Bio-Inspired Perception Encoder

• Perception encoder is a neural network module that extracts functionally meaningful representations from sensory data via bio-inspired operations.
• It employs self-supervised tasks like autoencoding, denoising, and deblurring to learn efficient, sparse, and perceptually aligned features.
• Robust human perceptual alignment is achieved at a V1-like layer, indicating an optimal balance of regularization for emergent perceptual metrics.

A perception encoder is a neural network module or system designed to extract functionally meaningful, perceptual representations from raw sensory inputs—typically images or videos—such that the internal encoding aligns closely with human perceptual similarity and supports efficient downstream computation and robust metric learning. Recent research has demonstrated that, under certain bio-inspired and self-supervised paradigms, the architecture and learning objectives of a perception encoder can lead to emergent properties mirroring human visual perception, even without relying on explicit perceptual supervision.

1. Bio-Inspired Architecture and Efficient Coding Principles

Perception encoder design has drawn substantial inspiration from biological vision, particularly the structural and computational organization of the early mammalian visual system (retina and V1 cortex). In this context, the encoder applies cascades of operations analogous to those observed in neurophysiology:

• Localized, oriented convolutional filters to emulate receptive fields (akin to those of the retina and simple cells in V1).
• Spatial pooling to capture local invariance and downsample signal representations.
• Divisive normalization to model neural gain control and adaptation, consistent with findings from Schwartz and Simoncelli (2001) and Laparra et al. (2012).

The architecture typically consists of a "retinal" stage for preprocessing followed by a V1-like stage implementing orientation selectivity and normalization. The decoder is constructed as an approximate inverse of the encoder, substituting pooling with upsampling and divisive normalization with its functional inverse (multiplicative scaling). Encoder and decoder parameters are learned jointly in an end-to-end framework.

2. Self-Supervised Optimization Tasks

A defining trait of modern perception encoders is their training via self-supervised objectives on large natural image corpora, eschewing the need for labelled perceptual ground-truth. The canonical approach involves:

• Autoencoding: Minimizing mean squared error (MSE) between the input and reconstructed images.
• Denoising: Reconstructing clean images from those corrupted with Gaussian noise ($\sigma$ parameterized).
• Deblurring: Restoring original images from blurred versions (Gaussian blur with specified SD).
• Sparsity regularization: Imposing an $L_1$ penalty on encoder activations, parameterized by $\lambda$ to promote efficient, sparse representations.

The loss functions are aggregated as: $$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{reconstruction}} + \mathcal{L}_{\text{denoising}} + \mathcal{L}_{\text{deblurring}} + \mathcal{L}_{\text{sparsity}},$$ where

$$\mathcal{L}_\text{sparsity} = \lambda \cdot \frac{1}{N}\sum_{i=1}^N |h_i|.$$

These tasks capitalize on the assumption that the efficient coding of natural image statistics, as posited by Barlow (1961) and Olshausen & Field (1996), shapes perceptual representations in the early visual cortex.

3. Alignment of Encoder Representations with Human Perceptual Judgments

A central empirical finding is that internal feature activations at specific intermediate (V1-like) layers of the perception encoder closely align with human perceptual similarity judgments. Alignment is quantitatively evaluated via the correlation (specifically, Spearman's $\rho$) between distances in model feature space and subjective human Mean Opinion Score (MOS) ratings on standard image quality assessment datasets (e.g., TID2013):

$$\rho = \text{Spearman}(d_\text{model}(\mathbf{x}_1, \mathbf{x}_2), \text{MOS}(\mathbf{x}_1, \mathbf{x}_2))$$

where $d_\text{model}$ is the $L_2$ norm between activations at the examined layer.

The maximal correspondence to human judgments does not occur at the network’s output or deepest layer, but at the V1-analogous encoder stage. This correlation peaks at moderate levels of noise, blurring, and sparsity regularization—exceeding both high and low extremes—demonstrating the existence of an optimal regime for perceptual metric emergence.

4. Implications for the Nature and Learning of Perceptual Metrics

The perception encoder phenomenon elucidates several theoretical and practical implications:

• Emergent Perceptual Metrics: Self-supervision and biologically-inspired inductive biases are sufficient for a model to develop a functionally relevant perceptual metric, even when explicit perceptual data is withheld during training.
• Optimal Regularization Regime: Perceptual alignment is non-monotonic with respect to noise, blur, or sparsity; there exists an optimal intermediate setting, suggesting the visual system is tuned not for maximal regularization but for an optimal level corresponding to natural ecological distortions.
• Task Generalization: The geometry of learned representations at the encoder aligns with human perception robustly across tasks and datasets, indicating broad applicability without reliance on task-specific supervision.

These insights corroborate the Efficient Coding Hypothesis at both computational and representational levels and support the hypothesis that the primate visual system may be evolutionarily tuned for optimal removal of moderate levels of noise and distortion encountered in natural environments.

5. Summary Table: Perception Encoder Findings

Aspect	Implementation/Result
Architecture	Retinal + V1-inspired encoder: oriented filters, pooling, divisive normalization
Training Tasks	Autoencoding, Denoising (varied $\sigma$), Deblurring (varied $\sigma$), Sparsity ($\lambda$)
Perceptual Metric	Spearman correlation between feature distances and human MOS
Biological Inspiration	Efficient coding, orientation selectivity, gain control (retina-V1 cortex mimicry)
Key Result	Encoder (V1-like) layer achieves highest human alignment at moderate regularization
Generalization	Robust across distortions/tasks, matches human perceptual similarity without supervision

6. Future Directions and General Significance

The concept of the perception encoder, validated in a bio-inspired, self-supervised setting, provides a new design principle for constructing perceptual metrics and encoders in both computational neuroscience and artificial vision. Models designed along these lines could obviate the need for tedious perceptual data annotation and enable strong zero-shot or few-shot generalization in perceptual tasks—suggesting a pathway toward interpretable, robust, and ecologically-informed computer vision systems.

The broader significance lies in demonstrating that internal representational geometry aligned with human perception can naturally emerge from efficient coding constraints and appropriate self-supervised objectives—opening avenues for the principled development of perception-centered AI models and further investigation into the computational principles of biological perception (Hernández-Cámara et al., 14 Aug 2025).