A Bidirectional Feedforward Neural Network Architecture Using the Discretized Neural Memory Ordinary Differential...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Feedforward Neural Networks (FNNs) with skip connections are pivotal in image recognition.
• Existing architectures face limitations in feature aggregation and adaptability.

Purpose of the Study:

• Introduce a novel Bidirectional FNN (BiFNN) architecture.
• Enhance feature aggregation and model flexibility in deep learning for image recognition.

Main Methods:

• Developed BiFNN, integrating a forward FNN with a non-parameter backward path using a discretized neural memory Ordinary Differential Equation ([Formula: see text]-net).
• Provided mathematical proof of convergence for the [Formula: see text]-net.
• Evaluated BiFNN on diverse image recognition datasets: Fashion-MNIST, SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet.

Main Results:

• BiFNNs demonstrated significant performance improvements over ConvMixer, ResNet, ResNeXt, and Vision Transformer.
• The architecture showed flexibility by accepting various FNNs as plugins with minimal parameter increase.
• Fine-tuning BiFNNs on Tiny-ImageNet and ImageNet-1K achieved performance comparable to embedded models.

Conclusions:

• BiFNNs represent a significant advancement in deep learning architectures for image recognition.
• The proposed model offers enhanced feature aggregation and adaptability, outperforming established models.
• BiFNNs provide a promising direction for future research in neural network design and image analysis.