Unlocking AI for Psychology
What can AI's current successes tell us about how it can help psychology?
Posted January 24, 2022 | Reviewed by Devon Frye
- By looking at the successes of Artificial Intelligence (AI), we can understand how it could be applied to psychology research.
- AI is largely dependent on Deep Learning, a technique which is most effective when it has many data points—millions, not thousands.
- Deep Learning has been most successful in analyzing unstructured data, like pixels in an image or words in a blog, as opposed to survey results.
- AI is likely to have the biggest benefits in psychology research using mobile sensors, social media data, and brain/physiological measurements.
What does the next generation of psychology research look like?
Artificial intelligence (AI) is changing the world. Largely, this is because of all the ways it’s being used by tech companies. AI is behind facial identification (like when Facebook automatically recognizes someone in a photograph you post), translation (like the Google Translate function), speech recognition (like when you give instructions to Alexa, Siri, or other electronic assistants), and many other emerging services. Can psychologists leverage this powerful new technology to gain new insight about how people work?
The Role of Deep Learning
Key to the increased success of AI in solving practical problems for businesses has been one particular technique: Deep Learning. Where can Deep Learning help psychologists? First, let’s consider when it’s most successful.
Deep Learning is a way of developing predictive models that become very accurate when they have access to a lot of data. This is in contrast to more typical statistical tools used by psychologists, like regression. Using these tools, you can create a model that will make reasonably accurate predictions, but predictions don’t often improve much after you’ve collected, say, 1,000 data points. So going from 1,000 to 1 million data points isn’t that helpful.
With Deep Learning, however, predictions continue to get more and more accurate as bigger and bigger datasets—multiple millions of measurements!—are used. So Deep Learning is a technique that goes hand-in-hand with another recent buzzword: Big Data. With Big Data (lots of measurements) you can do very good Deep Learning, and that means computerized systems can solve very complex problems (like picking out who is in a photograph). For psychology to fully take advantage of Deep Learning, then, it needs to have large datasets.
Deep Learning is also particularly successful when it’s applied to “unstructured data.” This means data where there isn’t a specific meaning for each measurement included. For example, every response to a survey is the specific answer to a question. But every pixel in a picture isn’t inherently meaningful. Deep Learning has been a game changer for processing images, but not necessarily for analyzing surveys.
Deep Learning will therefore likely be best used on unstructured data relevant to psychology. This might mean analyzing neuroscience data, where each measurement is activation at one specific part of the brain, or social media data, where there are millions of words written about whatever people wanted to comment on.
Finally, it’s worth remembering what Deep Learning actually does well: make predictions, based on data it has already seen. Predictive algorithms can be useful for businesses delivering a specific product to customers, but their application in the scientific process can be a bit more nuanced.
Psychologists may want to know whether—or how well—you can predict an outcome (like a depression diagnosis or a hiring decision) based on a given set of inputs (like content on social media or a job interview). But they’re also going to want to know how it works. What is it about the interview that landed you the job? Interpreting a Deep Learning model is not impossible, but it is a lot more complicated than interpreting more traditional statistical models.
Given that Deep Learning is best when there is a large dataset, the data is unstructured, and the goal is prediction, here is a short list of places where I expect AI could yield huge benefits to psychology:
- Understanding Daily Life Outside of the Lab: Mobile sensors, like Fitbits or smart watches, are increasingly common consumer electronics. They allow for the collection of huge datasets, and that data is unstructured—it involves how the device moved in space at the millisecond level, or how quickly someone’s heart was beating for days and weeks on end. This type of data could be leveraged to predict when people want to spend time with friends versus alone, what personality type the individual has, or when the individual might be slipping into depression.
- Understanding (Digital) Culture Change: The online world is increasingly a place where culture is born, morphs, and grows. Unlike culture that exists just in the face-to-face interactions of people during a normal day, this digital culture is all (theoretically) available to be collected and analyzed. Again, this is big data (entire social media sites over years) that is unstructured (whatever people decide to post). Predictive models could, in this case, be a proxy for stable culture. When something is easy to predict from what came before, we’re in a more stable cultural period. When the next post, picture, or video is harder to predict, that could suggest that we’re in a place of cultural transition.
- Understanding Mind-Brain-Body Connections: The fields of cognitive neuroscience and psychophysiology have, for years, examined how specific psychological processes are related to brain and body processes. For example, research has explored how being stressed is related to heart rate and blood pressure, or how certain brain regions are related to viewing faces. Collecting data from certain types of sensors—like an EEG cap, which measures electrical activity at the scalp—is becoming cheaper and easier for researchers (and even consumers). That means it will be feasible to collect increasingly large sample sizes of data. Given that this type of research also often tends to focus on simple tasks that can be repeated over and over again (e.g., getting a person to look for a shape on screen dozens or hundreds of times), there are already many instances of people doing that specific task. Again, this is big, unstructured data. What will be useful here is asking smart, interesting research questions using prediction. This can certainly be done. For example, Stanislas Dehaene and colleagues used EEG recordings to develop a model of when a patient in a coma was likely conscious or not.
- Deep Fakes for Deep Study Materials: An emerging area of research in Deep Learning is generative modeling, where the model doesn't just learn to predict something from examples, but actually learns to create new examples from those it sees. For example, a Deep Learning model can be taught to generate realistic headshots of individuals who don’t exist, or to generate images in a specific style (e.g., a portrait in the style of Picasso). Deep Fakes allow for videos to translate one person’s actions onto a simulation of another person’s body. Although this has potentially scary implications for society, it allows for radically new ways of setting up psychology experiments. What if you wanted to manipulate a person’s perceived age, gender, or ethnicity? In a virtual environment, you could. A man could have their speech and actions mapped to a woman’s body, and they could experience how the interaction was different. A British accent could be added to an American speaker, to see how that changes perceptions. All kinds of social information that is expressed in an “unstructured way” could be manipulated, giving us greater insight into social processes.
This article provides a brief introduction to how psychologists can think of using Deep Learning as a tool for research. What I’ve provided here are just a few heuristics to guide your thinking, as you consider what benefits AI might have for understanding the mind. If this technology is embraced in psychology, the insights generated in the next couple of decades will be exciting!
Engemann, D. A., Raimondo, F., King, J. R., Rohaut, B., Louppe, G., Faugeras, F., ... & Sitt, J. D. (2018). Robust EEG-based cross-site and cross-protocol classification of states of consciousness. Brain, 141(11), 3179-3192.