05Neuroscience & artificial intelligence · Read 4 min

Do CNNs see like us? What the analogy helps us test

The layers of a convolutional neural network progressively learn edges, textures and shapes; this organization resembles part of the ventral visual pathway without reproducing human vision as a whole.

Article context

The comparison connects artificial feature hierarchies with several principles of human visual organization.

CONTEXT

This post revisits the neuro-inspired framework of my Grenoble thesis. A convolutional neural network (CNN) is an architecture that learns filters applied to images. I compare this organization with several mechanisms of human vision to identify principles that may improve model robustness.

SUMMARY

Convolutional networks learn a hierarchy from edges to objects, partially echoing the ventral pathway involved in identification. Human vision also uses a dorsal pathway related to space and action, recurrent loops and global-to-detail processing. The comparison provides concrete hypotheses for designing and evaluating new architectures.

Key diagram

One image, multiple processing pathways

VISUAL SCENE

CLASSIC CNNFeed-forward hierarchy

Edges→Textures→Objects

Partial similarity with the ventral pathway

HUMAN VISIONParallel pathways + feedback

VENTRALWHAT?Identify

DORSALWHERE / ACT?Locate & guide

Global · low frequencies→Details · high frequencies↶ feedback

A feed-forward CNN mainly follows a recognition hierarchy similar to the ventral pathway. Human vision adds a spatial/action pathway and global-to-detail processing with feedback.

A research-based explainer grounded in the sources listed below. Some mechanisms are simplified without changing the overall conclusions.

Product application05

Does your visual-search model remain reliable when texture, contrast or framing changes?

Complement average benchmarks with controlled tests of spatial frequency, shape, texture and domain shift. For image search or a product catalogue, these tests reveal visual shortcuts likely to fail in production.

What to measure in the experiment

Performance by transformation

Shape / texture sensitivity

Out-of-distribution generalization

Takeaways

Convolutional layers move from simple features to abstract representations, echoing part of the ventral visual hierarchy. The dorsal pathway, recurrent loops and coarse-to-fine processing complete the biological framework.

The analogy with human vision helps formulate experiments; it does not establish identical mechanisms.
Compare errors and the cues used, alongside average accuracy.
Retain a perception-inspired modification only when an ablation demonstrates its value in the protocol.

01 — The convolutional hierarchy

From contours to objects: a functional progression.

In a convolutional network, early filters learn local low-level properties such as edges and contrasts. Later layers combine these patterns into complex shapes and representations useful for categorization. Pooling, a spatial-reduction operation, progressively condenses information while final layers prioritize class identity.

This progression echoes the occipito-temporal, or ventral ‘what’, pathway involved in object identification. The comparison concerns a broad functional organization; mechanisms, training data and representations differ between the two systems.

02 — The parallel pathway

The dorsal pathway connects perception, space and action.

The dorsal pathway projects visual information toward parietal cortex and contributes to spatial location, action and visuomotor guidance. A feed-forward classifier passes information from layer to layer without feedback loops and mainly performs identification.

More complete visual artificial intelligence can combine recognition, space, motion, goals and active visual sampling — functions that are distributed and interactive in humans.

03 — The role of time

Coarse-to-fine processing organizes visual information over time.

Coarse-to-fine processing makes low spatial frequencies, which describe broad scene structure, available early. The magnocellular pathway may use them to form rapid hypotheses. High frequencies, carrying details and textures, then refine interpretation through the parvocellular pathway.

A standard feed-forward convolutional network does not explicitly model this temporal sequence or top-down influences from prior knowledge. Humans often use shape to classify images that place shape and texture in conflict, while classic networks may favor texture. This divergence sets a clear boundary for the analogy.

04 — Useful bio-inspiration

Turning biological principles into computational hypotheses.

Bio-inspiration supports testable hypotheses for model design: separate frequency channels, global-first processing, recurrence, predictive feedback, or cooperation between recognition and action.

My research explores this direction for adversarial robustness by training networks on frequency-filtered adversarial examples and transferring the resulting knowledge. The results, obtained with a constrained protocol and dataset, define a research direction to evaluate across other architectures and datasets.

05 — From concept to decision

Turn an analogy into a falsifiable experiment.

If a model appears texture-dependent, I would compare its predictions for the same objects with and without controlled texture changes. If global information seems useful, I would compare frequency bands under one shared protocol.

The criterion would be the resulting robustness, together with cost and errors, rather than an architecture’s resemblance to a brain diagram. Biological inspiration becomes useful when it leads to a measurable comparison.

Primary source

Grenoble research dissertation — Bio-inspired visual processing in CNNs

Synthesis of my research dissertation on adversarial robustness, complemented by the Fondements sensori-moteurs de l’attention chapter for the dorsal/ventral distinction.

Selected references

Goodale & Milner (1995)
Bar (2003)
Geirhos et al. (2019) ↗
Hegdé (2008)