Terahertz computed tomography (THz CT) has drawn significant attention because of its unique capability to bring multi-dimensional object information from invisible to visible. However, current physics-model-based THz CT modalities present low data use efficiency on time-resolved THz signals and low model fusion extensibility, limiting their application fields' practical use. In this paper, we propose a supervised THz deep learning computed tomography (THz DL-CT) framework based on time-domain information. THz DL-CT restores superior THz tomographic images of 3D objects by extracting features from spatio-temporal THz signals without any prior material information. Compared with conventional and machine learning based methods, THz DL-CT delivers at least 50.2%, and 52.6% superior in root mean square error (RMSE) and structural similarity index (SSIM), respectively. Additionally, we have experimentally demonstrated that the pretrained THz DL-CT model can generalize to reconstruct multi-material systems with no prerequisite information. THz CT through the DL data fusion approach provides a new pathway for non-invasive functional imaging in object investigation.
For deep learning, size is power. Massive neural nets trained on broad data for a spectrum of tasks are at the forefront of artificial intelligence. These large pre-trained models or Jacks of All Trades (JATs), when fine-tuned for downstream tasks, are gaining importance in driving deep learning advancements. However, environments with tight resource constraints, changing objectives and intentions, or varied task requirements, could limit the real-world utility of a singular JAT. Hence, in tandem with current trends towards building increasingly large JATs, this paper conducts an initial exploration into concepts underlying the creation of a diverse set of compact machine learning model sets. Composed of many smaller and specialized models, the Set of Sets is formulated to simultaneously fulfil many task settings and environmental conditions. A means to arrive at such a set tractably in one pass of a neuroevolutionary multitasking algorithm is presented for the first time, bringing us closer to models that are collectively Masters of All Trades.
This paper investigates the problem of forecasting multivariate aggregated human mobility while preserving the privacy of the individuals concerned. Differential privacy, a state-of-the-art formal notion, has been used as the privacy guarantee in two different and independent steps when training deep learning models. On one hand, we considered \textit{gradient perturbation}, which uses the differentially private stochastic gradient descent algorithm to guarantee the privacy of each time series sample in the learning stage. On the other hand, we considered \textit{input perturbation}, which adds differential privacy guarantees in each sample of the series before applying any learning. We compared four state-of-the-art recurrent neural networks: Long Short-Term Memory, Gated Recurrent Unit, and their Bidirectional architectures, i.e., Bidirectional-LSTM and Bidirectional-GRU. Extensive experiments were conducted with a real-world multivariate mobility dataset, which we published openly along with this paper. As shown in the results, differentially private deep learning models trained under gradient or input perturbation achieve nearly the same performance as non-private deep learning models, with loss in performance varying between $0.57\%$ to $2.8\%$. The contribution of this paper is significant for those involved in urban planning and decision-making, providing a solution to the human mobility multivariate forecast problem through differentially private deep learning models.
Accurate uncertainty quantification is a major challenge in deep learning, as neural networks can make overconfident errors and assign high confidence predictions to out-of-distribution (OOD) inputs. The most popular approaches to estimate predictive uncertainty in deep learning are methods that combine predictions from multiple neural networks, such as Bayesian neural networks (BNNs) and deep ensembles. However their practicality in real-time, industrial-scale applications are limited due to the high memory and computational cost. Furthermore, ensembles and BNNs do not necessarily fix all the issues with the underlying member networks. In this work, we study principled approaches to improve uncertainty property of a single network, based on a single, deterministic representation. By formalizing the uncertainty quantification as a minimax learning problem, we first identify distance awareness, i.e., the model's ability to quantify the distance of a testing example from the training data, as a necessary condition for a DNN to achieve high-quality (i.e., minimax optimal) uncertainty estimation. We then propose Spectral-normalized Neural Gaussian Process (SNGP), a simple method that improves the distance-awareness ability of modern DNNs with two simple changes: (1) applying spectral normalization to hidden weights to enforce bi-Lipschitz smoothness in representations and (2) replacing the last output layer with a Gaussian process layer. On a suite of vision and language understanding benchmarks, SNGP outperforms other single-model approaches in prediction, calibration and out-of-domain detection. Furthermore, SNGP provides complementary benefits to popular techniques such as deep ensembles and data augmentation, making it a simple and scalable building block for probabilistic deep learning. Code is open-sourced at https://github.com/google/uncertainty-baselines
Machine learned tasks on seismic data are often trained sequentially and separately, even though they utilize the same features (i.e. geometrical) of the data. We present StorSeismic, as a framework for seismic data processing, which consists of neural network pre-training and fine-tuning procedures. We, specifically, utilize a neural network as a preprocessing model to store seismic data features of a particular dataset for any downstream tasks. After pre-training, the resulting model can be utilized later, through a fine-tuning procedure, to perform tasks using limited additional training. Used often in Natural Language Processing (NLP) and lately in vision tasks, BERT (Bidirectional Encoder Representations from Transformer), a form of a Transformer model, provides an optimal platform for this framework. The attention mechanism of BERT, applied here on a sequence of traces within the shot gather, is able to capture and store key geometrical features of the seismic data. We pre-train StorSeismic on field data, along with synthetically generated ones, in the self-supervised step. Then, we use the labeled synthetic data to fine-tune the pre-trained network in a supervised fashion to perform various seismic processing tasks, like denoising, velocity estimation, first arrival picking, and NMO. Finally, the fine-tuned model is used to obtain satisfactory inference results on the field data.
The success of deep learning (DL) fostered the creation of unifying frameworks such as tensorflow or pytorch as much as it was driven by their creation in return. Having common building blocks facilitates the exchange of, e.g., models or concepts and makes developments easier replicable. Nonetheless, robust and reliable evaluation and assessment of DL models has often proven challenging. This is at odds with their increasing safety relevance, which recently culminated in the field of "trustworthy ML". We believe that, among others, further unification of evaluation and safeguarding methodologies in terms of toolkits, i.e., small and specialized framework derivatives, might positively impact problems of trustworthiness as well as reproducibility. To this end, we present the first survey on toolkits for uncertainty estimation (UE) in DL, as UE forms a cornerstone in assessing model reliability. We investigate 11 toolkits with respect to modeling and evaluation capabilities, providing an in-depth comparison for the three most promising ones, namely Pyro, Tensorflow Probability, and Uncertainty Quantification 360. While the first two provide a large degree of flexibility and seamless integration into their respective framework, the last one has the larger methodological scope.
Deep learning has shown great promise in diverse areas of computer vision, such as image classification, object detection and semantic segmentation, among many others. However, as it has been repeatedly demonstrated, deep learning methods trained on a dataset do not generalize well to datasets from other domains or even to similar datasets, due to data distribution shifts. In this work, we propose the use of a meta-learning based few-shot learning approach to alleviate these problems. In order to demonstrate its efficacy, we use two datasets of kidney stones samples acquired with different endoscopes and different acquisition conditions. The results show how such methods are indeed capable of handling domain-shifts by attaining an accuracy of 74.38% and 88.52% in the 5-way 5-shot and 5-way 20-shot settings respectively. Instead, in the same dataset, traditional Deep Learning (DL) methods attain only an accuracy of 45%.
In recent years, there has been an increasing interest in exploiting logically specified background knowledge in order to obtain neural models (i) with a better performance, (ii) able to learn from less data, and/or (iii) guaranteed to be compliant with the background knowledge itself, e.g., for safety-critical applications. In this survey, we retrace such works and categorize them based on (i) the logical language that they use to express the background knowledge and (ii) the goals that they achieve.
The ever-growing demand for further advances in artificial intelligence motivated research on unconventional computation based on analog physical devices. While such computation devices mimic brain-inspired analog information processing, learning procedures still relies on methods optimized for digital processing such as backpropagation. Here, we present physical deep learning by extending a biologically plausible training algorithm called direct feedback alignment. As the proposed method is based on random projection with arbitrary nonlinear activation, we can train a physical neural network without knowledge about the physical system. In addition, we can emulate and accelerate the computation for this training on a simple and scalable physical system. We demonstrate the proof-of-concept using a hierarchically connected optoelectronic recurrent neural network called deep reservoir computer. By constructing an FPGA-assisted optoelectronic benchtop, we confirmed the potential for accelerated computation with competitive performance on benchmarks. Our results provide practical solutions for the training and acceleration of neuromorphic computation.
Uncertainty estimation bears the potential to make deep learning (DL) systems more reliable. Standard techniques for uncertainty estimation, however, come along with specific combinations of strengths and weaknesses, e.g., with respect to estimation quality, generalization abilities and computational complexity. To actually harness the potential of uncertainty quantification, estimators are required whose properties closely match the requirements of a given use case. In this work, we propose a framework that, firstly, structures and shapes these requirements, secondly, guides the selection of a suitable uncertainty estimation method and, thirdly, provides strategies to validate this choice and to uncover structural weaknesses. By contributing tailored uncertainty estimation in this sense, our framework helps to foster trustworthy DL systems. Moreover, it anticipates prospective machine learning regulations that require, e.g., in the EU, evidences for the technical appropriateness of machine learning systems. Our framework provides such evidences for system components modeling uncertainty.
Image segmentation has been increasingly applied in medical settings as recent developments have skyrocketed the potential applications of deep learning. Urology, specifically, is one field of medicine that is primed for the adoption of a real-time image segmentation system with the long-term aim of automating endoscopic stone treatment. In this project, we explored supervised deep learning models to annotate kidney stones in surgical endoscopic video feeds. In this paper, we describe how we built a dataset from the raw videos and how we developed a pipeline to automate as much of the process as possible. For the segmentation task, we adapted and analyzed three baseline deep learning models -- U-Net, U-Net++, and DenseNet -- to predict annotations on the frames of the endoscopic videos with the highest accuracy above 90\%. To show clinical potential for real-time use, we also confirmed that our best trained model can accurately annotate new videos at 30 frames per second. Our results demonstrate that the proposed method justifies continued development and study of image segmentation to annotate ureteroscopic video feeds.
Various attack methods against recommender systems have been proposed in the past years, and the security issues of recommender systems have drawn considerable attention. Traditional attacks attempt to make target items recommended to as many users as possible by poisoning the training data. Benifiting from the feature of protecting users' private data, federated recommendation can effectively defend such attacks. Therefore, quite a few works have devoted themselves to developing federated recommender systems. For proving current federated recommendation is still vulnerable, in this work we probe to design attack approaches targeting deep learning based recommender models in federated learning scenarios. Specifically, our attacks generate poisoned gradients for manipulated malicious users to upload based on two strategies (i.e., random approximation and hard user mining). Extensive experiments show that our well-designed attacks can effectively poison the target models, and the attack effectiveness sets the state-of-the-art.
Human activity recognition using deep learning techniques has become increasing popular because of its high effectivity with recognizing complex tasks, as well as being relatively low in costs compared to more traditional machine learning techniques. This paper surveys some state-of-the-art human activity recognition models that are based on deep learning architecture and has layers containing Convolution Neural Networks (CNN), Long Short-Term Memory (LSTM), or a mix of more than one type for a hybrid system. The analysis outlines how the models are implemented to maximize its effectivity and some of the potential limitations it faces.
Non-Hermitian systems offer new platforms for unusual physical properties that can be flexibly manipulated by redistribution of the real and imaginary parts of refractive indices, whose presence breaks conventional wave propagation symmetries, leading to asymmetric reflection and symmetric transmission with respect to the wave propagation direction. Here, we use supervised and unsupervised learning techniques for knowledge acquisition in non-Hermitian systems which accelerate the inverse design process. In particular, we construct a deep learning model that relates the transmission and asymmetric reflection in non-conservative settings and proposes sub-manifold learning to recognize non-Hermitian features from transmission spectra. The developed deep learning framework determines the feasibility of a desired spectral response for a given structure and uncovers the role of effective gain-loss parameters to tailor the spectral response. These findings pave the way for intelligent inverse design and shape our understanding of the physical mechanism in general non-Hermitian systems.
Code summarization with deep learning has been widely studied in recent years. Current deep learning models for code summarization generally follow the principle in neural machine translation and adopt the encoder-decoder framework, where the encoder learns the semantic representations from source code and the decoder transforms the learnt representations into human-readable text that describes the functionality of code snippets. Despite they achieve the new state-of-the-art performance, we notice that current models often either generate less fluent summaries, or fail to capture the core functionality, since they usually focus on a single type of code representations. As such we propose GypSum, a new deep learning model that learns hybrid representations using graph attention neural networks and a pre-trained programming and natural language model. We introduce particular edges related to the control flow of a code snippet into the abstract syntax tree for graph construction, and design two encoders to learn from the graph and the token sequence of source code, respectively. We modify the encoder-decoder sublayer in the Transformer's decoder to fuse the representations and propose a dual-copy mechanism to facilitate summary generation. Experimental results demonstrate the superior performance of GypSum over existing code summarization models.
On-road bicycle lanes improve safety for cyclists, and encourage participation in cycling for active transport and recreation. With many local authorities responsible for portions of the infrastructure, official maps and datasets of bicycle lanes may be out-of-date and incomplete. Even "crowdsourced" databases may have significant gaps, especially outside popular metropolitan areas. This thesis presents a method to create a map of bicycle lanes in a survey area by taking sample street scene images from each road, and then applying a deep learning model that has been trained to recognise bicycle lane symbols. The list of coordinates where bicycle lane markings are detected is then correlated to geospatial data about the road network to record bicycle lane routes. The method was applied to successfully build a map for a survey area in the outer suburbs of Melbourne. It was able to identify bicycle lanes not previously recorded in the official state government dataset, OpenStreetMap, or the "biking" layer of Google Maps.
This paper investigates a new online learning problem with doubly-streaming data, where the data streams are described by feature spaces that constantly evolve, with new features emerging and old features fading away. The challenges of this problem are two folds: 1) Data samples ceaselessly flowing in may carry shifted patterns over time, requiring learners to update hence adapt on-the-fly. 2) Newly emerging features are described by very few samples, resulting in weak learners that tend to make error predictions. A plausible idea to overcome the challenges is to establish relationship between the pre-and-post evolving feature spaces, so that an online learner can leverage the knowledge learned from the old features to better the learning performance on the new features. Unfortunately, this idea does not scale up to high-dimensional media streams with complex feature interplay, which suffers an tradeoff between onlineness (biasing shallow learners) and expressiveness(requiring deep learners). Motivated by this, we propose a novel OLD^3S paradigm, where a shared latent subspace is discovered to summarize information from the old and new feature spaces, building intermediate feature mapping relationship. A key trait of OLD^3S is to treat the model capacity as a learnable semantics, yields optimal model depth and parameters jointly, in accordance with the complexity and non-linearity of the input data streams in an online fashion. Both theoretical analyses and empirical studies substantiate the viability and effectiveness of our proposal.
Nowadays, we are more and more reliant on Deep Learning (DL) models and thus it is essential to safeguard the security of these systems. This paper explores the security issues in Deep Learning and analyses, through the use of experiments, the way forward to build more resilient models. Experiments are conducted to identify the strengths and weaknesses of a new approach to improve the robustness of DL models against adversarial attacks. The results show improvements and new ideas that can be used as recommendations for researchers and practitioners to create increasingly better DL algorithms.
Time Series Forecasting (TSF) is used to predict the target variables at a future time point based on the learning from previous time points. To keep the problem tractable, learning methods use data from a fixed length window in the past as an explicit input. In this paper, we study how the performance of predictive models change as a function of different look-back window sizes and different amounts of time to predict into the future. We also consider the performance of the recent attention-based Transformer models, which has had good success in the image processing and natural language processing domains. In all, we compare four different deep learning methods (RNN, LSTM, GRU, and Transformer) along with a baseline method. The dataset (hourly) we used is the Beijing Air Quality Dataset from the UCI website, which includes a multivariate time series of many factors measured on an hourly basis for a period of 5 years (2010-14). For each model, we also report on the relationship between the performance and the look-back window sizes and the number of predicted time points into the future. Our experiments suggest that Transformer models have the best performance with the lowest Mean Average Errors (MAE = 14.599, 23.273) and Root Mean Square Errors (RSME = 23.573, 38.131) for most of our single-step and multi-steps predictions. The best size for the look-back window to predict 1 hour into the future appears to be one day, while 2 or 4 days perform the best to predict 3 hours into the future.
On-device deep learning is rapidly gaining popularity in mobile applications. Compared to offloading deep learning from smartphones to the cloud, on-device deep learning enables offline model inference while preserving user privacy. However, such mechanisms inevitably store models on users' smartphones and may invite adversarial attacks as they are accessible to attackers. Due to the characteristic of the on-device model, most existing adversarial attacks cannot be directly applied for on-device models. In this paper, we introduce a grey-box adversarial attack framework to hack on-device models by crafting highly similar binary classification models based on identified transfer learning approaches and pre-trained models from TensorFlow Hub. We evaluate the attack effectiveness and generality in terms of four different settings including pre-trained models, datasets, transfer learning approaches and adversarial attack algorithms. The results demonstrate that the proposed attacks remain effective regardless of different settings, and significantly outperform state-of-the-art baselines. We further conduct an empirical study on real-world deep learning mobile apps collected from Google Play. Among 53 apps adopting transfer learning, we find that 71.7\% of them can be successfully attacked, which includes popular ones in medicine, automation, and finance categories with critical usage scenarios. The results call for the awareness and actions of deep learning mobile app developers to secure the on-device models. The code of this work is available at https://github.com/Jinxhy/SmartAppAttack