МНОГОСТОЛБЦОВЫЕ ГЛУБОКИЕ НЕЙРОННЫЕ СЕТИ ДЛЯ КЛАССИФИКАЦИИ ИЗОБРАЖЕНИЙ

Traditional computer vision and machine learning methods have historically struggled to match human performance on tasks such as handwritten digit recognition and traffic sign recognition. However, recent advancements in artificial neural network architectures have shown promise in overcoming these limitations. This article presents a novel approach that combines biologically plausible, wide, and deep neural networks to achieve remarkable results on various image classification benchmarks. The proposed neural network architecture leverages small receptive fields of convolutional winner-take-all neurons to achieve large network depth, resembling the organization of neural layers found in mammalian visual processing systems. Graphics processing units (GPUs) are utilized for fast training, enabling the training of deep neural networks in a fraction of the time previously required on traditional central processing units (CPUs). Experimental results demonstrate the effectiveness of the proposed approach, with the method achieving near-human performance on the MNIST handwriting benchmark and outperforming humans by a factor of two in traffic sign recognition. Furthermore, the architecture improves the state-of-the-art on a variety of common image classification benchmarks. Overall, this study highlights the potential of biologically inspired neural network architectures and GPU acceleration in advancing the field of computer vision and pattern recognition, offering insights into the development of more efficient and accurate image recognition systems.