sold.”
Beatrice Garcia, a software developer in Pedro’s group at the Lab, offers a slight, quiet chuckle when asked about the process to become project manager for the Lab’s Humanitarian Assistance for Disaster Relief Program in 2019. Garcia, who has more recently been an integral member of the APL team providing essential data collection, curation and aggregation for the Johns Hopkins Coronavirus Resource Center, describes her leadership experience before HADR as “very minor.”
“I think Pedro just took a risk,” she says, noting that there is an established mentorship pipeline in their group, which she easily tapped into upon taking the role.
The real reason is a bit more layered, the outermost one being that Garcia’s technical expertise and track record made her name one leadership quickly identified and brought to Pedro as a possibility. His reasoning, though, reveals more about his management style than it may about Garcia.
“ One thing many people struggle to appreciate about deep learning is that it has made things that used to take teams of 50 people — 20 of them Ph.D.s — years and millions of dollars to do, something a high schooler can do now.”
“The reason was twofold,” Pedro explains. “One thing many people struggle to appreciate about deep learning is that it has made things that used to take teams of 50 people — 20 of them Ph.D.s — years and millions of dollars to do, something a high schooler can do now. A high schooler can buy a quadcopter, put in a small graphics processing unit, fly it in the air and do car detection. Twenty years ago, that was a $5 million project. Somebody with 20 years of experience might say, ‘This is stupid, I’ve tried this for years and it never worked,’ but now, in 2021, it works.
“Sometimes experience lends itself to preconceived notions. I wanted an aspect of inexperience, somebody with fresh eyes, because I felt it could be an asset. I trusted the people who suggested Beatrice for this job, but the second part of this answer is that I strongly believe you learn by doing. You need to take risks. And in doing so, I knew we could exemplify that APL could put together young, diverse teams to tackle these kinds of projects, just like many of the deep learning teams in Silicon Valley.”
What Pedro hints at with his answer is not something he pinpoints himself, even later when asked about it directly: how does he, as a minority and hiring manager, approach diversity in a field that is historically not diverse?
“I kind of love being a minority,” he says nonchalantly, expanding to say that’s particularly pertinent in his role now as a supervisor in the field of artificial intelligence, in which researchers have identified technical issues relating to racial bias. “But what I am always seeking is diversity of thought.”
He’s noticed, for example, that he gets a high number of applicants from Puerto Rico and posits that it’s possible that’s related to his representation. There are times, too, when he’s wondered if that’s the pendulum swinging too far the other direction. “John Piorkowski once told me, ‘If somebody is super talented and wants to come to your group, and you being a minority became a magnet for them, that’s great.’ That’s very fulfilling to me.”
Within Pedro’s group, staff members have self-identified as members of at least five different races (including Hispanic, Black, Asian and Indian), and women make up 29% of the group — a number that nearly mirrored APL as a whole in 2020.
“He incorporates diversity without specifically trying to do so,” Laura Rodriguez says of her husband. “It happens naturally. When he started at APL, you could easily see that diversity wasn’t necessarily there. He stuck out as it. But now, there’s a little bit of everything, and everyone works well together. It’s the best group I’ve seen him work with in terms of everyone getting along. That says to me that his qualities are coming through. His representation isn’t ceremonial.”
The Humanitarian Assistance for Disaster Relief program may be just one piece of the work that Pedro leads at the Laboratory, but it’s an apt lens through which to view this topic. Disasters do not discriminate. They are not a problem central to only one group of people; they are an affliction the world over.
Solving them, or at least figuring out how help for those in harm’s way can become more expedient and effective, should be a challenge tackled by a group representative of that fact. Whether it’s this specific problem or the next one the group tackles, that won’t change.
“Pedro has fostered a community here in the group,” Amundsen said. “He will challenge your technical abilities, but you will get opportunities based on where he’s taking that trajectory of technical expertise. And really, he helps you understand that you should make time do that.”
More than a year after the earthquakes in Puerto Rico began, the aftershocks continue. In 2020, the Live data, Integration, Validation and Experimentation (LIVE) Lab at APL, where the Humanitarian Assistance for Disaster Relief program is run, added the code for the Johns Hopkins COVID-19 dashboard (tracking a disaster of a different kind). Meanwhile, Pedro’s focus remains steadfast.
The coronavirus pandemic has altered the way the world views just about everything, but one disaster does not cede to another. In 2020, fires still raged, and hurricanes still stormed. Even deep freezes so catastrophic as to short Texas’ power grid didn’t take a year off. Through it all, Pedro’s belief in building the Laboratory into the place for AI in disaster response — human-made, natural, medical or otherwise — only grew.
“I really hope to see a future where AI is automatically applied to all kinds of sensor modalities, and we can respond to emergencies faster than ever,” he said in late February this year. “I want a future where the sensors looking at the world were designed for AI, instead of sensors designed for human decision-making.
“In that world, we could respond faster because we wouldn’t even need a human in the loop. The AI could simply report ‘flooding in this road’ and recovery teams could just go, without a need for human confirmation.”
Those hopes are also deeply personal. Natalia and Elena, 6 and 9 years old, will live in the world Pedro is helping to create. It’s natural to think about how that will affect those you love. He comes back to a central thought: will his lived experience help improve theirs?
“I feel worry and hope about the future,” Pedro admits. “What will disasters — and recovery from them — look like in 50 years?
“But I also get excited about the possibility that my daughters will one day continue the work that we’re doing, that maybe they’ll have a role in that future. That would be really cool.”
Media contact: Amanda Zrebiec, 240-592-2794, Amanda.Zrebiec@jhuapl.edu
The Applied Physics Laboratory, a not-for-profit division of The Johns Hopkins University, meets critical national challenges through the innovative application of science and technology. For more information, visit www.jhuapl.edu.