Visible gorilla study finds we are better than we thought at spotting unexpected objects
Published in the Proceedings of the National Academy of Sciences, a research paper led by Clinical Associate Professor of Data Science and Psychology Pascal Wallisch adds new insights to “inattentional blindness”
Humans are good at noticing unexpected objects while focusing on another activity if the object is moving fast, a study by a team of NYU researchers found. The results challenge a widely held concept based on the 1999 “invisible gorilla experiment” that showed our ability to notice unexpected objects becomes impaired when attention is directed elsewhere.
The NYU study “The visible gorilla: Unexpected fast — not physically salient — Objects are noticeable” published in the Proceedings of the National Academy of Sciences, was led by Clinical Associate Professor of Data Science and Psychology Pascal Wallisch. Co-authors include Aimlab CEO and founder Wayne Mackey, Data Engineer at Mount Sinai Health System Michael Karlovich, and NYU Silver Professor of Psychology and Neural Science David Heeger.
In the original “invisible gorilla experiment,” participants watching a video of students passing basketballs were tasked with counting passes between players with white shirts. Their attention on the task they reported not noticing a person wearing a gorilla costume in the mix.
The recent PNAS study found that while engaged in the same pass-counting task, participants were more likely to see the ‘NYU gorilla’ if it was moving faster than the original experiment or if it was leaping instead of walking. To ensure the phenomenon generalizes beyond gorillas, the NYU team conducted an additional series of experiments. Over 3,000 participants were asked to count how many randomly moving dots of a specific color crossed a central line while an unexpected moving triangle crossed the screen at various speeds.
To learn more about the research study, what data science principles were at work, and who got to wear the gorilla costume, CDS interviewed Pascal. Read our Q&A with the psychologist and data scientist below!
The focus of the study is on “inattentional blindness”. Could you briefly explain what this is, perhaps giving an example?
A good example is riding a bike. Over time the task becomes what’s called ‘automatic’, meaning that it requires no conscious thought anymore. In contrast, while you are learning how to ride a bike, it takes up all your mental resources. Inattentional blindness is the idea that tasks that take up mental energy suck up all this attention in your head, and you fail to notice things in plain sight. Inattentional blindness is a consequence of paying attention to something very deeply; the idea is that you become literally blind to things right in front of you that would otherwise be perfectly visible.
Could you talk a bit about deciding what adjustments to make to the “invisible gorilla experiment”?
The way the findings from the invisible gorilla experiment are commonly interpreted is that if your attention is engaged in a task, you will inevitably miss other things. Even if it’s moving or a gorilla! This has been the dogma in psychology for over twenty-five years. However, this never made sense to me, for the following reason.
If it is true that you will not notice clearly visible things if your attention is engaged elsewhere, this would leave organisms vulnerable to attack from unexpected things. Let’s say you’re engaged in a task, like writing an article, and then there’s a tiger right there! This fits in with the prevailing view in contemporary psychology that people are fundamentally incompetent. They don’t perceive things correctly, which we know from visual illusions; they don’t remember things right, which we know from studies on false memories; and they can’t make rational decisions, which we know from research on heuristics and biases. And yet, here we are, the result of an unbroken evolutionary chain reaching back a billion years, so this is just not plausible.
Instead of saying the organism is fundamentally incompetent, we should be asking what the system is optimizing for. Our research shows that when you’re focusing on a task the brain hedges that there might be something more relevant out there that you didn’t anticipate, and the fast motion will flag it. So this system works beautifully because it’s filtering for relevancy under uncertainty.
In addition, while I was completing my PhD in visual neuroscience, I noticed that few of the studies on inattentional blindness varied the speed of the unexpected object, and none tried fast speeds. Fast speed is an especially good candidate for an evolutionary hedge because motion signals the presence of lifeforms and as most object speeds are slow, fast motion would stand out as potentially particularly relevant. Thus, we hypothesized that fast motion would be likely to break inattentional blindness.
How did data science play a role in the research process?
There are four different data science relevant aspects of this research. The first one might be the most straightforward. We had a large sample size (~4,500) which was about an order of magnitude more than typical inattentional blindness research. Sample size was particularly crucial because we could only do one trial per person — once participants notice the unexpected object, it is no longer unexpected. So this illustrates a key principle of data science: statistical power matters.
The second thing is once you have this sample size instead of saying these conditions are different, we can estimate the slope using bootstrap analysis. We can resample that data set to get a quantitative understanding of the detectability of unexpected objects as a function of speed.
The next thing is a little more subtle. I was teaching the Big Data class this spring and was struck by how similar the demands are between what a big data processing system must handle and how the brain handles things. The idea being that to effectively process big data you must reduce the data as quickly as possible, ideally by filtering upfront by relevance. Otherwise, you will get overwhelmed by the data. With so much information always coming at you, the human visual system, and the brain work in a similar way by triaging what’s relevant. It gives us a new understanding of the brain from a data science perspective. Psychologists would say people are inattentionally blind, but data scientists would say their filtering for relevance, as they must.
The fourth thing would be an understanding of the objective function. In a machine learning system and reinforcement learning framework, every organism has an objective function. I think the objective function of a human being is not necessarily to represent the outside world perfectly but to manage uncertainty elegantly.
The study highlights the importance of testing under the right conditions. What lesson can be taken away from this study about scientific research processes?
The research highlights the importance of testing systems fairly in the environment they were designed to be in. For example, imagine you don’t know what a propeller plane is, and you come across one in a hangar. You might never figure out what it’s for, if you just leave it in the hangar. To figure out its purpose, you would have to test the plane under regular conditions. That’s a big lesson for data science too. If you don’t test your system under realistic conditions, you might be completely off.
Did you dress up in the gorilla costume?
In hindsight, I wish I had! Or added it to my collection of vision science props. We had a former student who volunteered for the study, and they still have the costume. Maybe I’ll ask them to donate it to a science museum!
by Meryl Phair
