Learning Structure and Strength of CNN Filters for Small Sample Size Training

Abstract

Convolutional Neural Networks have provided state-ofthe-art results in several computer vision problems. However, due to a large number of parameters in CNNs, they require a large number of training samples which is a limiting factor for small sample size problems. To address this limitation, we propose SSF-CNN which focuses on learning the "structure" and "strength" of filters. The structure of the filter is initialized using a dictionary based filter learning algorithm and the strength of the filter is learned using the small sample training data. The architecture provides the flexibility of training with both small and large training databases, and yields good accuracies even with small size training data. The effectiveness of the algorithm is first demonstrated on MNIST, CIFAR10, and NORB databases, with varying number of training samples. The results show that SSF-CNN significantly reduces the number of parameters required for training while providing high accuracies on the test databases. On small sample size problems such as newborn face recognition and Omniglot, it yields stateof-the-art results. Specifically, on the IIITD Newborn Face Database, the results demonstrate improvement in rank-1 identification accuracy by at least 10%.

A Hybrid Geometric Morphometric Deep Learning Approach for Cut and Trampling Mark Classification

“…Nevertheless, Geometric Morphometric data has still proven to be a powerful type of input data for training, proving relatively fast to learn patterns from >1 min. Furthermore, considering the nature of the landmark data involved and its consequent transformation through GPA and PCA dimensionality reduction methods, this type of input data is less prone to issues presented by sample size as opposed to studies concerning, for example, Computer Vision and image processing-based techniques [54,[73][74][75][76], the latter requiring large amounts of parameters (usually in the millions), which are hard to learn from small datasets [76].…”

Meta-Transfer Learning for Few-Shot Learning

“…Algorithm 1 summarizes the training process of two main stages: large-scale DNN training (line 1-5) and metatransfer learning (line [6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22]. HT meta-batch re-sampling and continuous training phases are shown in lines 16-20, for which the failure classes are returned by Algorithm 2, see line 14.…”

A serialized classification method for pulmonary nodules based on lightweight cascaded convolutional neural network‐long short‐term memory

“…Oyallon et al 15 applied hybrid network to overcome the drawback of handcrafted filters. Rohit Keshari et al 16 proposed Structure and Strength Filtered CNN (SSF‐CNN). The second is to initialize the filter.…”

“…However, when these CNNs are trained, these architectures are not applicable to other datasets. Because these architectures learn filters in stack‐wise manner, once the network (filter) is trained, fine‐tuning filters on other databases is usually not allowed 16 . To address the gap, we built the Lightweight CNN‐LSTM network based on Rohit Keshari's method.…”