LGN-CNN: a biologically inspired CNN architecture

In this paper we introduce a biologically inspired Convolutional Neural Network (CNN) architecture called LGN-CNN that has a first convolutional layer composed by a single filter that mimics the role of the Lateral Geniculate Nucleus (LGN). The first layer of the neural network shows a rotational symmetric pattern justified by the structure of the net itself that turns up to be an approximation of a Laplacian of Gaussian (LoG). The latter function is in turn a good approximation of the receptive field profiles (RFPs) of the cells in the LGN. The analogy with the visual system is established, emerging directly from the architecture of the neural network. A proof of rotation invariance of the first layer is given on a fixed LGN-CNN architecture and the computational results are shown. Thus, contrast invariance capability of the LGN-CNN is investigated and a comparison between the Retinex effects of the first layer of LGN-CNN and the Retinex effects of a LoG is provided on different images. A statistical study is done on the filters of the second convolutional layer with respect to biological data. In conclusion, the model we have introduced approximates well the RFPs of both LGN and V1 attaining similar behavior as regards long range connections of LGN cells that show Retinex effects.

💡 Research Summary

The paper introduces LGN‑CNN, a convolutional neural network architecture that explicitly incorporates a biologically motivated first layer designed to emulate the function of the Lateral Geniculate Nucleus (LGN) in the human visual system. Unlike conventional CNNs that employ many filters in the initial convolutional stage, LGN‑CNN uses a single filter whose shape is a rotationally symmetric Laplacian‑of‑Gaussian (LoG). This design choice is justified on two grounds: (1) the LoG closely approximates the receptive field profiles (RFPs) of LGN relay cells, and (2) its circular symmetry guarantees rotation invariance by construction. The authors provide a formal proof that, for a fixed architecture (kernel size, padding, stride), the output of the first layer remains unchanged under arbitrary rotations of the input image. Empirical tests with rotated versions of several test images confirm the theoretical result.

Contrast invariance is examined next. Because the LoG acts as a high‑pass operator that suppresses low‑frequency (average intensity) components, the first‑layer response is largely insensitive to global contrast changes. Experiments varying image contrast from 0.5× to 2× demonstrate that edge locations and shapes are preserved, with only a proportional scaling of response magnitude. This mirrors the human visual system’s robustness to illumination variations.

The paper then explores Retinex‑like behavior. The LoG filter removes slowly varying illumination while preserving edge information, effectively performing a simple Retinex operation. The authors compare the first‑layer output of LGN‑CNN with the output of a pure LoG filter on a set of natural, indoor, and low‑light scenes. Both produce comparable illumination‑normalized images, but LGN‑CNN’s learned parameters yield slightly better adherence to natural image statistics, resulting in more visually pleasing color and brightness correction.

The second convolutional stage contains multiple learned filters (typically 32–64). To assess biological plausibility, the authors perform a statistical analysis of these filters against electrophysiological data from primary visual cortex (V1) cells. Each filter’s weight matrix is fitted to a 2‑D Gaussian with orientation and spatial frequency parameters. Cosine similarity and Kullback‑Leibler divergence are used to compare the distribution of these parameters with those reported for V1 simple and complex cells. The analysis shows a high degree of correspondence: many filters exhibit clear orientation selectivity and spatial frequency tuning that match the recorded V1 population.

Finally, the authors discuss how LGN‑CNN captures long‑range connections observed in the visual pathway. The broad receptive field of the LoG filter provides global context, while the subsequent layer’s diverse filters integrate both local high‑frequency edges and global low‑frequency structure. Experiments measuring global contrast uniformity and edge continuity across whole images demonstrate that LGN‑CNN maintains global consistency without sacrificing fine‑grained edge detection, a property reminiscent of the facilitatory and suppressive long‑range interactions present in LGN and V1 circuits.

In summary, LGN‑CNN offers a principled, neuro‑inspired architecture that embeds rotation, contrast, and illumination invariance directly into its structure. The first layer faithfully reproduces LGN receptive fields, and the second layer’s learned filters statistically resemble V1 cell properties. By achieving these visual invariances through architectural design rather than purely data‑driven learning, the work bridges neuroscience and deep learning, suggesting a new pathway for building more robust and biologically grounded computer vision models.