ProgressiveSpinalNet architecture for FC layers
In deeplearning models the FC (fully connected) layer has biggest important role for classification of the input based on the learned features from previous layers. The FC layers has highest numbers of parameters and fine-tuning these large numbers of parameters, consumes most of the computational resources, so in this paper it is aimed to reduce these large numbers of parameters significantly with improved performance. The motivation is inspired from SpinalNet and other biological architecture. The proposed architecture has a gradient highway between input to output layers and this solves the problem of diminishing gradient in deep networks. In this all the layers receives the input from previous layers as well as the CNN layer output and this way all layers contribute in decision making with last layer. This approach has improved classification performance over the SpinalNet architecture and has SOTA performance on many datasets such as Caltech101, KMNIST, QMNIST and EMNIST. The source code is available at https://github.com/praveenchopra/ProgressiveSpinalNet.
PDF AbstractDatasets
Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Benchmark |
---|---|---|---|---|---|---|
Fine-Grained Image Classification | Bird-225 | Pre trained wide-resnet-101 | Accuracy | 99.55 | # 2 | |
Fine-Grained Image Classification | Caltech-101 | Pre trained wide-resnet-101 | Accuracy | 97.76 | # 7 | |
Fine-Grained Image Classification | EMNIST-Digits | VGG-5 | Accuracy | 99.82 | # 1 | |
Fine-Grained Image Classification | EMNIST-Letters | VGG-5 | Accuracy | 95.86 | # 1 | |
Fine-Grained Image Classification | Fruits-360 | Pre trained wide-resnet-101 | Accuracy | 99.97 | # 1 | |
Fine-Grained Image Classification | Kuzushiji-MNIST | VGG-5 | Accuracy | 98.98 | # 1 | |
Fine-Grained Image Classification | MNIST | Vanilla FC layer only | Accuracy | 98.19 | # 1 | |
Fine-Grained Image Classification | QMNIST | VGG-5 | Accuracy | 99.6867 | # 1 | |
Fine-Grained Image Classification | STL-10 | Pre trained wide-resnet-101 | Accuracy | 98.18 | # 1 |