Fast Recurrent Processing via Ventrolateral Prefrontal Cortex Is Needed by the Primate Ventral Stream for Robust Core Visual Object Recognition [video]
Date Posted:
October 19, 2020
Date Recorded:
October 19, 2020
CBMM Speaker(s):
Kohitij Kar ,
James DiCarlo Speaker(s):
Arash Afraz, NIH
All Captioned Videos Publication Releases
Description:
Authors Kohitij Kar and James DiCarlo discuss the research behind their newly published paper in the Neuron journal with a statement from Dr. Arash Afraz of the NIH on his outside perspective of the impacts of this study.
[MUSIC PLAYING] KOHITIJ KAR: For the last decade or so, our lab has pioneered the experimental and computational approaches to study the brain systems that give rise to visual object recognition. We do that by recording hundreds of neurons in the primate visual stream as well as testing humans and monkeys on multiple behavioral tasks.
At the same time, we are building brain tissue mapped computational models that can predict and explain this data. These models are getting better every day. However, we still don't have models that perfectly match the primate brain. So what are the missing pieces?
One such missing piece is the fact that these models are predominantly feedforward in nature, which means that the information flow is unidirectional. For example, information flows from A to B to C, but not from B to A or C to B. However, we know that our brain is heavily recurrently connected.
So last year we published a study where we showed evidence that the inferior temporal cortex, which lies at the top of the ventral visual cortical hierarchy-- that brain area has neural signals that are supportive of recurrent computations in the brain. But, at the time, we did not know where these recurrent signals are coming from. Is it coming from areas downstream of IT? Areas within IT? Or areas within the entire ventral stream?
So in this study we specifically targeted a downstream area of IT, which is the ventral lateral prefrontal cortex. Our question was what happens when we temporarily inactivate this area? What happens to the behavior of the monkey? What happens to the neural data that we had previously seen in the IT cortex?
When primates look at an image, the information travels through the eye and through the ventral stream to reach the ventral lateral prefrontal cortex, which has reciprocal connections to the ventral stream. In this study, we injected a muscimol in the ventral lateral prefrontal cortex to weaken its feedback to the ventral stream and test its role in core object recognition. We implanted [INAUDIBLE] in IT test how this inactivation affects IT population activity while the monkeys perform the object recognition tasks.
From our earlier work, we had identified two sets of images, ones that are solved early in IT and ones that are solved late. Our hypothesis essentially was that recurrence is more critical for images that are solved later. Therefore, we predicted that if we disrupt the relevant recurrent circuits, it is the late solved images that will show higher deficits and this is what we observed both in IT decodes and behavior.
ARASH AFRAZ: Well, the paper is about feedback. And feedback is a big old question in systems neuroscience. We know that feedback exists. We just didn't know how to study that with the principle of parsimony that you'd like to explain things with the least number of variables. When you deal with a complex variable like feedback, you kind of like to ignore it and explain things without it, mainly because it is hard to study feedback. You need to make interventions in two different parts of the brain and all that. This paper is doing that.
So, in that sense, not only I think the paper is studying the role of feedback, which is a very important question, but also they are setting a role model for the rest of us on how to study the effect of feedback. They are creating a paradigm for studying feedback. And I think that's quite valuable.
JAMES DICARLO: This is really far more the beginning of the story than the end. Now that we have these kind of results, we now need to upgrade our models to go from these feedforward models of the ventral visual stream to these more recurrent models of ventral stream that include prefrontal cortex as one of several circuit nodes that are now shown to be critical to this very detailed dance that happens between IT and other brain areas to produce these higher and more impressive patterns of results.
And this will provide better models of both the monkey brain itself and, therefore, the human visual system but also is clearly going to lead to better computer vision algorithms as we can see that circuit add value relative to the current feedforward element computer vision algorithms. So it's really quite an exciting time at the interface between brain science and figuring the scientific hypotheses there and translating that into impacts on things like computer vision and perhaps even AI more broadly.
[MUSIC PLAYING]
Associated Research Module:
have an interactive transcript feature enabled, which appears below the video when playing. Viewers can search for keywords in the video or click on any word in the transcript to jump to that point in the video. When searching, a dark bar with white vertical lines appears below the video frame. Each white line is an occurrence of the searched term and can be clicked on to jump to that spot in the video.
