Auditing trained deep learning (DL) models prior to deployment is vital in preventing unintended consequences. One of the biggest challenges in auditing is in understanding how we can obtain human-interpretable specifications that are directly useful to the end-user. We address this challenge through a sequence of semantically-aligned unit tests, where each unit test verifies whether a predefined specification (e.g., accuracy over 95%) is satisfied with respect to controlled and semantically aligned variations in the input space (e.g., in face recognition, the angle relative to the camera). We perform these unit tests by directly verifying the semantically aligned variations in an interpretable latent space of a generative model. Our framework, AuditAI, bridges the gap between interpretable formal verification and scalability. With evaluations on four different datasets, covering images of towers, chest X-rays, human faces, and ImageNet classes, we show how AuditAI allows us to obtain controlled variations for verification and certified training while addressing the limitations of verifying using only pixel-space perturbations. A blog post accompanying the paper is at this link https://developer.nvidia.com/blog/nvidia-research-auditing-ai-models-for-verified-deployment-under-semantic-specifications

Blind and visually challenged face multiple issues with navigating the world independently. Some of these challenges include finding the shortest path to a destination and detecting obstacles from a distance. To tackle this issue, this paper proposes ViT Cane, which leverages a vision transformer model in order to detect obstacles in real-time. Our entire system consists of a Pi Camera Module v2, Raspberry Pi 4B with 8GB Ram and 4 motors. Based on tactile input using the 4 motors, the obstacle detection model is highly efficient in helping visually impaired navigate unknown terrain and is designed to be easily reproduced. The paper discusses the utility of a Visual Transformer model in comparison to other CNN based models for this specific application. Through rigorous testing, the proposed obstacle detection model has achieved higher performance on the Common Object in Context (COCO) data set than its CNN counterpart. Comprehensive field tests were conducted to verify the effectiveness of our system for holistic indoor understanding and obstacle avoidance.

Although researchers increasingly adopt machine learning to model travel behavior, they predominantly focus on prediction accuracy, ignoring the ethical challenges embedded in machine learning algorithms. This study introduces an important missing dimension - computational fairness - to travel behavior analysis. We first operationalize computational fairness by equality of opportunity, then differentiate between the bias inherent in data and the bias introduced by modeling. We then demonstrate the prediction disparities in travel behavior modeling using the 2017 National Household Travel Survey (NHTS) and the 2018-2019 My Daily Travel Survey in Chicago. Empirically, deep neural network (DNN) and discrete choice models (DCM) reveal consistent prediction disparities across multiple social groups: both over-predict the false negative rate of frequent driving for the ethnic minorities, the low-income and the disabled populations, and falsely predict a higher travel burden of the socially disadvantaged groups and the rural populations than reality. Comparing DNN with DCM, we find that DNN can outperform DCM in prediction disparities because of DNN's smaller misspecification error. To mitigate prediction disparities, this study introduces an absolute correlation regularization method, which is evaluated with synthetic and real-world data. The results demonstrate the prevalence of prediction disparities in travel behavior modeling, and the disparities still persist regarding a variety of model specifics such as the number of DNN layers, batch size and weight initialization. Since these prediction disparities can exacerbate social inequity if prediction results without fairness adjustment are used for transportation policy making, we advocate for careful consideration of the fairness problem in travel behavior modeling, and the use of bias mitigation algorithms for fair transport decisions.

Self-supervised methods play an increasingly important role in monocular depth estimation due to their great potential and low annotation cost. To close the gap with supervised methods, recent works take advantage of extra constraints, e.g., semantic segmentation. However, these methods will inevitably increase the burden on the model. In this paper, we show theoretical and empirical evidence that the potential capacity of self-supervised monocular depth estimation can be excavated without increasing this cost. In particular, we propose (1) a novel data augmentation approach called data grafting, which forces the model to explore more cues to infer depth besides the vertical image position, (2) an exploratory self-distillation loss, which is supervised by the self-distillation label generated by our new post-processing method - selective post-processing, and (3) the full-scale network, designed to endow the encoder with the specialization of depth estimation task and enhance the representational power of the model. Extensive experiments show that our contributions can bring significant performance improvement to the baseline with even less computational overhead, and our model, named EPCDepth, surpasses the previous state-of-the-art methods even those supervised by additional constraints.

Recently, surrogate models based on deep learning have attracted much attention for engineering analysis and optimization. As the construction of data pairs in most engineering problems is time-consuming, data acquisition is becoming the predictive capability bottleneck of most deep surrogate models, which also exists in surrogate for thermal analysis and design. To address this issue, this paper develops a physics-informed convolutional neural network (CNN) for the thermal simulation surrogate. The network can learn a mapping from heat source layout to the steady-state temperature field without labeled data, which equals solving an entire family of partial difference equations (PDEs). To realize the physics-guided training without labeled data, we employ the heat conduction equation and finite difference method to construct the loss function. Since the solution is sensitive to boundary conditions, we properly impose hard constraints by padding in the Dirichlet and Neumann boundary conditions. In addition, the neural network architecture is well-designed to improve the prediction precision of the problem at hand, and pixel-level online hard example mining is introduced to overcome the imbalance of optimization difficulty in the computation domain. The experiments demonstrate that the proposed method can provide comparable predictions with numerical method and data-driven deep learning models. We also conduct various ablation studies to investigate the effectiveness of the network component and training methods proposed in this paper.

It is well known that natural language models are vulnerable to adversarial attacks, which are mostly input-specific in nature. Recently, it has been shown that there also exist input-agnostic attacks in NLP models, called universal adversarial triggers. However, existing methods to craft universal triggers are data intensive. They require large amounts of data samples to generate adversarial triggers, which are typically inaccessible by attackers. For instance, previous works take 3000 data samples per class for the SNLI dataset to generate adversarial triggers. In this paper, we present a novel data-free approach, MINIMAL, to mine input-agnostic adversarial triggers from models. Using the triggers produced with our data-free algorithm, we reduce the accuracy of Stanford Sentiment Treebank's positive class from 93.6% to 9.6%. Similarly, for the Stanford Natural Language Inference (SNLI), our single-word trigger reduces the accuracy of the entailment class from 90.95% to less than 0.6\%. Despite being completely data-free, we get equivalent accuracy drops as data-dependent methods.

Multi-source Domain Adaptation (MDA) aims to transfer predictive models from multiple, fully-labeled source domains to an unlabeled target domain. However, in many applications, relevant labeled source datasets may not be available, and collecting source labels can be as expensive as labeling the target data itself. In this paper, we investigate Multi-source Few-shot Domain Adaptation (MFDA): a new domain adaptation scenario with limited multi-source labels and unlabeled target data. As we show, existing methods often fail to learn discriminative features for both source and target domains in the MFDA setting. Therefore, we propose a novel framework, termed Multi-Source Few-shot Adaptation Network (MSFAN), which can be trained end-to-end in a non-adversarial manner. MSFAN operates by first using a type of prototypical, multi-domain, self-supervised learning to learn features that are not only domain-invariant but also class-discriminative. Second, MSFAN uses a small, labeled support set to enforce feature consistency and domain invariance across domains. Finally, prototypes from multiple sources are leveraged to learn better classifiers. Compared with state-of-the-art MDA methods, MSFAN improves the mean classification accuracy over different domain pairs on MFDA by 20.2%, 9.4%, and 16.2% on Office, Office-Home, and DomainNet, respectively.

With the hardware development of mobile devices, it is possible to build the recommendation models on the mobile side to utilize the fine-grained features and the real-time feedbacks. Compared to the straightforward mobile-based modeling appended to the cloud-based modeling, we propose a Slow-Fast learning mechanism to make the Mobile-Cloud Collaborative recommendation (MC$^2$-SF) mutual benefit. Specially, in our MC$^2$-SF, the cloud-based model and the mobile-based model are respectively treated as the slow component and the fast component, according to their interaction frequency in real-world scenarios. During training and serving, they will communicate the prior/privileged knowledge to each other to help better capture the user interests about the candidates, resembling the role of System I and System II in the human cognition. We conduct the extensive experiments on three benchmark datasets and demonstrate the proposed MC$^2$-SF outperforms several state-of-the-art methods.

This paper attacks an emerging challenge of multi-modal retinal disease recognition. Given a multi-modal case consisting of a color fundus photo (CFP) and an array of OCT B-scan images acquired during an eye examination, we aim to build a deep neural network that recognizes multiple vision-threatening diseases for the given case. As the diagnostic efficacy of CFP and OCT is disease-dependent, the network's ability of being both selective and interpretable is important. Moreover, as both data acquisition and manual labeling are extremely expensive in the medical domain, the network has to be relatively lightweight for learning from a limited set of labeled multi-modal samples. Prior art on retinal disease recognition focuses either on a single disease or on a single modality, leaving multi-modal fusion largely underexplored. We propose in this paper Multi-Modal Multi-Instance Learning (MM-MIL) for selectively fusing CFP and OCT modalities. Its lightweight architecture (as compared to current multi-head attention modules) makes it suited for learning from relatively small-sized datasets. For an effective use of MM-MIL, we propose to generate a pseudo sequence of CFPs by over sampling a given CFP. The benefits of this tactic include well balancing instances across modalities, increasing the resolution of the CFP input, and finding out regions of the CFP most relevant with respect to the final diagnosis. Extensive experiments on a real-world dataset consisting of 1,206 multi-modal cases from 1,193 eyes of 836 subjects demonstrate the viability of the proposed model.

Though recent end-to-end neural models have shown promising progress on Conversational Recommender System (CRS), two key challenges still remain. First, the recommended items cannot be always incorporated into the generated replies precisely and appropriately. Second, only the items mentioned in the training corpus have a chance to be recommended in the conversation. To tackle these challenges, we introduce a novel framework called NTRD for recommender dialogue system that decouples the dialogue generation from the item recommendation. NTRD has two key components, i.e., response template generator and item selector. The former adopts an encoder-decoder model to generate a response template with slot locations tied to target items, while the latter fills in slot locations with the proper items using a sufficient attention mechanism. Our approach combines the strengths of both classical slot filling approaches (that are generally controllable) and modern neural NLG approaches (that are generally more natural and accurate). Extensive experiments on the benchmark ReDial show our NTRD significantly outperforms the previous state-of-the-art methods. Besides, our approach has the unique advantage to produce novel items that do not appear in the training set of dialogue corpus. The code is available at \url{https://github.com/jokieleung/NTRD}.