Brain's Focus: How It Stays Sharp

Jim Crocker
23rd June, 2025

Brain's Focus: How It Stays Sharp

The proposed two-layer model, which incorporates quadratic computations (a), accurately accounts for neural responses to natural stimuli across visual areas V1, V2, and V4, achieving predictive performance comparable to a leading non-interpretable machine learning model (b).

Image adapted from: Rowekamp et al. / CC BY (Source)

Key Findings

  • Scientists at The Salk Institute and collaborators found that the brain uses "quadratic computations" to process visual information, involving complex interactions between visual features
  • These computations precisely balance and coordinate signals that excite and suppress neuron activity, allowing the brain to sharpen its focus on specific visual details
  • This unique processing helps neurons become highly selective for specific features in natural images, explaining how our brains robustly recognize objects
The human brain's remarkable ability to recognize objects with precision and reliability is a cornerstone of our perception, yet the exact neural processes that make this possible have largely remained a mystery. Understanding how the brain transforms raw visual information into meaningful perceptions is a fundamental challenge in neuroscience. Recent research conducted by scientists at The Salk Institute for Biological Studies, University of California - San Diego, and Centre National de la Recherche Scientifique has begun to unravel these complex mechanisms[1]. Their study investigates how neurons in different visual processing areas of the brain—specifically V1, V2, and V4—respond to natural visual inputs. They propose a framework that incorporates "quadratic computations," which essentially means looking at how different features interact with each other in a non-linear way, rather than just summing them up. These computations are designed to capture the intricate local interactions between neurons, both those that excite (stimulate) and those that suppress (inhibit) activity. The study's core finding is that these quadratic computations, along with precise coordination among their different elements, significantly improve how well scientists can predict neural responses and, crucially, how selective neurons become to specific features in natural images. Neural selectivity refers to a neuron's tendency to respond strongly to particular types of stimuli while ignoring others, which is essential for distinguishing objects. One key aspect highlighted by the researchers is the coordination between excitatory and suppressive features. These features are often arranged to represent "mutually exclusive hypotheses" about what the brain is seeing. For instance, in visual area V4, this might involve opposite motion directions or orientations that are perpendicular to each other. This finding resonates with earlier work on visual processing, particularly in area V2[2]. Previous research noted that while the first visual processing stage (V1) is relatively well understood, the complexity of V2's transformations makes it harder to grasp what it computes. However, it was observed that V2 neurons are selective for "multi-edge features," and that "excitatory edges have nearby suppressive edges with orthogonal orientations." This "cross-orientation suppression" was found to make neural responses more "sparse" (meaning neurons respond only to very specific inputs), which helps with position invariance across different scales. The current study provides a computational framework—the quadratic computations—that helps explain how this coordination between excitatory and suppressive elements leads to such precise feature selectivity. Another important principle identified in the new research is the delicate balance between excitatory and suppressive components. This balance needs to be maintained at similar levels as visual information progresses through different processing stages (from V1 to V2 to V4). This concept aligns closely with the idea of "normalization," a widely recognized "canonical neural computation"[3]. Normalization describes a process where a neuron's response is divided by a common factor, often representing the summed activity of a pool of surrounding neurons. This mechanism is thought to operate throughout the visual system and in many other sensory and brain regions, playing a role in processes like visual attention and multisensory integration. The current study's emphasis on maintaining a balance between excitation and suppression through quadratic computations offers a more detailed, mechanistic understanding of how normalization-like operations might be implemented at a fundamental computational level within the brain. Furthermore, the study observed a refinement of feature selectivity across processing stages. Earlier stages, like V1, tend to respond to a broader category of inputs, while later stages, such as V2 and V4, become increasingly specialized and selective for more complex features. This hierarchical processing, where information is progressively refined, is a hallmark of the visual system. Analyzing neural responses to natural stimuli, such as real-world images, presents significant challenges. Unlike simplified laboratory stimuli, natural images are "non-Gaussian," meaning their statistical properties are complex and do not follow a simple bell-curve distribution. Conventional methods for characterizing neural feature selectivity, like spike-triggered covariance (STC), often struggle with such complex, non-Gaussian data or are limited in the number of features they can identify due to the "curse of dimensionality" (where the amount of data needed to analyze increases exponentially with the number of features)[4]. To overcome these issues, previous work has proposed advanced "dimensionality reduction methods" that use information theory to find relevant patterns in complex data. While the new study doesn't explicitly state it used these specific methods, its use of "quadratic computations" to analyze responses to natural stimuli represents a sophisticated approach to tackle the very challenges highlighted by earlier research[4]—namely, understanding how neurons process the rich and complex information found in our everyday visual world. In essence, this work from The Salk Institute for Biological Studies, University of California - San Diego, and Centre National de la Recherche Scientifique describes how the brain utilizes multiple non-linear mechanisms, particularly through coordinated excitatory and suppressive interactions, to significantly enhance the selectivity of neural responses to the complex visual information we encounter daily. By providing a computational framework for these interactions, the study sheds new light on the fundamental processes underlying our ability to recognize objects.

HealthMental Health

References

Main Study

1) Computations that sustain neural feature selectivity across processing stages

Published 20th June, 2025

https://doi.org/10.1371/journal.pcbi.1013075

Related Studies

2) Cross-orientation suppression in visual area V2.

https://doi.org/10.1038/ncomms15739

3) Normalization as a canonical neural computation.

https://doi.org/10.1038/nrn3136

4) Second order dimensionality reduction using minimum and maximum mutual information models.

https://doi.org/10.1371/journal.pcbi.1002249


Related Articles

An unhandled error has occurred. Reload 🗙