Artificial neural networks (ANNs), particularly those employing deep learning, are valuable tools for modeling nervous system function. While successful in controlled settings (e.g., object recognition, predicting responses in visual cortex and retina), their performance in naturalistic settings is less clear. A major concern is the limited dynamic adaptation of typical ANNs to changing input conditions, unlike the nearly universal adaptation found in biological neurons and circuits. Sensory systems, such as the visual system, provide clear examples of adaptation's importance. Retinal adaptation, specifically within photoreceptors, dynamically adjusts neural responses to match prevailing light levels, which can vary dramatically in natural vision (from fractions of a second to hours). This adaptation is crucial because individual neurons have limited dynamic range, making adaptation essential for efficient encoding across a wide range of light intensities. This study investigates whether incorporating photoreceptor adaptation into ANNs enhances their ability to predict neural responses under varying conditions.

Existing research has demonstrated the success of ANNs, specifically CNNs, in modeling neural activity in the visual cortex and retina under relatively controlled conditions. Models such as Deep Retina have shown promise in predicting retinal ganglion cell (RGC) activity. However, a gap remains in understanding how these models perform under naturalistic conditions, where the statistics of sensory input vary significantly over time. The importance of adaptation in biological neural circuits is well-established, with numerous studies highlighting its role in various sensory systems. However, the integration of biophysical adaptation mechanisms into ANNs to improve their performance under dynamic conditions has been less explored. This research aims to bridge this gap by incorporating a biophysical model of photoreceptor adaptation into a CNN model of the retina.

The researchers developed a new deep learning model that incorporates a biophysical model of photoreceptor adaptation at the input stage of a conventional CNN. This biophysical model, previously validated, represents the phototransduction cascade with a set of differential equations that capture the dynamic feedback mechanisms influencing the conversion of light into electrical signals. The model has twelve parameters, some of which were trainable (allowing them to be learned during the training process) and some which were fixed based on experimental values derived from primate rods. The hybrid model, termed the photoreceptor-CNN, was trained and evaluated on data from primate and rat retinas. The experimental data included recordings of RGC spiking activity in response to various stimuli, including checkerboard noise and naturalistic movies presented at different mean luminance levels. The conventional CNN model, used as a control, was similar in architecture but lacked the biophysical photoreceptor layer. The models were trained using a two-stage process: first on checkerboard noise data, then fine-tuned on naturalistic movies. Model performance was assessed using the fraction of explainable variance (FEV), which quantifies the percentage of variance in the RGC's actual responses explained by the model. Multiple statistical tests (Wilcoxon signed-rank test and Wilcoxon rank-sum test) were used to compare model performance.

The photoreceptor-CNN model significantly outperformed the conventional CNN model in several key aspects: 1. **Predicting RGC responses to naturalistic movies:** The photoreceptor-CNN model achieved a substantially higher median FEV (49% ± 15%) compared to the conventional CNN (38% ± 8%) when predicting RGC responses to naturalistic movies containing dynamic local luminance variations. This improvement was statistically significant (p = 0.002). 2. **Generalization across light levels:** When trained at two light levels and tested at a third, the photoreceptor-CNN model demonstrated significantly better generalization than the conventional CNN. The conventional CNN performed poorly at the untested low light level (24% ± 15% FEV), while the photoreceptor-CNN model showed a much improved performance (54% ± 11% FEV) at the same low light level (p = 5 × 10⁻⁸). 3. **Capturing light-level-dependent changes in response kinetics:** The photoreceptor-CNN model, unlike the conventional CNN, accurately captured the light-level-dependent changes in RGC response kinetics (latency), showing shorter latencies at higher light levels, consistent with experimental observations. 4. **Generalization across photopic and scotopic light levels:** The photoreceptor-CNN model showed a remarkable ability to generalize across extremely different light levels (photopic and scotopic), while the conventional CNN failed to predict RGC responses at the scotopic light level when trained only at the photopic level. The FEV at the scotopic light level in the photoreceptor-CNN was 54% ± 8% versus -52% ± 9% in the conventional CNN model. The superior performance of the photoreceptor-CNN model was not attributable to increased model complexity; rather, it was directly linked to the inclusion of the biophysical photoreceptor adaptation layer, demonstrating the importance of incorporating biological realism into deep learning models.

The study's findings demonstrate the significant benefits of incorporating biophysically realistic adaptation mechanisms into ANNs for modeling neural activity. While ANNs are universal function approximators, the addition of the photoreceptor layer provides crucial inductive biases that improve model performance and generalizability, especially under dynamic and out-of-distribution conditions. The results highlight the limitations of conventional CNNs in capturing nonlinear adaptation dynamics present in biological systems. The success of the photoreceptor-CNN model in predicting RGC responses across a wide range of light levels and stimulus types underscores the value of integrating biophysical realism with trainable deep learning architectures. The improved performance in predicting both steady-state and dynamic changes highlights the importance of photoreceptor adaptation in the retina.

This research introduces a novel deep learning model of the retina incorporating a biophysical model of photoreceptor adaptation. This photoreceptor-CNN model significantly outperforms conventional CNNs in predicting RGC responses to dynamic stimuli and generalizing across various light levels. The superior performance is attributed to the inclusion of biophysically realistic adaptation mechanisms. This work establishes a valuable framework for integrating neural dynamics into ANNs, improving model accuracy and biological interpretability. Future research could focus on incorporating additional retinal adaptation mechanisms, exploring other neural architectures, and applying this approach to other brain regions.

The study's limitations include the focus on specific RGC types and the simplification of complex retinal circuitry. The model does not explicitly incorporate adaptation mechanisms present in bipolar and amacrine cells, or spike frequency adaptation in RGCs. While the Layer Norm layer helps with training, it might inadvertently mitigate some of the sensitivity changes normally handled by the photoreceptor layer. Also, the training data for naturalistic movies was limited compared to that for the checkerboard noise data, potentially affecting model performance. Finally, the study primarily focused on parasol RGCs; further investigation across other cell types is needed for broader generalization.