AAAI 2024 Papers

Data-Free Knowledge Distillation (DFKD) enables knowledge transfer from a pretrained teacher to a light-weighted student without original training data. Existing works are limited by a strong assumption that samples used to pretrain the teacher model are balanced, which is, however, unrealistic for many real-world tasks. In this work, we investigated a pragmatic yet under-explored problem: how to perform DFKD from a teacher model pretrained from imbalanced data. We observe a seemingly counter-intuitive phenomenon, i.e., adversarial DFKD algorithms favour minority classes, while causing a disastrous impact on majority classes. We theoretically prove that a biased teacher could cause severe disparity on different groups of synthetic data in adversarial distillation, which further exacerbates the mode collapse of a generator and consequently degenerates the overall accuracy of a distilled student model. To tackle this problem, we propose a class-adaptive regularization method, aiming to encourage impartial representation learning of a generator among different classes under a constrained learning formulation. We devise a primal-dual algorithm to solve the target optimization problem. Through extensive experiments, we show that our method mitigates the biased learning of majority classes in DFKD and improves the overall performance compared with baselines. Code will be available at https://github.com/ldpbuaa/ipad.

Biologically Inspired Design (BID), or Biomimicry, is a problem-solving methodology that applies analogies from nature to solve engineering challenges. For example, Speedo engineers designed swimsuits based on shark skin. Finding relevant biological solutions for real-world problems poses significant challenges, both due to the limited biological knowledge engineers and designers typically possess and to the limited BID resources. Existing BID datasets are hand-curated and small, and scaling them up requires costly human annotations. In this paper, we introduce BARcode (Biological Analogy Retriever), a search engine for automatically mining bio-inspirations from the web at scale. Using advances in natural language understanding and data programming, BARcode identifies potential inspirations for engineering challenges. Our experiments demonstrate that BARcode can retrieve inspirations that are valuable to engineers and designers tackling real-world problems, as well as recover famous historical BID examples. We release data and code; we view BARcode as a step towards addressing the challenges that have historically hindered the practical application of BID to engineering innovation.

A popular framework for enforcing safe actions in Reinforcement Learning (RL) is Constrained RL, where trajectory based constraints on expected cost (or other cost measures) are employed to enforce safety and more importantly these constraints are enforced while maximizing expected reward. Most recent approaches for solving Constrained RL convert the trajectory based cost constraint into a surrogate problem that can be solved using minor modifications to RL methods. A key drawback with such approaches is an over or underestimation of the cost constraint at each state. Therefore, we provide an approach that does not modify the trajectory based cost constraint and instead imitates "good" trajectories and avoids "bad" trajectories generated from incrementally improving policies. We employ an oracle that utilizes a reward threshold (which is varied with learning) and the overall cost constraint to label trajectories as "good" or "bad". A key advantage of our approach is that we are able to work from any starting policy or set of trajectories and improve on it. In an exhaustive set of experiments, we demonstrate that our approach is able to outperform top benchmark approaches for solving Constrained RL problems, with respect to expected cost, CVaR cost, or even unknown cost constraints.

Effective exploration is crucial to discovering optimal strategies for multi-agent reinforcement learning (MARL) in complex coordination tasks. Existing methods mainly utilize intrinsic rewards to enable committed exploration or use role-based learning for decomposing joint action spaces instead of directly conducting a collective search in the entire action-observation space. However, they often face challenges obtaining specific joint action sequences to reach successful states in long-horizon tasks. To address this limitation, we propose Imagine, Initialize, and Explore (IIE), a novel method that offers a promising solution for efficient multi-agent exploration in complex scenarios. IIE employs a transformer model to imagine how the agents reach a critical state that can influence each other's transition functions. Then, we initialize the environment at this state using a simulator before the exploration phase. We formulate the imagination as a sequence modeling problem, where the states, observations, prompts, actions, and rewards are predicted autoregressively. The prompt consists of timestep-to-go, return-to-go, influence value, and one-shot demonstration, specifying the desired state and trajectory as well as guiding the action generation. By initializing agents at the critical states, IIE significantly increases the likelihood of discovering potentially important under-explored regions. Despite its simplicity, empirical results demonstrate that our method outperforms multi-agent exploration baselines on the StarCraft Multi-Agent Challenge (SMAC) and SMACv2 environments. Particularly, IIE shows improved performance in the sparse-reward SMAC tasks and produces more effective curricula over the initialized states than other generative methods, such as CVAE-GAN and diffusion models.

Artificial intelligence (AI) and its teaching in the K-12 grades has been championed as a vital need for the United States due to the technology's future prominence in the 21st century. However, there remain several barriers to effective AI lessons at these age groups including the broad range of interdisciplinary knowledge needed and the lack of formal training or preparation for teachers to implement these lessons. In this experience report, we present ImageSTEAM, a teacher professional development for creating lessons surrounding computer vision, machine learning, and computational photography/cameras targeted for middle school grades 6-8 classes. Teacher professional development workshops were conducted in the states of Arizona and Georgia from 2021-2023 where lessons were co-created with teachers to introduce various specific visual computing concepts while aligning to state and national standards. In addition, the use of a variety of computer vision and image processing software including custom designed Python notebooks were created as technology activities and demonstrations to be used in the classroom. Educational research showed that teachers improved their self-efficacy and outcomes for concepts in computer vision, machine learning, and artificial intelligence when participating in the program. Results from the professional development workshops highlight key opportunities and challenges in integrating this content into the standard curriculum, the benefits of a co-creation pedagogy, and the positive impact on teacher and student's learning experiences. The open-source program curriculum is available at www.imagesteam.org.

Most pre-trained learning systems are known to suffer from bias, which typically emerges from the data, the model, or both. Measuring and quantifying bias and its sources is a challenging task and has been extensively studied in image captioning. Despite the significant effort in this direction, we observed that existing metrics lack consistency in the inclusion of the visual signal. In this paper, we introduce a new bias assessment metric, dubbed ImageCaptioner2, for image captioning. Instead of measuring the absolute bias in the model or the data, ImageCaptioner2pay more attention to the bias introduced by the model w.r.t the data bias, termed bias amplification. Unlike the existing methods, which only evaluate the image captioning algorithms based on the generated captions only, ImageCaptioner2incorporates the image while measuring the bias. In addition, we design a formulation for measuring the bias of generated captions as prompt-based image captioning instead of using language classifiers. Finally, we apply our ImageCaptioner2metric across 11 different image captioning architectures on three different datasets, i.e., MS-COCO caption dataset, Artemis V1, and Artemis V2, and on three different protected attributes, i.e., gender, race, and emotions. Consequently, we verify the effectiveness of our ImageCaptioner2metric by proposing Anonymous-Bench, which is a novel human evaluation paradigm for bias metrics. Our metric shows significant superiority over the recent bias metric; LIC, in terms of human alignment, where the correlation scores are 80% and 54% for our metric and LIC, respectively. The code and more details are available at https://eslambakr.github.io/imagecaptioner2.github.io/.

Scene text recognition is inherently a vision-language task. However, previous works have predominantly focused either on extracting more robust visual features or designing better language modeling. How to effectively and jointly model vision and language to mitigate heavy reliance on a single modality remains a problem. In this paper, aiming to enhance vision-language reasoning in scene text recognition, we present a balanced, unified and synchronized vision-language reasoning network (BUSNet). Firstly, revisiting the image as a language by balanced concatenation along length dimension alleviates the issue of over-reliance on vision or language. Secondly, BUSNet learns an ensemble of unified external and internal vision-language model with shared weight by masked modality modeling (MMM). Thirdly, a novel vision-language reasoning module (VLRM) with synchronized vision-language decoding capacity is proposed. Additionally, BUSNet achieves improved performance through iterative reasoning, which utilizes the vision-language prediction as a new language input. Extensive experiments indicate that BUSNet achieves state-of-the-art performance on several mainstream benchmark datasets and more challenge datasets for both synthetic and real training data compared to recent outstanding methods. Code and dataset will be available at https://github.com/jjwei66/BUSNet.

Social media platforms are being increasingly used by malicious actors to share unsafe content, such as images depicting sexual activity, cyberbullying, and self-harm. Consequently, major platforms use artificial intelligence (AI) and human moderation to obfuscate such images to make them safer. Two critical needs for obfuscating unsafe images is that an accurate rationale for obfuscating image regions must be provided, and the sensitive regions should be obfuscated (e.g. blurring) for users' safety. This process involves addressing two key problems: (1) the reason for obfuscating unsafe images demands the platform to provide an accurate rationale that must be grounded in unsafe image-specific attributes, and (2) the unsafe regions in the image must be minimally obfuscated while still depicting the safe regions. In this work, we address these key issues by first performing visual reasoning by designing a visual reasoning model (VLM) conditioned on pre-trained unsafe image classifiers to provide an accurate rationale grounded in unsafe image attributes, and then proposing a counterfactual explanation algorithm that minimally identifies and obfuscates unsafe regions for safe viewing, by first utilizing an unsafe image classifier attribution matrix to guide segmentation for a more optimal subregion segmentation followed by an informed greedy search to determine the minimum number of subregions required to modify the classifier's output based on attribution score. Extensive experiments on uncurated data from social networks emphasize the efficacy of our proposed method. We make our code available at: https://github.com/SecureAIAutonomyLab/ConditionalVLM

The emergence of ChatGPT has once again sparked research in generative artificial intelligence (GAI). While people have been amazed by the generated results, they have also noticed the reasoning potential reflected in the generated textual content. However, this current ability for causal reasoning is primarily limited to the domain of language generation, such as in models like GPT-3. In visual modality, there is currently no equivalent research. Considering causal reasoning in visual content generation is significant. This is because visual information contains infinite granularity. Particularly, images can provide more intuitive and specific demonstrations for certain reasoning tasks, especially when compared to coarse-grained text. Hence, we propose a new image generation task called visual question answering with image (VQAI) and establish a dataset of the same name based on the classic Tom and Jerry animated series. Additionally, we develop a new paradigm for image generation to tackle the challenges of this task. Finally, we perform extensive experiments and analyses, including visualizations of the generated content and discussions on the potentials and limitations. The code and data are publicly available under the license of CC BY-NC-SA 4.0 for academic and non-commercial usage at: https://github.com/IEIT-AGI/MIX-Shannon/blob/main/projects/VQAI/lgd_vqai.md.

Image captioning requires numerous annotated image-text pairs, resulting in substantial annotation costs. Recently, large models (e.g. diffusion models and large language models) have excelled in producing high-quality images and text. This potential can be harnessed to create synthetic image-text pairs for training captioning models. Synthetic data can improve cost and time efficiency in data collection, allow for customization to specific domains, bootstrap generalization capability for zero-shot performance, and circumvent privacy concerns associated with real-world data. However, existing methods struggle to attain satisfactory performance solely through synthetic data. We identify the issue as generated images from simple descriptions mostly capture a solitary perspective with limited context, failing to align with the intricate scenes prevalent in real-world imagery. To tackle this, we present an innovative pipeline that introduces multi-context data generation. Beginning with an initial text corpus, our approach employs a large language model to extract multiple sentences portraying the same scene from diverse viewpoints. These sentences are then condensed into a single sentence with multiple contexts. Subsequently, we generate intricate images using the condensed captions through diffusion models. Our model is exclusively trained on synthetic image-text pairs crafted through this process. The effectiveness of our pipeline is validated through experimental results in both the in-domain and cross-domain settings, where it achieves state-of-the-art performance on well-known datasets such as MSCOCO, Flickr30k, and NoCaps.

Today's machine learning (ML) applications predominantly adhere to a standard paradigm: the decision maker designs the algorithm by optimizing a model for some objective function. While this has proven to be a powerful approach in many domains, it comes with inherent side effects: the power over the algorithmic outcomes lies solely in the hands of the algorithm designer, and alternative objectives, such as fairness, are often disregarded. This is particularly problematic if the algorithm is used to make consequential decisions that affect peoples lives. My research focuses on developing principled methods to characterize and address the mismatch between these different objectives.

Public libraries are an essential public good. We ask: are urban library systems providing equitable service to all residents, in terms of the books they have access to and check out? If not, what causes disparities: heterogeneous book collections, resident behavior and access, and/or operational policies? Existing methods leverage only system-level outcome data (such as overall checkouts per branch), and so cannot distinguish between these factors. As a result, it is difficult to use their results to guide interventions to increase equitable access. We propose a Bayesian framework to characterize book checkout behavior across multiple branches of a library system, learning heterogeneous book popularity, overall branch demand, and usage of the online hold system, while controlling for book availability. In collaboration with the New York Public Library, we apply our framework to granular data consisting of over 400,000 checkouts during 2022. We first show that our model significantly out-performs baseline methods in predicting checkouts at the book-branch level. Next, we study spatial and socioeconomic disparities. We show that disparities are largely driven by disparate use of the online holds system, which allows library patrons to receive books from any other branch through an online portal. This system thus leads to a large outflow of popular books from branches in lower income neighborhoods to those in high income ones. Finally, we illustrate the use of our model and insights to quantify the impact of potential interventions, such as changing how books are internally routed between branches to fulfill hold requests.

As algorithmic decision-making systems become more prevalent in society, ensuring the fairness of these systems is becoming increasingly important. Whilst there has been substantial research in building fair algorithmic decision-making systems, the majority of these methods require access to the training data, including personal characteristics, and are not transparent regarding which individuals are classified unfairly. In this paper, we propose a novel model-agnostic argumentation-based method to determine why an individual is classified differently in comparison to similar individuals. Our method uses a quantitative argumentation framework to represent attribute-value pairs of an individual and of those similar to them, and uses a well-known semantics to identify the attribute-value pairs in the individual contributing most to their different classification. We evaluate our method on two datasets commonly used in the fairness literature and illustrate its effectiveness in the identification of bias.

Most countries provide veterans with various benefits to reward their sacrifice. Unfortunately, many veterans have failed to prove their status due to loss of military records. Thus, some governments allow the verification of those veterans through "buddy statements" obtained from the people who can vouch for the buddy's participation in the war. However, it is still challenging for veterans to find guarantors directly. With this background, we suggest to utilizing historical war records of combined operations to increase the pool of potential guarantors for the buddy statements. However, a combined operation network among troops can have missing edges and perturbations on attributes of the troop due to inaccurate information. In this study, we learn from some recorded interactions which might be incomplete and noisy, and predict missing linkages among the troops that might have interacted together in the war, by proposing Robust-SEAL (learning from Subgraphs, Embeddings, and Attributes for Link prediction). It combines two Graph Neural Network (GNN) architectures: robust Graph Convolutional Network which considers the uncertainty of node attributes with a probabilistic approach, and SEAL which improves the expressive power of the GNN with a labeling trick. Our proposed approach was applied to Korean War data with perturbations. For experimentations, we hid some actual interactions and found that Robust-SEAL restores missing interactions better than other GNN-based baselines.

Image-text matching bridges vision and language, which is a fundamental task in multimodal intelligence. Its key challenge lies in how to capture visual-semantic relevance. Fine-grained semantic interactions come from fragment alignments between image regions and text words. However, not all fragments contribute to image-text relevance, and many existing methods are devoted to mining the vital ones to measure the relevance accurately. How well image and text relate depends on the degree of semantic sharing between them. Treating the degree as an effect and fragments as its possible causes, we define those indispensable causes for the generation of the degree as necessary undertakers, i.e., if any of them did not occur, the relevance would be no longer valid. In this paper, we revisit image-text matching in the causal view and uncover inherent causal properties of relevance generation. Then we propose a novel theoretical prototype for estimating the probability-of-necessity of fragments, PN_f, for the degree of semantic sharing by means of causal inference, and further design a Necessary Undertaker Identification Framework (NUIF) for image-text matching, which explicitly formalizes the fragment's contribution to image-text relevance by modeling PN_f in two ways. Extensive experiments show our method achieves state-of-the-art on benchmarks Flickr30K and MSCOCO.

Causal discovery with latent variables is a crucial but challenging task. Despite the emergence of numerous methods aimed at addressing this challenge, they are not fully identified to the structure that two observed variables are influenced by one latent variable and there might be a directed edge in between. Interestingly, we notice that this structure can be identified through the utilization of higher-order cumulants. By leveraging the higher-order cumulants of non-Gaussian data, we provide an analytical solution for estimating the causal coefficients or their ratios. With the estimated (ratios of) causal coefficients, we propose a novel approach to identify the existence of a causal edge between two observed variables subject to latent variable influence. In case when such a causal edge exits, we introduce an asymmetry criterion to determine the causal direction. The experimental results demonstrate the effectiveness of our proposed method.

Missing data are an unavoidable complication frequently encountered in many causal discovery tasks. When a missing process depends on the missing values themselves (known as self-masking missingness), the recovery of the joint distribution becomes unattainable, and detecting the presence of such self-masking missingness remains a perplexing challenge. Consequently, due to the inability to reconstruct the original distribution and to discern the underlying missingness mechanism, simply applying existing causal discovery methods would lead to wrong conclusions. In this work, we found that the recent advances additive noise model has the potential for learning causal structure under the existence of the self-masking missingness. With this observation, we aim to investigate the identification problem of learning causal structure from missing data under an additive noise model with different missingness mechanisms, where the `no self-masking missingness' assumption can be eliminated appropriately. Specifically, we first elegantly extend the scope of identifiability of causal skeleton to the case with weak self-masking missingness (i.e., no other variable could be the cause of self-masking indicators except itself). We further provide the sufficient and necessary identification conditions of the causal direction under additive noise model and show that the causal structure can be identified up to an IN-equivalent pattern. We finally propose a practical algorithm based on the above theoretical results on learning the causal skeleton and causal direction. Extensive experiments on synthetic and real data demonstrate the efficiency and effectiveness of the proposed algorithms.

Linear structural causal models (SCMs) are used to express and analyze the relationships between random variables. Direct causal effects are represented as directed edges and confounding factors as bidirected edges. Identifying the causal parameters from correlations between the nodes is an open problem in artificial intelligence. In this paper, we study SCMs whose directed component forms a tree. Van der Zander et al. give a PSPACE-algorithm for the identification problem in this case, which is a significant improvement over the general Gröbner basis approach, which has doubly-exponential time complexity in the number of structural parameters. However, they do not show that their algorithm is complete. In this work, we present a randomized polynomial-time algorithm, which solves the identification problem for tree-shaped SCMs. For every structural parameter, our algorithms decides whether it is generically identifiable, generically 2-identifiable, or generically unidentifiable. (No other cases can occur.) In the first two cases, it provides one or two fractional affine square root terms of polynomials (FASTPs) for the corresponding parameter, respectively. In particular, our algorithm is not only polynomial time, but also complete for for tree-shaped SCMs.

Dynamic structural causal models (SCMs) are a powerful framework for reasoning in dynamic systems about direct effects which measure how a change in one variable affects another variable while holding all other variables constant. The causal relations in a dynamic structural causal model can be qualitatively represented with an acyclic full-time causal graph. Assuming linearity and no hidden confounding and given the full-time causal graph, the direct causal effect is always identifiable. However, in many application such a graph is not available for various reasons but nevertheless experts have access to the summary causal graph of the full-time causal graph which represents causal relations between time series while omitting temporal information and allowing cycles. This paper presents a complete identifiability result which characterizes all cases for which the direct effect is graphically identifiable from a summary causal graph and gives two sound finite adjustment sets that can be used to estimate the direct effect whenever it is identifiable.