In this paper, we present our system for the RSNA Intracranial Hemorrhage Detection challenge. The proposed system is based on a lightweight deep neural network architecture composed of a convolutional neural network (CNN) that takes as input individual CT slices, and a Long Short-Term Memory (LSTM) network that takes as input feature embeddings provided by the CNN. For efficient processing, we consider various feature selection methods to produce a subset of useful CNN features for the LSTM. Furthermore, we reduce the CT slices by a factor of 2x, allowing ourselves to train the model faster. Even if our model is designed to balance speed and accuracy, we report a weighted mean log loss of 0.04989 on the final test set, which places us in the top 30 ranking (2%) from a total of 1345 participants. Although our computing infrastructure does not allow it, processing CT slices at their original scale is likely to improve performance. In order to enable others to reproduce our results, we provide our code as open source at https://github.com/warchildmd/ihd. After the challenge, we conducted a subjective intracranial hemorrhage detection assessment by radiologists, indicating that the performance of our deep model is on par with that of doctors specialized in reading CT scans. Another contribution of our work is to integrate Grad-CAM visualizations in our system, providing useful explanations for its predictions. We therefore consider our system as a viable option when a fast diagnosis or a second opinion on intracranial hemorrhage detection are needed.
In this work, we provide a follow-up on the Moldavian versus Romanian Cross-Dialect Topic Identification (MRC) shared task of the VarDial 2019 Evaluation Campaign. The shared task included two sub-task types: one that consisted in discriminating between the Moldavian and the Romanian dialects and one that consisted in classifying documents by topic across the two dialects of Romanian. Participants achieved impressive scores, e.g. the top model for Moldavian versus Romanian dialect identification obtained a macro F1 score of 0.895. We conduct a subjective evaluation by human annotators, showing that humans attain much lower accuracy rates compared to machine learning (ML) models. Hence, it remains unclear why the methods proposed by participants attain such high accuracy rates. Our goal is to understand (i) why the proposed methods work so well (by visualizing the discriminative features) and (ii) to what extent these methods can keep their high accuracy levels, e.g. when we shorten the text samples to single sentences or when use tweets at inference time. A secondary goal of our work is to propose an improved ML model using ensemble learning. Our experiments show that ML models can accurately identify the dialects, even at the sentence level and across different domains (news articles versus tweets). We also analyze the most discriminative features of the best performing models, providing some explanations behind the decisions taken by these models. Interestingly, we learn new dialectal patterns previously unknown to us or to our human annotators. Furthermore, we conduct experiments showing that the machine learning performance on the MRC shared task can be improved through an ensemble based on classifier stacking.
With the growing complexity of deep learning methods adopted in practical applications, there is an increasing and stringent need to explain and interpret the decisions of such methods. In this work, we focus on explainable AI and propose a novel generic and model-agnostic framework for synthesizing input exemplars that maximize a desired response from a machine learning model. To this end, we use a generative model, which acts as a prior for generating data, and traverse its latent space using a novel evolutionary strategy with momentum updates. Our framework is generic because (i) it can employ any underlying generator, e.g. Variational Auto-Encoders (VAEs) or Generative Adversarial Networks (GANs), and (ii) it can be applied to any input data, e.g. images, text samples or tabular data. Since we use a zero-order optimization method, our framework is model-agnostic, in the sense that the machine learning model that we aim to explain is a black-box. We stress out that our novel framework does not require access or knowledge of the internal structure or the training data of the black-box model. We conduct experiments with two generative models, VAEs and GANs, and synthesize exemplars for various data formats, image, text and tabular, demonstrating that our framework is generic. We also employ our prototype synthetization framework on various black-box models, for which we only know the input and the output formats, showing that it is model-agnostic. Moreover, we compare our framework (available at https://github.com/antoniobarbalau/exemplar) with a model-dependent approach based on gradient descent, proving that our framework obtains equally-good exemplars in a shorter computational time.
The task of detecting whether a person wears a face mask from speech is useful in modelling speech in forensic investigations, communication between surgeons or people protecting themselves against infectious diseases such as COVID-19. In this paper, we propose a novel data augmentation approach for mask detection from speech. Our approach is based on (i) training Generative Adversarial Networks (GANs) with cycle-consistency loss to translate unpaired utterances between two classes (with mask and without mask), and on (ii) generating new training utterances using the cycle-consistent GANs, assigning opposite labels to each translated utterance. Original and translated utterances are converted into spectrograms which are provided as input to a set of ResNet neural networks with various depths. The networks are combined into an ensemble through a Support Vector Machines (SVM) classifier. With this system, we participated in the Mask Sub-Challenge (MSC) of the INTERSPEECH 2020 Computational Paralinguistics Challenge, surpassing the baseline proposed by the organizers by 2.8%. Our data augmentation technique provided a performance boost of 0.9% on the private test set. Furthermore, we show that our data augmentation approach yields better results than other baseline and state-of-the-art augmentation methods.
In order to classify linearly non-separable data, neurons are typically organized into multi-layer neural networks that are equipped with at least one hidden layer. Inspired by some recent discoveries in neuroscience, we propose a new neuron model along with a novel activation function enabling learning of non-linear decision boundaries using a single neuron. We show that a standard neuron followed by the novel apical dendrite activation (ADA) can learn the XOR logical function with 100% accuracy. Furthermore, we conduct experiments on three benchmark data sets from computer vision and natural language processing, i.e. Fashion-MNIST, UTKFace and MOROCO, showing that the ADA and the leaky ADA functions provide superior results to Rectified Liner Units (ReLU) and leaky ReLU, for various neural network architectures, e.g. 1-hidden layer or 2-hidden layers multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs) such as LeNet, VGG, ResNet and Character-level CNN. We also obtain further improvements when we change the standard model of the neuron with our pyramidal neuron with apical dendrite activations (PyNADA).
CT scanners that are commonly-used in hospitals nowadays produce low-resolution images, up to 512 pixels in size. One pixel in the image corresponds to a one millimeter piece of tissue. In order to accurately segment tumors and make treatment plans, doctors need CT scans of higher resolution. The same problem appears in MRI. In this paper, we propose an approach for the single-image super-resolution of 3D CT or MRI scans. Our method is based on deep convolutional neural networks (CNNs) composed of 10 convolutional layers and an intermediate upscaling layer that is placed after the first 6 convolutional layers. Our first CNN, which increases the resolution on two axes (width and height), is followed by a second CNN, which increases the resolution on the third axis (depth). Different from other methods, we compute the loss with respect to the ground-truth high-resolution output right after the upscaling layer, in addition to computing the loss after the last convolutional layer. The intermediate loss forces our network to produce a better output, closer to the ground-truth. A widely-used approach to obtain sharp results is to add Gaussian blur using a fixed standard deviation. In order to avoid overfitting to a fixed standard deviation, we apply Gaussian smoothing with various standard deviations, unlike other approaches. We evaluate our method in the context of 2D and 3D super-resolution of CT and MRI scans from two databases, comparing it to relevant related works from the literature and baselines based on various interpolation schemes, using 2x and 4x scaling factors. The empirical results show that our approach attains superior results to all other methods. Moreover, our human annotation study reveals that both doctors and regular annotators chose our method in favor of Lanczos interpolation in 97.55% cases for 2x upscaling factor and in 96.69% cases for 4x upscaling factor.
In this paper, we propose a deep learning approach for smartphone user identification based on analyzing motion signals recorded by the accelerometer and the gyroscope, during a single tap gesture performed by the user on the screen. We transform the discrete 3-axis signals from the motion sensors into a gray-scale image representation which is provided as input to a convolutional neural network (CNN) that is pre-trained for multi-class user classification. In the pre-training stage, we benefit from different users and multiple samples per user. After pre-training, we use our CNN as feature extractor, generating an embedding associated to each single tap on the screen. The resulting embeddings are used to train a Support Vector Machines (SVM) model in a few-shot user identification setting, i.e. requiring only 20 taps on the screen during the registration phase. We compare our identification system based on CNN features with two baseline systems, one that employs handcrafted features and another that employs recurrent neural network (RNN) features. All systems are based on the same classifier, namely SVM. To pre-train the CNN and the RNN models for multi-class user classification, we use a different set of users than the set used for few-shot user identification, ensuring a realistic scenario. The empirical results demonstrate that our CNN model yields a top accuracy of 89.75% in multi-class user classification and a top accuracy of 96.72% in few-shot user identification. In conclusion, we believe that our system is ready for practical use, having a better generalization capacity than both baselines.
The goal of query performance prediction (QPP) is to automatically estimate the effectiveness of a search result for any given query, without relevance judgements. Post-retrieval features have been shown to be more effective for this task while being more expensive to compute than pre-retrieval features. Combining multiple post-retrieval features is even more effective, but state-of-the-art QPP methods are impossible to interpret because of the black-box nature of the employed machine learning models. However, interpretation is useful for understanding the predictive model and providing more answers about its behavior. Moreover, combining many post-retrieval features is not applicable to real-world cases, since the query running time is of utter importance. In this paper, we investigate a new framework for feature selection in which the trained model explains well the prediction. We introduce a step-wise (forward and backward) model selection approach where different subsets of query features are used to fit different models from which the system selects the best one. We evaluate our approach on four TREC collections using standard QPP features. We also develop two QPP features to address the issue of query-drift in the query feedback setting. We found that: (1) our model based on a limited number of selected features is as good as more complex models for QPP and better than non-selective models; (2) our model is more efficient than complex models during inference time since it requires fewer features; (3) the predictive model is readable and understandable; and (4) one of our new QPP features is consistently selected across different collections, proving its usefulness.
Training (source) domain bias affects state-of-the-art object detectors, such as Faster R-CNN, when applied to new (target) domains. To alleviate this problem, researchers proposed various domain adaptation methods to improve object detection results in the cross-domain setting, e.g. by translating images with ground-truth labels from the source domain to the target domain using Cycle-GAN or by applying self-paced learning. On top of combining Cycle-GAN transformations and self-paced learning, in this paper, we propose a novel self-paced algorithm that learns from easy to hard. To estimate the difficulty of each image, we use the number of detected objects divided by their average size. Our method is simple and effective, without any overhead during inference. It uses only pseudo-labels for samples taken from the target domain, i.e. the domain adaptation is unsupervised. We conduct experiments on two cross-domain benchmarks, showing better results than the state of the art. We also perform an ablation study demonstrating the utility of each component in our framework.