Convolutional Neural Networks (CNNs) are known to be brittle under various image transformations, including rotations, scalings, and changes of lighting conditions. We observe that the features of a transformed image are drastically different from the ones of the original image. To make CNNs more invariant to transformations, we propose "Feature Lenses", a set of ad-hoc modules that can be easily plugged into a trained model (referred to as the "host model"). Each individual lens reconstructs the original features given the features of a transformed image under a particular transformation. These lenses jointly counteract feature distortions caused by various transformations, thus making the host model more robust without retraining. By only updating lenses, the host model is freed from iterative updating when facing new transformations absent in the training data; as feature semantics are preserved, downstream applications, such as classifiers and detectors, automatically gain robustness without retraining. Lenses are trained in a self-supervised fashion with no annotations, by minimizing a novel "Top-K Activation Contrast Loss" between lens-transformed features and original features. Evaluated on ImageNet, MNIST-rot, and CIFAR-10, Feature Lenses show clear advantages over baseline methods.
Deep learning for medical image classification faces three major challenges: 1) the number of annotated medical images for training are usually small; 2) regions of interest (ROIs) are relatively small with unclear boundaries in the whole medical images, and may appear in arbitrary positions across the x,y (and also z in 3D images) dimensions. However often only labels of the whole images are annotated, and localized ROIs are unavailable; and 3) ROIs in medical images often appear in varying sizes (scales). We approach these three challenges with a Multi-Instance Multi-Scale (MIMS) CNN: 1) We propose a multi-scale convolutional layer, which extracts patterns of different receptive fields with a shared set of convolutional kernels, so that scale-invariant patterns are captured by this compact set of kernels. As this layer contains only a small number of parameters, training on small datasets becomes feasible; 2) We propose a "top-k pooling" to aggregate the feature maps in varying scales from multiple spatial dimensions, allowing the model to be trained using weak annotations within the multiple instance learning (MIL) framework. Our method is shown to perform well on three classification tasks involving two 3D and two 2D medical image datasets.
Deep Neural Networks (DNN) have been successful in en- hancing noisy speech signals. Enhancement is achieved by learning a nonlinear mapping function from the features of the corrupted speech signal to that of the reference clean speech signal. The quality of predicted features can be improved by providing additional side channel information that is robust to noise, such as visual cues. In this paper we propose a novel deep learning model inspired by insights from human audio visual perception. In the proposed unified hybrid architecture, features from a Convolution Neural Network (CNN) that processes the visual cues and features from a fully connected DNN that processes the audio signal are integrated using a Bidirectional Long Short-Term Memory (BiLSTM) network. The parameters of the hybrid model are jointly learned using backpropagation. We compare the quality of enhanced speech from the hybrid models with those from traditional DNN and BiLSTM models.
Most existing learning to hash methods assume that there are sufficient data, either labeled or unlabeled, on the domain of interest (i.e., the target domain) for training. However, this assumption cannot be satisfied in some real-world applications. To address this data sparsity issue in hashing, inspired by transfer learning, we propose a new framework named Transfer Hashing with Privileged Information (THPI). Specifically, we extend the standard learning to hash method, Iterative Quantization (ITQ), in a transfer learning manner, namely ITQ+. In ITQ+, a new slack function is learned from auxiliary data to approximate the quantization error in ITQ. We developed an alternating optimization approach to solve the resultant optimization problem for ITQ+. We further extend ITQ+ to LapITQ+ by utilizing the geometry structure among the auxiliary data for learning more precise binary codes in the target domain. Extensive experiments on several benchmark datasets verify the effectiveness of our proposed approaches through comparisons with several state-of-the-art baselines.
The Support Vector Machine using Privileged Information (SVM+) has been proposed to train a classifier to utilize the additional privileged information that is only available in the training phase but not available in the test phase. In this work, we propose an efficient solution for SVM+ by simply utilizing the squared hinge loss instead of the hinge loss as in the existing SVM+ formulation, which interestingly leads to a dual form with less variables and in the same form with the dual of the standard SVM. The proposed algorithm is utilized to leverage the additional web knowledge that is only available during training for the image categorization tasks. The extensive experimental results on both Caltech101 andWebQueries datasets show that our proposed method can achieve a factor of up to hundred times speedup with the comparable accuracy when compared with the existing SVM+ method.