From the scent of flowers to the taste of wine, our perception is shaped by prior knowledge and expectation — a cognitive mechanism called top-down control.
Researchers at the University of California, San Diego School of Medicine report that associative learning shifts how the visual cortex operates by increasing the influence of top-down signals. The study, led by Takaki Komiyama, PhD, assistant professor of neurosciences and neurobiology, was published online in Nature Neuroscience on July 13, 2015.
Using mouse models, the team found that when an animal assigns new meaning to a previously neutral visual stimulus, inputs from higher brain regions play a larger role in driving activity within the primary visual cortex. “When the mouse learns that a visual cue predicts something important, top-down control has a much stronger effect on activating the visual cortex,” said first author Hiroshi Makino, PhD, a postdoctoral researcher in Komiyama’s laboratory. “These top-down signals interact with specific neuron types in V1 to change how the circuit operates.”
Top-down processing refers to how expectations, memories and internal models guide what we perceive. In contrast, bottom-up processing begins with raw sensory input. Everyday examples include how our brain fills in missing letters in a familiar word or interprets ambiguous images based on context. This study examined how learning alters the balance between those two streams of information and which circuit elements mediate the shift.
To probe the circuitry, the researchers monitored activity in multiple cell types and pathways in mouse primary visual cortex (V1) while animals learned associations over several days. They used chronic two-photon calcium imaging to follow the same neurons through the learning process. The key elements under study were layer 2/3 (L2/3) excitatory neurons in V1, layer 4 (L4) excitatory neurons that provide the primary bottom-up sensory drive, long-range top-down projections from the retrosplenial cortex (RSC), and somatostatin-expressing inhibitory neurons (SOM-INs) that regulate local cortical activity.
The results reveal a coordinated, learning-dependent rebalancing of input streams. During training, responses in L4 — the main conduit for bottom-up sensory information — gradually weakened. At the same time, top-down inputs originating in the RSC grew stronger. Corresponding with these shifts, L2/3 neurons developed a ramping temporal profile of activity, which could encode the timing of the learned association. Parallel changes were observed in RSC inputs, suggesting that the ramping signal reflects enhanced top-down timing information arriving at V1.
Learning also reduced the activity of somatostatin-expressing inhibitory neurons in V1. Because SOM-INs can gate distal inputs to pyramidal cells, their suppression during training may open a window for top-down signals to exert greater influence. Supporting this idea, the study showed that either inactivating RSC inputs or artificially activating SOM-INs partially reversed the learning-induced changes in L2/3 activity. Together, these manipulations indicate that both enhanced top-down drive and altered inhibitory gating contribute to the observed shift in cortical operation modes.
These findings support the view that sensory cortices do not simply passively represent the world; instead, they combine incoming sensory data with internally generated predictions and expectations. By demonstrating specific circuit mechanisms — including weakening of bottom-up drive, strengthening of top-down inputs, and changes in inhibitory control — the study clarifies how learning can tip the balance toward prediction-driven perception.
Beyond basic neuroscience, the results may inform our understanding of disorders in which perception becomes abnormal. “In addition to revealing circuit mechanisms underlying these learning-related changes, our findings may have implications for the pathophysiology of psychiatric diseases, such as schizophrenia, that generate abnormal perception,” said Makino.
Funding: This research was supported in part by the National Institutes of Health (1R01NS091010-01, 1R01DC014690-01), Japan Science and Technology Agency (PRESTO), Pew Charitable Trusts, Alfred P. Sloan Foundation, David & Lucile Packard Foundation, Human Frontier Science Program, McKnight Foundation and New York Stem Cell Foundation.
Source: Yadira Galindo — UCSD
Image Credit: Polina Tishina (public domain)
Original Research: Abstract for “Learning enhances the relative impact of top-down processing in the visual cortex” by Hiroshi Makino and Takaki Komiyama, Nature Neuroscience. Published online July 13, 2015. DOI: 10.1038/nn.4061
Abstract
Learning enhances the relative impact of top-down processing in the visual cortex
Theories propose that sensory cortices adapt during learning by increasing modulation from higher brain areas while reducing bottom-up sensory drives. To investigate the circuit mechanisms behind this shift, the authors examined activity in layer 2/3 excitatory neurons of mouse primary visual cortex (V1), layer 4 excitatory neurons (the principal bottom-up source), and long-range top-down projections from retrosplenial cortex (RSC) during associative learning over several days, using chronic two-photon calcium imaging. During learning, responses in L4 progressively weakened while inputs from RSC became stronger. Concurrently, L2/3 neurons developed a ramping response profile, potentially encoding the timing of an associated event, a change that coincided with similar dynamics in RSC inputs. Learning also reduced activity of somatostatin-expressing inhibitory neurons in V1, which could gate top-down inputs. Finally, either inactivating RSC or activating SOM-INs partially reversed the learning-induced changes in L2/3. These results reveal a learning-dependent, dynamic shift in the balance between bottom-up and top-down information streams and identify SOM-expressing inhibitory neurons as a control point in this process.