[Paper Review] A neuromorphic model of the insect visual system for natural image processing

Adam Hines, Karin Nordström+1 more|arXiv (Cornell University)|Feb 6, 2026

Neurobiology and Insect Physiology Research0 citations

TL;DR

The paper presents a bio-inspired vision model (ANN and SNN) that converts dense images into sparse Kenyon cell codes via a self-supervised contrastive objective, and demonstrates generalization across flower recognition and VPR tasks.

ABSTRACT

Insect vision supports complex behaviors including associative learning, navigation, and object detection, and has long motivated computational models for understanding biological visual processing. However, many contemporary models prioritize task performance while neglecting biologically grounded processing pathways. Here, we introduce a bio-inspired vision model that captures principles of the insect visual system to transform dense visual input into sparse, discriminative codes. The model is trained using a fully self-supervised contrastive objective, enabling representation learning without labeled data and supporting reuse across tasks without reliance on domain-specific classifiers. We evaluated the resulting representations on flower recognition tasks and natural image benchmarks. The model consistently produced reliable sparse codes that distinguish visually similar inputs. To support different modelling and deployment uses, we have implemented the model as both an artificial neural network and a spiking neural network. In a simulated localization setting, our approach outperformed a simple image downsampling comparison baseline, highlighting the functional benefit of incorporating neuromorphic visual processing pathways. Collectively, these results advance insect computational modelling by providing a generalizable bio-inspired vision model capable of sparse computation across diverse tasks.

Motivation & Objective

• Motivate the use of insect visual processing principles to produce sparse, discriminative representations.
• Develop a generalizable vision module implemented as both ANN and SNN variants.
• Train with self-supervised contrastive learning to enable task reuse without domain-specific classifiers.
• Evaluate on flower recognition and natural image benchmarks to assess sparsity and discrimination.

Proposed method

• Model the insect visual pathway as a CNN with Activation–Normalization blocks and two normalization strategies to mimic local and global homeostasis.
• Generate a sparse Kenyon cell (KC) 1,024-dimensional code from projection neuron outputs via a sparsity mechanism including masking and adaptive k-Winner-Take-All.
• Train the vision module using NT-Xent contrastive loss in a SimCLR-like setup with two augmentations to maximize KC representation agreement.
• Implement a spiking neural network (visionSNN) variant by replacing Leaky ReLU with Leaky Integrate-and-Fire neurons and using temporal spiking inputs.
• Process inputs through retinal, lamina, medulla, and lobula-like stages to produce VPNs (asot, aiot, lot) feeding the KC layer.
• Use a two-stage normalization: LocalResponseNorm for local inhibition and GroupNorm for global feature normalization to ensure diversity and contrast.

Experimental results

Research questions

• RQ1 Can a bio-inspired, insect-vision-based module learn useful sparse representations from unlabeled natural images?
• RQ2 Do artificial and spiking implementations of the insect-inspired vision model yield discriminative, sparse codes suitable for downstream tasks?
• RQ3 How does temporal accumulation and sparse coding affect flower recognition and visual place recognition performance?
• RQ4 Is the KC representation sufficiently discriminative to support linear classifiers on natural image datasets?
• RQ5 How do the ANN and SNN variants compare in terms of sparsity, selectivity, and task performance?

Key findings

• The model produces sparse KC representations of about 1,024 dimensions with approximately 50 active neurons per input.
• In Experiment 2 on 17CFD, vision achieved 40% to 75% classification accuracy within three training epochs, peaking at 76.6%.
• The spiking vision (visionSNN) yields sparse KC codes and strong intra-class similarity but does not achieve linear classification performance as well as the ANN version (highest ~40% accuracy).
• Spiking KC outputs show a large drop in active neuron count from an average of 353 to 33 when trained on 17CFD, with KC neuron selectivity increasing from 14% to 76%.
• A KC Top-K selectivity and reduced overlap between classes were observed, indicating robust sparse representations despite natural image variability.
• In Visual Place Recognition tasks, cosine similarity-based matching aligned with ground truth, with Recall@K evaluated up to K=25.