Search

Successful deployment of geological carbon storage (GCS) requires an extensive use of reservoir simulators for screening, ranking and optimization of storage sites. However, the time scales of GCS are such that no sufficient long-term data is available yet to validate the simulators against. As a consequence, there is currently no solid basis for assessing the quality with which the dynamics of large-scale GCS operations can be forecasted. To meet this knowledge gap, we have conducted a major GCS validation benchmark study. To achieve reasonable time scales, a laboratory-size geological storage formation was constructed (the "FluidFlower"), forming the basis for both the experimental and computational work. A validation experiment consisting of repeated GCS operations was conducted in the FluidFlower, providing what we define as the true physical dynamics for this system. Nine different research groups from around the world provided forecasts, both individually and collaboratively, based on a detailed physical and petrophysical characterization of the FluidFlower sands. The major contribution of this paper is a report and discussion of the results of the validation benchmark study, complemented by a description of the benchmarking process and the participating computational models. The forecasts from the participating groups are compared to each other and to the experimental data by means of various indicative qualitative and quantitative measures. By this, we provide a detailed assessment of the capabilities of reservoir simulators and their users to capture both the injection and post-injection dynamics of the GCS operations.

Interpretable AI tools are often motivated by the goal of understanding model behavior in out-of-distribution (OOD) contexts. Despite the attention this area of study receives, there are comparatively few cases where these tools have identified previously unknown bugs in models. We argue that this is due, in part, to a common feature of many interpretability methods: they analyze model behavior by using a particular dataset. This only allows for the study of the model in the context of features that the user can sample in advance. To address this, a growing body of research involves interpreting models using \emph{feature synthesis} methods that do not depend on a dataset. In this paper, we benchmark the usefulness of interpretability tools on debugging tasks. Our key insight is that we can implant human-interpretable trojans into models and then evaluate these tools based on whether they can help humans discover them. This is analogous to finding OOD bugs, except the ground truth is known, allowing us to know when an interpretation is correct. We make four contributions. (1) We propose trojan discovery as an evaluation task for interpretability tools and introduce a benchmark with 12 trojans of 3 different types. (2) We demonstrate the difficulty of this benchmark with a preliminary evaluation of 16 state-of-the-art feature attribution/saliency tools. Even under ideal conditions, given direct access to data with the trojan trigger, these methods still often fail to identify bugs. (3) We evaluate 7 feature-synthesis methods on our benchmark. (4) We introduce and evaluate 2 new variants of the best-performing method from the previous evaluation. A website for this paper and its code is at https://benchmarking-interpretability.csail.mit.edu/, Comment: In Proceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS 2023)

Human beings have rich ways of emotional expressions, including facial action, voice, and natural languages. Due to the diversity and complexity of different individuals, the emotions expressed by various modalities may be semantically irrelevant. Directly fusing information from different modalities may inevitably make the model subject to the noise from semantically irrelevant modalities. To tackle this problem, we propose a multimodal relevance estimation network to capture the relevant semantics among modalities in multimodal emotions. Specifically, we take advantage of an attention mechanism to reflect the semantic relevance weights of each modality. Moreover, we propose a relevant semantic estimation loss to weakly supervise the semantics of each modality. Furthermore, we make use of contrastive learning to optimize the similarity of category-level modality-relevant semantics across different modalities in feature space, thereby bridging the semantic gap between heterogeneous modalities. In order to better reflect the emotional state in the real interactive scenarios and perform the semantic relevance analysis, we collect a single-label discrete multimodal emotion dataset named SDME, which enables researchers to conduct multimodal semantic relevance research with large category bias. Experiments on continuous and discrete emotion datasets show that our model can effectively capture the relevant semantics, especially for the large deviations in modal semantics. The code and SDME dataset will be publicly available.

Given two graphs $H_1$ and $H_2$, a graph is $(H_1,H_2)$-free if it contains no induced subgraph isomorphic to $H_1$ or $H_2$. Let $P_t$ and $C_t$ be the path and the cycle on $t$ vertices, respectively. A bull is the graph obtained from a triangle with two disjoint pendant edges. In this paper, we show that there are finitely many 5-vertex-critical ($P_5$,bull)-free graphs., Comment: 21 pages

Beyond achieving higher compression efficiency over classical image compression codecs, deep image compression is expected to be improved with additional side information, e.g., another image from a different perspective of the same scene. To better utilize the side information under the distributed compression scenario, the existing method (Ayzik and Avidan 2020) only implements patch matching at the image domain to solve the parallax problem caused by the difference in viewing points. However, the patch matching at the image domain is not robust to the variance of scale, shape, and illumination caused by the different viewing angles, and can not make full use of the rich texture information of the side information image. To resolve this issue, we propose Multi-Scale Feature Domain Patch Matching (MSFDPM) to fully utilizes side information at the decoder of the distributed image compression model. Specifically, MSFDPM consists of a side information feature extractor, a multi-scale feature domain patch matching module, and a multi-scale feature fusion network. Furthermore, we reuse inter-patch correlation from the shallow layer to accelerate the patch matching of the deep layer. Finally, we nd that our patch matching in a multi-scale feature domain further improves compression rate by about 20% compared with the patch matching method at image domain (Ayzik and Avidan 2020)., Comment: This work is accepted by the AAAI 2023

Cyber threat hunting is a proactive search process for hidden threats in the organization's information system. It is a crucial component of active defense against advanced persistent threats (APTs). However, most of the current threat hunting methods rely on Cyber Threat Intelligence(CTI), which can find known attacks but cannot find unknown attacks that have not been disclosed by CTI. In this paper, we propose LogKernel, a threat hunting method based on graph kernel clustering which can effectively separates attack behaviour from benign activities. LogKernel first abstracts system audit logs into Behaviour Provenance Graphs (BPGs), and then clusters graphs by embedding them into a continuous space using a graph kernel. In particular, we design a new graph kernel clustering method based on the characteristics of BPGs, which can capture structure information and rich label information of the BPGs. To reduce false positives, LogKernel further quantifies the threat of abnormal behaviour. We evaluate LogKernel on the malicious dataset which includes seven simulated attack scenarios and the DAPRA CADETS dataset which includes four attack scenarios. The result shows that LogKernel can hunt all attack scenarios among them, and compared to the state-of-the-art methods, it can find unknown attacks.

Recent trends in AIGC effectively boosted the application of visual inspection. However, most of the available systems work in a human-in-the-loop manner and can not provide long-term support to the online application. To make a step forward, this paper outlines an automatic annotation system called SsaA, working in a self-supervised learning manner, for continuously making the online visual inspection in the manufacturing automation scenarios. Benefit from the self-supervised learning, SsaA is effective to establish a visual inspection application for the whole life-cycle of manufacturing. In the early stage, with only the anomaly-free data, the unsupervised algorithms are adopted to process the pretext task and generate coarse labels for the following data. Then supervised algorithms are trained for the downstream task. With user-friendly web-based interfaces, SsaA is very convenient to integrate and deploy both of the unsupervised and supervised algorithms. So far, the SsaA system has been adopted for some real-life industrial applications., Comment: 4 pages, 3 figures, ICCV 2023 Demo Track

Weitzman (1979) introduced the Pandora Box problem as a model for sequential search with inspection costs, and gave an elegant index-based policy that attains provably optimal expected payoff. In various scenarios, the searching agent may select an option without making a costly inspection. The variant of the Pandora box problem with non-obligatory inspection has attracted interest from both economics and algorithms researchers. Various simple algorithms have proved suboptimal, with the best known 0.8-approximation algorithm due to Guha et al. (2008). No hardness result for the problem was known. In this work, we show that it is NP-hard to compute an optimal policy for Pandora's problem with nonobligatory inspection. We also give a polynomial-time approximation scheme (PTAS) that computes policies with an expected payoff at least $(1 - \epsilon)$-fraction of the optimal, for arbitrarily small $\epsilon > 0$. On the side, we show the decision version of the problem to be in NP.

Adversarial attacks on data-driven algorithms applied in the power system will be a new type of threat to grid security. Literature has demonstrated that the adversarial attack on the deep-neural network can significantly mislead the load fore-cast of a power system. However, it is unclear how the new type of attack impacts the operation of the grid system. In this research, we manifest that the adversarial algorithm attack induces a significant cost-increase risk which will be exacerbated by the growing penetration of intermittent renewable energy. In Texas, a 5% adversarial attack can increase the total generation cost by 17% in a quarter, which accounts for around $20 million. When wind-energy penetration increases to over 40%, the 5% adversarial attack will inflate the genera-tion cost by 23%. Our research discovers a novel approach to defending against the adversarial attack: investing in the energy-storage system. All current literature focuses on developing algorithms to defend against adversarial attacks. We are the first research revealing the capability of using the facility in a physical system to defend against the adversarial algorithm attack in a system of the Internet of Things, such as a smart grid system., Comment: 5 pages, 5 main figures. Published in PESGM 2022

Few-shot abstractive summarization has become a challenging task in natural language generation. To support it, we designed a novel soft prompts architecture coupled with a prompt pre-training plus fine-tuning paradigm that is effective and tunes only extremely light parameters. The soft prompts include continuous input embeddings across an encoder and a decoder to fit the structure of the generation models. Importantly, a novel inner-prompt placed in the text is introduced to capture document-level information. The aim is to devote attention to understanding the document that better prompts the model to generate document-related content. The first step in the summarization procedure is to conduct prompt pre-training with self-supervised pseudo-data. This teaches the model basic summarizing capabilities. The model is then fine-tuned with few-shot examples. Experimental results on the CNN/DailyMail and XSum datasets show that our method, with only 0.1% of the parameters, outperforms full-model tuning where all model parameters are tuned. It also surpasses Prompt Tuning by a large margin and delivers competitive results against Prefix-Tuning with 3% of the parameters., Comment: Accepted as a long paper in COLING 2022

Checking how well a fitted model explains the data is one of the most fundamental parts of a Bayesian data analysis. However, existing model checking methods suffer from trade-offs between being well-calibrated, automated, and computationally efficient. To overcome these limitations, we propose split predictive checks (SPCs), which combine the ease-of-use and speed of posterior predictive checks with the good calibration properties of predictive checks that rely on model-specific derivations or inference schemes. We develop an asymptotic theory for two types of SPCs: single SPCs and the divided SPCs. Our results demonstrate that they offer complementary strengths. Single SPCs work well with smaller datasets and provide excellent power when there is substantial misspecification, such as when the estimate uncertainty in the test statistic is significantly underestimated. When the sample size is large, divided SPCs can provide better power and are able to detect more subtle form of misspecification. We validate the finite-sample utility of SPCs through extensive simulation experiments in exponential family and hierarchical models, and provide three real-data examples where SPCs offer novel insights and additional flexibility beyond what is available when using posterior predictive checks., Comment: v3: updated some discussion of model criticism and predictive checks; improved some figures

Multi-hop Question Answering (QA) requires the machine to answer complex questions by finding scattering clues and reasoning from multiple documents. Graph Network (GN) and Question Decomposition (QD) are two common approaches at present. The former uses the "black-box" reasoning process to capture the potential relationship between entities and sentences, thus achieving good performance. At the same time, the latter provides a clear reasoning logical route by decomposing multi-hop questions into simple single-hop sub-questions. In this paper, we propose a novel method to complete multi-hop QA from the perspective of Question Generation (QG). Specifically, we carefully design an end-to-end QG module on the basis of a classical QA module, which could help the model understand the context by asking inherently logical sub-questions, thus inheriting interpretability from the QD-based method and showing superior performance. Experiments on the HotpotQA dataset demonstrate that the effectiveness of our proposed QG module, human evaluation further clarifies its interpretability quantitatively, and thorough analysis shows that the QG module could generate better sub-questions than QD methods in terms of fluency, consistency, and diversity., Comment: 10 pages, Work-in-Progress

As Deep learning (DL) systems continuously evolve and grow, assuring their quality becomes an important yet challenging task. Compared to non-DL systems, DL systems have more complex team compositions and heavier data dependency. These inherent characteristics would potentially cause DL systems to be more vulnerable to bugs and, in the long run, to maintenance issues. Code smells are empirically tested as efficient indicators of non-DL systems. Therefore, we took a step forward into identifying code smells, and understanding their impact on maintenance in this comprehensive study. This is the first study on investigating code smells in the context of DL software systems, which helps researchers and practitioners to get a first look at what kind of maintenance modification made and what code smells developers have been dealing with. Our paper has three major contributions. First, we comprehensively investigated the maintenance modifications that have been made by DL developers via studying the evolution of DL systems, and we identified nine frequently occurred maintenance-related modification categories in DL systems. Second, we summarized five code smells in DL systems. Third, we validated the prevalence, and the impact of our newly identified code smells through a mixture of qualitative and quantitative analysis. We found that our newly identified code smells are prevalent and impactful on the maintenance of DL systems from the developer's perspective.