Artificial intelligence-based computer vision systems are widely used for image recognition, classification, and reconstruction. However, balancing high image quality with low computational cost remains a major challenge In particular,detail loss during compression and blurring in reconstructed images limit current models. Biological neural networks, on the other hand, can process information in a distributed and dynamic manner through electrical signals. Mammalian neural circuits can process complex visual scenes in real time under severe energy constraints, utilizing local connectivity and distributed encoding in this process. These properties suggest that living neural networks could support both biological computing research and visual information processing in AI systems.
Two research studies described by The Biological Computing Co. examined the integration of biological neural networks into computer vision systems at various stages. The first study investigated whether living neural networks could improve classification performance by processing visual data before feeding them into artificial neural networks. The second study evaluated whether a software component derived from biological connection patterns observed in living neural networks could enhance the reconstruction quality of a pre-trained AI image model. The researchers demonstrated that a lightweight adapter—specifically, one that leverages the biological connection architecture—can help preserve image details without the need to retrain the entire existing model. Thus, the studies focused on the contributions that computational features derived from biological systems can make to both the representation of visual information and the reconstruction of images.
In the first study, researchers used living neural networks cultured on multi-electrode arrays (MEAs) as a computational platform. Visual images were converted into patterns of electrical stimulation applied via the electrodes, and the electrical responses of the neurons to these stimuli were recorded. The resulting neural activity was represented using metrics based on action potential frequencies (spike rates) over short time intervals and transferred to traditional artificial neural networks. The researchers evaluated the approach first on the MNIST dataset, which consists of handwritten digits, and then on the CIFAR dataset, which contains more complex natural images. In the second study, a lightweight adapter was developed using connectivity patterns observed in living neural networks, drawing on the principles of local connectivity and distributed encoding associated with mammalian visual systems. This adapter was designed to interact with the intermediate representations of a pre-trained Variational Autoencoder (VAE) model that transforms images into compressed representations. The adapter’s structure was derived from connection patterns observed in living neural networks cultured on MEA platforms containing 4,096 electrodes, as well as from recorded spatiotemporal neural response features. 5,000 images from five different video games were used to train the adapter; only the adapter’s parameters were trained, while the encoder and decoder weights of the base VAE model were kept fixed. In this approach, biological features derived from living neural networks are converted into a software component, and the system operates entirely in a computer environment (in silico) during the application phase.
In the first study, a comparison using the same artificial neural network architecture for the classification of MNIST images processed with biological neural activity yielded an average accuracy increase of 4.7%. This increase was calculated based on the results of 10 different network initializations during evaluations conducted after 10 training epochs. Furthermore, it was demonstrated that signals recorded from electrodes outside the directly stimulated region also carried information, and that using these signals, a classification accuracy of approximately 13% could be achieved; this rate is above the chance-level accuracy of approximately 10%. On the CIFAR dataset, models trained using biological neural responses achieved a final classification accuracy of approximately 43.5%, converging to this performance approximately three times faster than a comparison condition that did not use biological responses. The researchers noted that this convergence speed indicated a roughly threefold reduction in the computational and energy requirements needed to bring the classifier to its peak performance. The biological connectivity-based adapter developed in the second study achieved a 27% relative reduction in image reconstruction loss compared to the baseline VAE model. This improvement was achieved with only 2.36 million additional parameters—representing an increase of approximately 1%—compared to the base model’s 229 million parameters. Furthermore, it has been reported that adapters created using biological connection patterns derived from 11 different MEA-based neural networks outperformed non-biological fixed-reservoir models and came close to the results of fully trainable models.
When both studies are evaluated together, it becomes evident that biological neural networks can be integrated into artificial intelligence systems, not merely as biological models that serve as inspiration, but as computational components capable of directly contributing to information processing. In the first approach, living neural networks contribute to classification performance by physically participating in the processing step before visual data are fed into artificial neural networks. In the second approach, biological connection patterns observed in living neural networks are converted into a software-based adapter, thereby enhancing the reconstruction capacity of a pre-trained image model. However, there are some limitations at the current stage of the research. In the first study, the MNIST experiments were designed as small-scale experiments aimed at investigating the fundamental mechanism of the approach; furthermore, summarizing neural activity solely based on action potential frequency does not allow for a full assessment of the temporal structure of the signals and spike timing. In the second study, the current MEA approach is limited to the spatial scope of a single array; to apply it to wider fields of view, it is necessary to integrate connection patterns obtained from multiple MEAs and address the boundaries between these arrays. Furthermore, the exclusion of nonlinear interactions between simultaneously stimulated regions from the current model, as well as the variability of biological responses both spatially and across experiments, are areas that require further development. These studies provide an experimental foundation for how the information-processing mechanisms of biological neural systems can be leveraged in the development of computer vision and artificial intelligence algorithms.
Translated by: Yağmur Aktepe
Editor: Hazel Birgül
Reference:
The Biological Computing Co. (T.y.). Biological neural dynamics for computer vision. https://www.tbc.co/post/biological-neural-dynamics-for-computer-vision
The Biological Computing Co. (t.y.). Using neural connectivity patterns to improve image reconstruction quality. https://www.tbc.co/post/using-neural-connectivity-patterns-to-improve-image-reconstruction-quality