How Robots Could Teach Us to Trust AI
Jason Hattrick-Simpers, Materials Research Engineer, National Institute of Standards and Technology (NIST)
The influence of artificial intelligence (AI) is everywhere in our lives. It helps us pick movies to stream, recommends old friends to reconnect with, and autocompletes our most pressing questions in search engines.
Modern AIs are powerful machines, ingesting reams of data and outputting useful relationships between data points. If you manage a baseball team, you can create an AI model that will take factors like the weather, the opposing pitcher and other aspects of the game situation to determine which batter to substitute to improve your probability of winning. The AI won’t tell you why it selected batter A over batter B over batter C, only that batter A provides the best chance for an optimal outcome like hitting a home run. But, if you had to select the next batter from the stands, its recommendations would be useless.
These are both known problems in modern AIs: They can’t tell you the “why” underlying the relationships they discover and are notoriously poor at predicting outcomes they haven’t seen before. For this reason, scientists often find it hard to trust AI systems because scientists constantly seek to understand the “why” of our observations.
The inverse of a modern AI is the traditional scientific theory, in which physical laws are condensed into elegant mathematical models that can be applied to a wide variety of similar problems. The same classical or “Newtonian” equations that describe the path a baseball takes can be used to describe the motion of the heavens. Moreover, if you know the initial velocity of any object (baseball, football, ham sandwich), you can predict its path without needing to measure something similar first.
However, if you tried to use the same set of equations to describe the path of an electron in a molecule, which obeys the bizarre rules of quantum physics, then the results would be less than satisfactory. This is a known limitation of traditional science: Different phenomena can be dominated by different sets of rules, and sometimes the connective tissue between physical regimes isn’t well known. For instance, we don’t understand how physics transformed into biology or how biology leads to the rise and fall of civilizations.
Moreover, there are often multiple potential conflicting explanations for an observation. Did the ball curve left because of its spin or because of a turbulent gust of wind? In some sense, science is brittle: The rules work until something unexpected happens, and a great deal of research effort is devoted to teasing out what exactly caused the ball to curve the way it did.
Traditional Science vs. AI
This is the origin of the current tension between traditional AI and science: AI can fit any dataset with exquisite accuracy but without any causality, while traditional science is built around rules that teach us why but can be broken. Truthfully, however, the tension is the result of a flawed decision to view the use of these techniques as mutually exclusive. Instead, a more fruitful approach is to combine the strengths of AI (flexible correlation masters) with the strengths of traditional scientific models (explainers of why) to create a symbiotic relationship.
Think of it this way: With a good understanding of Newton’s laws and AI models that correlate the weather, the pitcher’s arm motion, and how the bat is swung to the likelihood of a batter hitting a home run, you could create a hybrid model that could power a national championship. As an added benefit, knowing the scientific underpinnings would allow confident extrapolation to any batter/pitcher combination (maybe even someone from the stands if they took a few practice swings), and the AI could provide statistical tools to differentiate the effect of ball spin versus a gust of wind.
These exact ideas are currently driving the emerging field of scientific AI. In scientific AI, physical laws are incorporated into AI algorithms, creating a whole that is greater than the sum of its parts. By putting the “why” into our AI, we can generate predictions that allow us to see what influenced how the predictions were made or maybe even explain how the predictions were made in a human-understandable way. This would enable us as scientists to trust those predictions enough to try them out, even if they challenge our worldview.
Here, a new problem emerges. Scientific AI is so powerful, flexible and curious that testing its new ideas and separating genuine insights from extrapolation error is now the work of many lifetimes. A scientific AI doesn’t care if it’s wrong; each “error” just means the next set of predictions is better. But as human scientists, we don’t have many lifetimes to accomplish our work. We also like to sleep, eat and spend time relaxing at home, but the AI will update its model and make predictions as fast as we can provide it with fresh data.
Robots to the Rescue
Our scientific AI robot for formulating, making and testing corrosion-resistant alloys during a proof of principle in situ synchrotron experiment at the NIST Beam for Materials Measurement beamline located at the Brookhaven National Laboratory National Synchrotron Light Source II in January 2020. During a weeklong experiment, the robot taught itself to make corrosion-resistant zinc-nickel alloy coatings using measurements of structure, chemical composition and electrochemical polarization.
To keep up with the AI, our team has been designing robots that automatically perform the experiments recommended by our scientific AIs with minimal human intervention. The idea is simple. If you build a robot that can control the parameters that affect your experiment and understands the physical rules that lead to your final observation, then you can arrive at your desired outcome (new optimum or new insight) in days instead of decades. In the process, we get to know when our model is venturing away from solid ground if a series of observations are not explainable by any known models, while at the same time potentially finding and learning something about the blind spots in the conventional wisdom.
Our robots don’t look anything like C-3PO. They more closely resemble a collection of syringes, tubes, electrodes and stages stitched together with parts generated at your local maker studio. Our robots are powered by an AI that controls every aspect of our experiments. In our case, this means that our robot identifies a corrosion-resistant coating to test through an understanding of every previous experiment performed in the lab. It then mixes the ingredients, coats them onto a surface, tests the coating properties, and then decides upon the next coating to make based on the results, all without the humans needing to touch a single beaker. Why corrosion-resistant coatings? That is a topic for a different blog post, but suffice it to say that since corrosion mitigation and remediation costs 3.1% of the U.S. GDP, it is a big-time problem.
Using AIs, we have already demonstrated a hundredfold increase in the rate of the discovery of new coating materials, including a coating that could be implemented to increase the lifetime of medical stents.
Building and testing scientific AI with robots is exciting! At this moment, dozens of labs across the world are spinning up different and competing AI-driven robotic platforms to discover new, potentially more conductive solar cell coatings, safer electrolytes for lithium-ion batteries, and stronger 3D-printed structures, for example.
As scientists begin to develop interpretable and trustworthy scientific AIs, we have to remember that our models will be influenced by the uncertainty and errors contained in our measurements in ways that are not yet clearly understood. Uncertainty comes from inaccuracy and imprecision either in our observations or in how we make measurements. For instance, a radar gun in need of calibration may measure pitch speed as 100 mph versus 95 mph. Uncertainties tend to get carried through calculations in unexpected ways, and so the radar gun uncertainty could result in a model that predicts the ball will travel 193 meters (643 feet) plus or minus 193 m, meaning we have no idea where the ball will go.
As a neutral and nonregulatory agency, NIST has an important role to play in this new field by providing guidance for how to include and measure the impact of uncertainty and errors on scientific AI’s predictions. We do this by generating reference datasets that are designed to put a new algorithm’s ability to handle and explain sources of uncertainty or possible erroneous data through its paces. We also make our robots directly available to U.S. businesses and educational institutes to help them see the benefits of adopting best practices in robot and scientific-AI design.
Success in this area promises immense payoffs: A community consensus for generating AIs that are interpretable, explainable and trustworthy! More scientific discoveries! A better understanding of our world! And the freedom to ask high-risk questions whose answers could be revolutionary! It is awesome to know when I walk into the lab every morning that our work here at NIST has the potential to transform scientific AI into a trustworthy and sustainable tool for maintaining American technological competitiveness.
This post originally appeared on Taking Measure, the official blog of the National Institute of Standards and Technology (NIST) on April 28, 2020.
To make sure you never miss our blog posts or other news from NIST, sign up for our email alerts.
About the Author
Jason Hattrick-Simpers is a materials research engineer in the Materials and Manufacturing for Sustainable Development group at the National Institute of Standards and Technology. He got his B.S. in mathematics and physics at Rowan University and his Ph.D. in materials science and engineering at the University of Maryland. Prior to joining NIST, he was an assistant professor of chemical engineering at the University of South Carolina. When he isn’t doing science, he is either running, eating, playing videogames or spending time with family.