Hamed Firooz

Song Jiang, Zahra Shakeri, Aaron Chan et al. · berkeley, meta-ai

Chain-of-thought (CoT) prompting, which offers step-by-step problem-solving rationales, has impressively unlocked the reasoning potential of large language models (LLMs). Yet, the standard CoT is less effective in problems demanding multiple reasoning steps. This limitation arises from the complex reasoning process in multi-step problems: later stages often depend on the results of several steps earlier, not just the results of the immediately preceding step. Such complexities suggest the reasoning process is naturally represented as a graph. The almost linear and straightforward structure of CoT prompting, however, struggles to capture this complex reasoning graph. To address this challenge, we propose Residual Connection Prompting (RESPROMPT), a new prompting strategy that advances multi-step reasoning in LLMs. Our key idea is to reconstruct the reasoning graph within prompts. We achieve this by integrating necessary connections-links present in the reasoning graph but missing in the linear CoT flow-into the prompts. Termed "residual connections", these links are pivotal in morphing the linear CoT structure into a graph representation, effectively capturing the complex reasoning graphs inherent in multi-step problems. We evaluate RESPROMPT on six benchmarks across three diverse domains: math, sequential, and commonsense reasoning. For the open-sourced LLaMA family of models, RESPROMPT yields a significant average reasoning accuracy improvement of 12.5% on LLaMA-65B and 6.8% on LLaMA2-70B. Breakdown analysis further highlights RESPROMPT particularly excels in complex multi-step reasoning: for questions demanding at least five reasoning steps, RESPROMPT outperforms the best CoT based benchmarks by a remarkable average improvement of 21.1% on LLaMA-65B and 14.3% on LLaMA2-70B. Through extensive ablation studies and analyses, we pinpoint how to most effectively build residual connections.

Neha Kalibhat, Shweta Bhardwaj, Bayan Bruss et al.

We propose Automatic Feature Explanation using Contrasting Concepts (FALCON), an interpretability framework to explain features of image representations. For a target feature, FALCON captions its highly activating cropped images using a large captioning dataset (like LAION-400m) and a pre-trained vision-language model like CLIP. Each word among the captions is scored and ranked leading to a small number of shared, human-understandable concepts that closely describe the target feature. FALCON also applies contrastive interpretation using lowly activating (counterfactual) images, to eliminate spurious concepts. Although many existing approaches interpret features independently, we observe in state-of-the-art self-supervised and supervised models, that less than 20% of the representation space can be explained by individual features. We show that features in larger spaces become more interpretable when studied in groups and can be explained with high-order scoring concepts through FALCON. We discuss how extracted concepts can be used to explain and debug failures in downstream tasks. Finally, we present a technique to transfer concepts from one (explainable) representation space to another unseen representation space by learning a simple linear transformation. Code available at https://github.com/NehaKalibhat/falcon-explain.

Khalil Mrini, Shaoliang Nie, Jiatao Gu et al. · meta-ai

We propose an autoregressive entity linking model, that is trained with two auxiliary tasks, and learns to re-rank generated samples at inference time. Our proposed novelties address two weaknesses in the literature. First, a recent method proposes to learn mention detection and then entity candidate selection, but relies on predefined sets of candidates. We use encoder-decoder autoregressive entity linking in order to bypass this need, and propose to train mention detection as an auxiliary task instead. Second, previous work suggests that re-ranking could help correct prediction errors. We add a new, auxiliary task, match prediction, to learn re-ranking. Without the use of a knowledge base or candidate sets, our model sets a new state of the art in two benchmark datasets of entity linking: COMETA in the biomedical domain, and AIDA-CoNLL in the news domain. We show through ablation studies that each of the two auxiliary tasks increases performance, and that re-ranking is an important factor to the increase. Finally, our low-resource experimental results suggest that performance on the main task benefits from the knowledge learned by the auxiliary tasks, and not just from the additional training data.

Ajinkya Tejankar, Maziar Sanjabi, Qifan Wang et al.

Recently, self-supervised learning (SSL) was shown to be vulnerable to patch-based data poisoning backdoor attacks. It was shown that an adversary can poison a small part of the unlabeled data so that when a victim trains an SSL model on it, the final model will have a backdoor that the adversary can exploit. This work aims to defend self-supervised learning against such attacks. We use a three-step defense pipeline, where we first train a model on the poisoned data. In the second step, our proposed defense algorithm (PatchSearch) uses the trained model to search the training data for poisoned samples and removes them from the training set. In the third step, a final model is trained on the cleaned-up training set. Our results show that PatchSearch is an effective defense. As an example, it improves a model's accuracy on images containing the trigger from 38.2% to 63.7% which is very close to the clean model's accuracy, 64.6%. Moreover, we show that PatchSearch outperforms baselines and state-of-the-art defense approaches including those using additional clean, trusted data. Our code is available at https://github.com/UCDvision/PatchSearch

Neha Kalibhat, Kanika Narang, Hamed Firooz et al. · meta-ai

Self-supervised learning (SSL) has shown impressive results in downstream classification tasks. However, there is limited work in understanding their failure modes and interpreting their learned representations. In this paper, we study the representation space of state-of-the-art self-supervised models including SimCLR, SwaV, MoCo, BYOL, DINO, SimSiam, VICReg and Barlow Twins. Without the use of class label information, we discover discriminative features that correspond to unique physical attributes in images, present mostly in correctly-classified representations. Using these features, we can compress the representation space by up to 40% without significantly affecting linear classification performance. We then propose Self-Supervised Representation Quality Score (or Q-Score), an unsupervised score that can reliably predict if a given sample is likely to be mis-classified during linear evaluation, achieving AUPRC of 91.45 on ImageNet-100 and 78.78 on ImageNet-1K. Q-Score can also be used as a regularization term on pre-trained encoders to remedy low-quality representations. Fine-tuning with Q-Score regularization can boost the linear probing accuracy of SSL models by up to 5.8% on ImageNet-100 and 3.7% on ImageNet-1K compared to their baselines. Finally, using gradient heatmaps and Salient ImageNet masks, we define a metric to quantify the interpretability of each representation. We show that discriminative features are strongly correlated to core attributes and, enhancing these features through Q-score regularization makes SSL representations more interpretable.

Aaron Chan, Shaoliang Nie, Liang Tan et al. · meta-ai

Following how humans communicate, free-text rationales aim to use natural language to explain neural language model (LM) behavior. However, free-text rationales' unconstrained nature makes them prone to hallucination, so it is important to have metrics for free-text rationale quality. Existing free-text rationale metrics measure how consistent the rationale is with the LM's predicted label, but there is no protocol for assessing such metrics' reliability. Thus, we propose FRAME, a framework for evaluating rationale-label consistency (RLC) metrics for free-text rationales. FRAME is based on three axioms: (1) good metrics should yield highest scores for reference rationales, which maximize RLC by construction; (2) good metrics should be appropriately sensitive to semantic perturbation of rationales; and (3) good metrics should be robust to variation in the LM's task performance. Across three text classification datasets, we show that existing RLC metrics cannot satisfy all three FRAME axioms, since they are implemented via model pretraining which muddles the metric's signal. Then, we introduce a non-pretraining RLC metric that greatly outperforms baselines on (1) and (3), while performing competitively on (2). Finally, we discuss the limitations of using RLC to evaluate free-text rationales.

Shivam Sharma, Firoj Alam, Md. Shad Akhtar et al.

The automatic identification of harmful content online is of major concern for social media platforms, policymakers, and society. Researchers have studied textual, visual, and audio content, but typically in isolation. Yet, harmful content often combines multiple modalities, as in the case of memes, which are of particular interest due to their viral nature. With this in mind, here we offer a comprehensive survey with a focus on harmful memes. Based on a systematic analysis of recent literature, we first propose a new typology of harmful memes, and then we highlight and summarize the relevant state of the art. One interesting finding is that many types of harmful memes are not really studied, e.g., such featuring self-harm and extremism, partly due to the lack of suitable datasets. We further find that existing datasets mostly capture multi-class scenarios, which are not inclusive of the affective spectrum that memes can represent. Another observation is that memes can propagate globally through repackaging in different languages and that they can also be multilingual, blending different cultures. We conclude by highlighting several challenges related to multimodal semiotics, technological constraints, and non-trivial social engagement, and we present several open-ended aspects such as delineating online harm and empirically examining related frameworks and assistive interventions, which we believe will motivate and drive future research.

Brihi Joshi, Aaron Chan, Ziyi Liu et al. · meta-ai

By explaining how humans would solve a given task, human rationales can provide strong learning signal for neural language models (LMs). Explanation regularization (ER) aims to improve LM generalization by pushing the LM's machine rationales (Which input tokens did the LM focus on?) to align with human rationales (Which input tokens would humans focus on?). Though prior works primarily study ER via in-distribution (ID) evaluation, out-of-distribution (OOD) generalization is often more critical in real-world scenarios, yet ER's effect on OOD generalization has been underexplored. In this paper, we introduce ER-Test, a framework for evaluating ER models' OOD generalization along three dimensions: unseen dataset tests, contrast set tests, and functional tests. Using ER-Test, we extensively analyze how ER models' OOD generalization varies with different ER design choices. Across two tasks and six datasets, ER-Test shows that ER has little impact on ID performance but can yield large OOD performance gains. Also, we find that ER can improve OOD performance even with limited rationale supervision. ER-Test's results help demonstrate ER's utility and establish best practices for using ER effectively.

Aaron Mueller, Kanika Narang, Lambert Mathias et al. · meta-ai

Large language models show impressive results on few-shot NLP tasks. However, these models are memory and computation-intensive. Meta-training allows one to leverage smaller models for few-shot generalization in a domain-general and task-agnostic manner; however, these methods alone results in models that may not have sufficient parameterization or knowledge to adapt quickly to a large variety of tasks. To overcome this issue, we propose meta-training with demonstration retrieval, where we use a dense passage retriever to retrieve semantically similar labeled demonstrations to each example for more varied supervision. By separating external knowledge from model parameters, we can use meta-training to train parameter-efficient models that generalize well on a larger variety of tasks. We construct a meta-training set from UnifiedQA and CrossFit, and propose a demonstration bank based on UnifiedQA tasks. To our knowledge, our work is the first to combine retrieval with meta-training, to use DPR models to retrieve demonstrations, and to leverage demonstrations from many tasks simultaneously, rather than randomly sampling demonstrations from the training set of the target task. Our approach outperforms a variety of targeted parameter-efficient and retrieval-augmented few-shot methods on QA, NLI, and text classification tasks (including SQuAD, QNLI, and TREC). Our approach can be meta-trained and fine-tuned quickly on a single GPU.

Nan Wang, Qifan Wang, Yi-Chia Wang et al.

As language models become increasingly integrated into our digital lives, Personalized Text Generation (PTG) has emerged as a pivotal component with a wide range of applications. However, the bias inherent in user written text, often used for PTG model training, can inadvertently associate different levels of linguistic quality with users' protected attributes. The model can inherit the bias and perpetuate inequality in generating text w.r.t. users' protected attributes, leading to unfair treatment when serving users. In this work, we investigate fairness of PTG in the context of personalized explanation generation for recommendations. We first discuss the biases in generated explanations and their fairness implications. To promote fairness, we introduce a general framework to achieve measure-specific counterfactual fairness in explanation generation. Extensive experiments and human evaluations demonstrate the effectiveness of our method.

Vignesh Kothapalli, Ata Fatahibaarzi, Hamed Firooz et al.

Chain-of-thought (CoT) prompting combined with few-shot in-context learning (ICL) has unlocked significant reasoning capabilities in large language models (LLMs). However, ICL with CoT examples is ineffective on novel tasks when the pre-training knowledge is insufficient. We study this problem in a controlled setting using the CoT-ICL Lab framework, and propose meta-training techniques to learn novel abstract reasoning tasks in-context. Although CoT examples facilitate reasoning, we noticed that their excessive inclusion during meta-training degrades performance when CoT supervision is limited. To mitigate such behavior, we propose CoT-Recipe, a formal approach to modulate the mix of CoT and non-CoT examples in meta-training sequences. We demonstrate that careful modulation via CoT-Recipe can increase the accuracy of transformers on novel tasks by up to 300% even when there are no CoT examples available in-context. We confirm the broader effectiveness of these techniques by applying them to pretrained LLMs (Qwen2.5 series) for symbolic reasoning tasks and observing gains of up to 130% in accuracy.

Hamed Firooz, Rui Liu, Yuchen Lu et al.

Content moderation at scale remains one of the most pressing challenges in today's digital ecosystem, where billions of user- and AI-generated artifacts must be continuously evaluated for policy violations. Although recent advances in large language models (LLMs) have demonstrated strong potential for policy-grounded moderation, the practical challenges of training these systems to achieve expert-level accuracy in real-world settings remain largely unexplored, particularly in regimes characterized by label sparsity, evolving policy definitions, and the need for nuanced reasoning beyond shallow pattern matching. In this work, we present a comprehensive empirical investigation of scaling reinforcement learning (RL) for content classification, systematically evaluating multiple RL training recipes and reward-shaping strategies-including verifiable rewards and LLM-as-judge frameworks-to transform general-purpose language models into specialized, policy-aligned classifiers across three real-world content moderation tasks. Our findings provide actionable insights for industrial-scale moderation systems, demonstrating that RL exhibits sigmoid-like scaling behavior in which performance improves smoothly with increased training data, rollouts, and optimization steps before gradually saturating. Moreover, we show that RL substantially improves performance on tasks requiring complex policy-grounded reasoning while achieving up to 100x higher data efficiency than supervised fine-tuning, making it particularly effective in domains where expert annotations are scarce or costly.

Polina Kirichenko, Mark Ibrahim, Randall Balestriero et al.

Data augmentation (DA) encodes invariance and provides implicit regularization critical to a model's performance in image classification tasks. However, while DA improves average accuracy, recent studies have shown that its impact can be highly class dependent: achieving optimal average accuracy comes at the cost of significantly hurting individual class accuracy by as much as 20% on ImageNet. There has been little progress in resolving class-level accuracy drops due to a limited understanding of these effects. In this work, we present a framework for understanding how DA interacts with class-level learning dynamics. Using higher-quality multi-label annotations on ImageNet, we systematically categorize the affected classes and find that the majority are inherently ambiguous, co-occur, or involve fine-grained distinctions, while DA controls the model's bias towards one of the closely related classes. While many of the previously reported performance drops are explained by multi-label annotations, our analysis of class confusions reveals other sources of accuracy degradation. We show that simple class-conditional augmentation strategies informed by our framework improve performance on the negatively affected classes.

Hamed Firooz, Maziar Sanjabi, Adrian Englhardt et al.

Ranking and recommendation systems are the foundation for numerous online experiences, ranging from search results to personalized content delivery. These systems have evolved into complex, multilayered architectures that leverage vast datasets and often incorporate thousands of predictive models. The maintenance and enhancement of these models is a labor intensive process that requires extensive feature engineering. This approach not only exacerbates technical debt but also hampers innovation in extending these systems to emerging problem domains. In this report, we present our research to address these challenges by utilizing a large foundation model with a textual interface for ranking and recommendation tasks. We illustrate several key advantages of our approach: (1) a single model can manage multiple predictive tasks involved in ranking and recommendation, (2) decoder models with textual interface due to their comprehension of reasoning capabilities, can generalize to new recommendation surfaces and out-of-domain problems, and (3) by employing natural language interfaces for task definitions and verbalizing member behaviors and their social connections, we eliminate the need for feature engineering and the maintenance of complex directed acyclic graphs of model dependencies. We introduce our research pre-production model, 360Brew V1.0, a 150B parameter, decoder-only model that has been trained and fine-tuned on LinkedIn's data and tasks. This model is capable of solving over 30 predictive tasks across various segments of the LinkedIn platform, achieving performance levels comparable to or exceeding those of current production systems based on offline metrics, without task-specific fine-tuning. Notably, each of these tasks is conventionally addressed by dedicated models that have been developed and maintained over multiple years by teams of a similar or larger size than our own.

Sudarshan Srinivasa Ramanujam, Antonio Alonso, Saurabh Kataria et al.

In large scale recommendation systems like the LinkedIn Feed, the retrieval stage is critical for narrowing hundreds of millions of potential candidates to a manageable subset for ranking. LinkedIn's Feed serves suggested content from outside of the member's network (based on the member's topical interests), where 2000 candidates are retrieved from a pool of hundreds of millions candidate with a latency budget of a few milliseconds and inbound QPS of several thousand per second. This paper presents a novel retrieval approach that fine-tunes a large causal language model (Meta's LLaMA 3) as a dual encoder to generate high quality embeddings for both users (members) and content (items), using only textual input. We describe the end to end pipeline, including prompt design for embedding generation, techniques for fine-tuning at LinkedIn's scale, and infrastructure for low latency, cost effective online serving. We share our findings on how quantizing numerical features in the prompt enables the information to get properly encoded in the embedding, facilitating greater alignment between the retrieval and ranking layer. The system was evaluated using offline metrics and an online A/B test, which showed substantial improvements in member engagement. We observed significant gains among newer members, who often lack strong network connections, indicating that high-quality suggested content aids retention. This work demonstrates how generative language models can be effectively adapted for real time, high throughput retrieval in industrial applications.

Vignesh Kothapalli, Hamed Firooz, Maziar Sanjabi

We introduce CoT-ICL Lab, a framework and methodology to generate synthetic tokenized datasets and systematically study chain-of-thought (CoT) in-context learning (ICL) in language models. CoT-ICL Lab allows fine grained control over the complexity of in-context examples by decoupling (1) the causal structure involved in chain token generation from (2) the underlying token processing functions. We train decoder-only transformers (up to 700M parameters) on these datasets and show that CoT accelerates the accuracy transition to higher values across model sizes. In particular, we find that model depth is crucial for leveraging CoT with limited in-context examples, while more examples help shallow models match deeper model performance. Additionally, limiting the diversity of token processing functions throughout training improves causal structure learning via ICL. We also interpret these transitions by analyzing transformer embeddings and attention maps. Overall, CoT-ICL Lab serves as a simple yet powerful testbed for theoretical and empirical insights into ICL and CoT in language models.

Yun Dai, Tejas Dharamsi, Byron Hsu et al.

Training extremely large language models (LLMs) with billions of parameters is a computationally intensive task that pushes the limits of current data parallel training systems. While techniques like ZeRO++ have enabled efficient distributed training of such giant models on inexpensive low-bandwidth clusters, they can suffer from convergence issues due to potential race conditions in the hierarchical partitioning (hpZ) scheme employed to reduce cross-machine communication. In this work, we first show how these race conditions cause instability when training models with billions of parameters. We then propose a modification to the partitioning algorithm that addresses these convergence challenges while maintaining competitive training efficiency. Empirical evaluation on training the multi-billion parameters Falcon Models and Llama-2 models demonstrates the updated algorithm's ability to achieve reliable convergence on these massive models, where stock ZeRO++ hpZ fails to converge. The updated algorithm enables robust training of larger models with 98\% throughput and model training speed improvement without sacrificing the quality of convergence.

Nimit S. Sohoni, Maziar Sanjabi, Nicolas Ballas et al.

While neural networks have shown remarkable success on classification tasks in terms of average-case performance, they often fail to perform well on certain groups of the data. Such group information may be expensive to obtain; thus, recent works in robustness and fairness have proposed ways to improve worst-group performance even when group labels are unavailable for the training data. However, these methods generally underperform methods that utilize group information at training time. In this work, we assume access to a small number of group labels alongside a larger dataset without group labels. We propose BARACK, a simple two-step framework to utilize this partial group information to improve worst-group performance: train a model to predict the missing group labels for the training data, and then use these predicted group labels in a robust optimization objective. Theoretically, we provide generalization bounds for our approach in terms of the worst-group performance, which scale with respect to both the total number of training points and the number of training points with group labels. Empirically, our method outperforms the baselines that do not use group information, even when only 1-33% of points have group labels. We provide ablation studies to support the robustness and extensibility of our framework.

Ajinkya Tejankar, Maziar Sanjabi, Bichen Wu et al.

Using natural language as a supervision for training visual recognition models holds great promise. Recent works have shown that if such supervision is used in the form of alignment between images and captions in large training datasets, then the resulting aligned models perform well on zero-shot classification as downstream tasks2. In this paper, we focus on teasing out what parts of the language supervision are essential for training zero-shot image classification models. Through extensive and careful experiments, we show that: 1) A simple Bag-of-Words (BoW) caption could be used as a replacement for most of the image captions in the dataset. Surprisingly, we observe that this approach improves the zero-shot classification performance when combined with word balancing. 2) Using a BoW pretrained model, we can obtain more training data by generating pseudo-BoW captions on images that do not have a caption. Models trained on images with real and pseudo-BoW captions achieve stronger zero-shot performance. On ImageNet-1k zero-shot evaluation, our best model, that uses only 3M image-caption pairs, performs on-par with a CLIP model trained on 15M image-caption pairs (31.5% vs 31.3%).

Aaron Chan, Maziar Sanjabi, Lambert Mathias et al.

An extractive rationale explains a language model's (LM's) prediction on a given task instance by highlighting the text inputs that most influenced the prediction. Ideally, rationale extraction should be faithful (reflective of LM's actual behavior) and plausible (convincing to humans), without compromising the LM's (i.e., task model's) task performance. Although attribution algorithms and select-predict pipelines are commonly used in rationale extraction, they both rely on certain heuristics that hinder them from satisfying all three desiderata. In light of this, we propose UNIREX, a flexible learning framework that generalizes rationale extractor optimization as follows: (1) specify architecture for a learned rationale extractor; (2) select explainability objectives (i.e., faithfulness and plausibility criteria); and (3) jointly the train task model and rationale extractor on the task using the selected objectives. UNIREX enables replacing prior works' heuristic design choices with a generic learned rationale extractor in (1) and optimizing it for all three desiderata in (2)-(3). To facilitate comparison between methods with respect to multiple desiderata, we introduce the Normalized Relative Gain (NRG) metric. Across five text classification datasets, our best UNIREX configuration outperforms baselines by an average of 32.9% NRG. Plus, we find that UNIREX-trained rationale extractors can even generalize to unseen datasets and tasks.

Dimitar Dimitrov, Bishr Bin Ali, Shaden Shaar et al.

Propaganda can be defined as a form of communication that aims to influence the opinions or the actions of people towards a specific goal; this is achieved by means of well-defined rhetorical and psychological devices. Propaganda, in the form we know it today, can be dated back to the beginning of the 17th century. However, it is with the advent of the Internet and the social media that it has started to spread on a much larger scale than before, thus becoming major societal and political issue. Nowadays, a large fraction of propaganda in social media is multimodal, mixing textual with visual content. With this in mind, here we propose a new multi-label multimodal task: detecting the type of propaganda techniques used in memes. We further create and release a new corpus of 950 memes, carefully annotated with 22 propaganda techniques, which can appear in the text, in the image, or in both. Our analysis of the corpus shows that understanding both modalities together is essential for detecting these techniques. This is further confirmed in our experiments with several state-of-the-art multimodal models.

Dimitar Dimitrov, Bishr Bin Ali, Shaden Shaar et al.

We describe SemEval-2021 task 6 on Detection of Persuasion Techniques in Texts and Images: the data, the annotation guidelines, the evaluation setup, the results, and the participating systems. The task focused on memes and had three subtasks: (i) detecting the techniques in the text, (ii) detecting the text spans where the techniques are used, and (iii) detecting techniques in the entire meme, i.e., both in the text and in the image. It was a popular task, attracting 71 registrations, and 22 teams that eventually made an official submission on the test set. The evaluation results for the third subtask confirmed the importance of both modalities, the text and the image. Moreover, some teams reported benefits when not just combining the two modalities, e.g., by using early or late fusion, but rather modeling the interaction between them in a joint model.

Firoj Alam, Stefano Cresci, Tanmoy Chakraborty et al.

Recent years have witnessed the proliferation of offensive content online such as fake news, propaganda, misinformation, and disinformation. While initially this was mostly about textual content, over time images and videos gained popularity, as they are much easier to consume, attract more attention, and spread further than text. As a result, researchers started leveraging different modalities and combinations thereof to tackle online multimodal offensive content. In this study, we offer a survey on the state-of-the-art on multimodal disinformation detection covering various combinations of modalities: text, images, speech, video, social media network structure, and temporal information. Moreover, while some studies focused on factuality, others investigated how harmful the content is. While these two components in the definition of disinformation (i) factuality, and (ii) harmfulness, are equally important, they are typically studied in isolation. Thus, we argue for the need to tackle disinformation detection by taking into account multiple modalities as well as both factuality and harmfulness, in the same framework. Finally, we discuss current challenges and future research directions

Woojeong Jin, Maziar Sanjabi, Shaoliang Nie et al.

To reduce a model size but retain performance, we often rely on knowledge distillation (KD) which transfers knowledge from a large "teacher" model to a smaller "student" model. However, KD on multimodal datasets such as vision-language tasks is relatively unexplored, and digesting multimodal information is challenging since different modalities present different types of information. In this paper, we perform a large-scale empirical study to investigate the importance and effects of each modality in knowledge distillation. Furthermore, we introduce a multimodal knowledge distillation framework, modality-specific distillation (MSD), to transfer knowledge from a teacher on multimodal tasks by learning the teacher's behavior within each modality. The idea aims at mimicking a teacher's modality-specific predictions by introducing auxiliary loss terms for each modality. Furthermore, because each modality has different saliency for predictions, we define saliency scores for each modality and investigate saliency-based weighting schemes for the auxiliary losses. We further study a weight learning approach to learn the optimal weights on these loss terms. In our empirical analysis, we examine the saliency of each modality in KD, demonstrate the effectiveness of the weighting scheme in MSD, and show that it achieves better performance than KD on four multimodal datasets.

Ivan Evtimov, Russel Howes, Brian Dolhansky et al.

This work examines the vulnerability of multimodal (image + text) models to adversarial threats similar to those discussed in previous literature on unimodal (image- or text-only) models. We introduce realistic assumptions of partial model knowledge and access, and discuss how these assumptions differ from the standard "black-box"/"white-box" dichotomy common in current literature on adversarial attacks. Working under various levels of these "gray-box" assumptions, we develop new attack methodologies unique to multimodal classification and evaluate them on the Hateful Memes Challenge classification task. We find that attacking multiple modalities yields stronger attacks than unimodal attacks alone (inducing errors in up to 73% of cases), and that the unimodal image attacks on multimodal classifiers we explored were stronger than character-based text augmentation attacks (inducing errors on average in 45% and 30% of cases, respectively).

Douwe Kiela, Hamed Firooz, Aravind Mohan et al.

This work proposes a new challenge set for multimodal classification, focusing on detecting hate speech in multimodal memes. It is constructed such that unimodal models struggle and only multimodal models can succeed: difficult examples ("benign confounders") are added to the dataset to make it hard to rely on unimodal signals. The task requires subtle reasoning, yet is straightforward to evaluate as a binary classification problem. We provide baseline performance numbers for unimodal models, as well as for multimodal models with various degrees of sophistication. We find that state-of-the-art methods perform poorly compared to humans (64.73% vs. 84.7% accuracy), illustrating the difficulty of the task and highlighting the challenge that this important problem poses to the community.

Douwe Kiela, Suvrat Bhooshan, Hamed Firooz et al.

Self-supervised bidirectional transformer models such as BERT have led to dramatic improvements in a wide variety of textual classification tasks. The modern digital world is increasingly multimodal, however, and textual information is often accompanied by other modalities such as images. We introduce a supervised multimodal bitransformer model that fuses information from text and image encoders, and obtain state-of-the-art performance on various multimodal classification benchmark tasks, outperforming strong baselines, including on hard test sets specifically designed to measure multimodal performance.