The Role of Bottom-up and Top-Down Information Cycles in Visual Processing
Written By: Ayush Saha, Edited By: Sherin Lajevardi
Intro
One very popular area of interest in neuroscience research currently is top-down processing, or “the interpretation of incoming information based on prior knowledge, experiences, and expectations” (Rousay, 2021). Understanding how the brain interprets sensory stimuli is key to understanding how we interact with our environment. It is also important for computer vision work, as it allows for the development of AIs with more human-like processing of the world. That is why, in this article, I shall discuss recent neuroscience research on top-down processing. My aim is to explain three distinct research projects from the National Center for Biotechnology Information (NCBI) and their findings in a manner that allows the general public to understand.
Key terms: Top-down projections, contextual feedback, bottom-up projections, visual cortex, RSC, higher order processing, neural synchrony, neural response amplitude
The Visual Cortex
The visual cortex (VC) is one of the major brain regions of interest for this domain of research. This region comprises six different layers, namely, L1, L2, L3, L4, L5, and L6. The way the VC works is so: as light signals pertaining to a scene are passed into the retina, they are converted to electrical signals, which traverse to L1. A well-known theoretical underpinning of the working of the scene perception process is that L1 projects the information up to certain higher-order regions,1 which, in turn, send projections of information back. These feedback projections play an important role in modifying the bottom-up sensory information drive occurring throughout the layers of the VC. L2 receives the partially processed image from L1 and processes more details of it. L2 will also send feedforward projections to higher-order regions and receive feedback projections. The phenomenon repeats throughout the layers of the VC until the brain’s perception of the scene sufficiently matches its sensation of the scene. (Kveraga et al., 2007).
Bottom-up Processing Versus Top-Down Processing
Bottom up processing is the projection of sensory information upward in the brain, from regions dedicated to basic information-processing to regions which carry out higher-level processing, or higher-order regions. Top down processing is the perceptual analysis of the sensory input carried out by the higher-order regions. In simpler terms, bottom up processing is related to sensation, while top down processing may be thought of as perception.
A practical example of this process would be the retina’s light receptors’ bottom-up sensory drive of visual information. The information would first be sent to the visual cortex (VC). It would traverse multiple layers of the VC, then be sent to a number of higher order regions, including the cerebral cortex. These regions give rise to the top down perceptual analysis of the sensory input.
The functioning of their perceptual analysis is quite sophisticated. The top-down information feedback projections have complex interactions with the bottom up sensory drive. In fact, object detection classifications are formed immediately as the sensory information drive reaches the visual cortex, before even being fully processed. These classifications are important to the brain’s perception of a visual scene due to the heavy noise of the incoming sensory data. (Kveraga et al., 2007).
Neural Synchrony As a method to test brain functioning
The analysis of brain waves and neural firing patterns is a common method used to attain information on the brain’s functioning. Numerous studies have found that groups of neurons which fire synchronously tend to be linked to one specific function. (Kveraga et al., 2007).
Researchers Kveraga et al. (2007)2 found neural synchrony to occur in both regions of local processing and across disparate neural regions. The locally synchronized neurons were working to bind together local stimulus features (for example, to put together a representation of an edge of an object), and the globally synchronized neurons were working to exchange bottom-up and top-down information across the brain. The researchers believed that in bottom-up and top-down neuronal interactions, “synchrony reflects match in stimulus detection, while asynchrony represents mismatch or error.” (Kveraga et al., 2007).
Neural Activity Changes Pertaining to Top-Down Processing in the Visual Cortex
The interactions between top down neural projections and bottom up sensory drives have been found to improve the processing of numerous types of visual stimuli (Kveraga et al., 2007). In other words, contextual feedback projections that are fired in response to semi-processed bottom-up sensory drives can lead to faster processing of a visual stimulus- specifically a visual stimulus that is a memorized member of a certain category of entities.
An example of this phenomenon is the findings of researchers Hiroshi Makino and Takaki Komiyama (2015)3 . The test subjects, mice, were given a mild tail shock if they incorrectly recognized a certain type of visual stimuli. Within a couple days of training, the number of milliseconds it required for them to process the various onset test stimuli categories decreased. At a microscopic level, the neurons of each mouse’s retrosplenial cortex (RSC)4 were activating more rapidly in response to the presented visual stimuli. Over the days of training, these neurons’ firing response increased in amplitude, which signified enhanced activity of the region for categorizing visual stimuli. This change corresponded with a decrease of the neural response amplitudes of the layers L2, L3, and L4 of the primary visual cortex. This meant that less neurons in L2, L3, and L4 were responding to the onset stimuli. However, at the same time, the remaining active neurons in L2, L3, and L4 were responding faster to the stimuli.
In sum, as the RSC of each mouse’s brain improved at categorizing the visual stimuli, it sent top-down feedback projections that made the visual cortex process the stimuli faster and require fewer active neurons.
Research findings by (Kveraga et al., 2007) add to our understanding of the bottom up-versus-top down information processing system in the visual cortex. Local neuronal synchrony (i.e. between L1 and L2) was observed to occur at higher frequencies, while global synchrony (related to long-range information projections, i.e. between the RSC and L2) was observed to occur at lower frequencies, mainly in the theta/alpha range. Furthermore, the researchers found that bottom-up signals and top-down signals were mapped onto separate interacting neural populations.
Visualization
To better visualize these findings and understand their meanings, picture a man watching a moving object inside an office room. The entire scene before his eyes is full of complex details, such as shadows, textures, etc., so he can’t process all of it at once. That is why top-down projections are important to his perception of the scene. Narrowing down on his perception of the moving object alone, there are multiple steps which occur:
- As the electrical signals pertaining to the object are sent into his visual cortex (for low level processing), basic information is first extracted, such as the general shape of the object.
- This information is projected to a couple higher order regions (likely including the RSC), which generate contextual associations. In turn, the higher order regions send back down hypothetical contextual projections.
- As the lower level regions are guided by information on the possible categories of the object, they are able to process the sensory data with more efficiency and are less sensitive to scene noise.
Recall that this process will occur at a very high speed- within milliseconds- due to the facts that the VC and RSC acquire ramp-up responses to familiar stimuli and the RSC stores memorized representations of various objects.
Discussion
“The brain is regarded as proactive in nature: rather than waiting to be activated by sensations, it is constantly generating predictions that help interpret the sensory environment in the most efficient manner” (Kveraga et al., 2007). Each person’s ability to perceive the world is truly a multi-year-long development, for the brain constantly configures the higher order networks which source top down projections so as to enable maximum interpretation efficiency. These configurations coincide with changes to lower-level sensory drive dynamics, in that the brain will attempt to acquire ramp-up response profiles to memorized stimuli while limiting the number of bottom-up sensory neurons in use. This ramp up response profile coincides with a greater response amplitude of the higher order regions in use.
The phenomena I described in this article do not only pertain to perceiving objects. High-level classifications of scene properties “have been demonstrated to manifest themselves in processes ranging from perception and memory to stereotypic judgments and prejudice.” Gist information can also be used to rapidly predict actions and movement patterns following a stimulus (i.e. a man running from a bear) (Kveraga et al., 2007).
The information processing that occurs in the visual cortex holds roots to many of our fundamental cognitive processes. Thus it explains much about our ability to perceive and respond to the world.
Footnotes
1 Higher-order regions are regions of the brain which carry out higher-level, or more complex, information processing.
2 More about this work may be found here.
3 More about this work may be found here.
4 A higher-order region that has been observed to send top-down feedback projections.
References
- Kveraga, K., Ghuman, A. S., & Bar, M. (2007, November). Top-down predictions in the cognitive brain. Brain and cognition. Retrieved November 26, 2021, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2099308/#!po=28.8360.
- Makino, H., & Komiyama, T. (2015, August). Learning enhances the relative impact of top-down processing in the visual cortex. Nature neuroscience. Retrieved November 26, 2021, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4523093/.
- Marek, S., & Dosenbach, N. U. F. (2018, June). The frontoparietal network: Function, electrophysiology, and importance of individual precision mapping. Dialogues in clinical neuroscience. Retrieved November 25, 2021, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6136121/.
- Rousay, V. (2021, January 21). Bottom-up processing. Bottom-Up Processing | Simply Psychology. Retrieved November 25, 2021, from https://www.simplypsychology.org/bottom-up-processing.html.
