Lisa Anne Hendricks

Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch et al.

We investigate the optimal model size and number of tokens for training a transformer language model under a given compute budget. We find that current large language models are significantly undertrained, a consequence of the recent focus on scaling language models whilst keeping the amount of training data constant. By training over 400 language models ranging from 70 million to over 16 billion parameters on 5 to 500 billion tokens, we find that for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled. We test this hypothesis by training a predicted compute-optimal model, Chinchilla, that uses the same compute budget as Gopher but with 70B parameters and 4$\times$ more more data. Chinchilla uniformly and significantly outperforms Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks. This also means that Chinchilla uses substantially less compute for fine-tuning and inference, greatly facilitating downstream usage. As a highlight, Chinchilla reaches a state-of-the-art average accuracy of 67.5% on the MMLU benchmark, greater than a 7% improvement over Gopher.

Amelia Glaese, Nat McAleese, Maja Trębacz et al.

We present Sparrow, an information-seeking dialogue agent trained to be more helpful, correct, and harmless compared to prompted language model baselines. We use reinforcement learning from human feedback to train our models with two new additions to help human raters judge agent behaviour. First, to make our agent more helpful and harmless, we break down the requirements for good dialogue into natural language rules the agent should follow, and ask raters about each rule separately. We demonstrate that this breakdown enables us to collect more targeted human judgements of agent behaviour and allows for more efficient rule-conditional reward models. Second, our agent provides evidence from sources supporting factual claims when collecting preference judgements over model statements. For factual questions, evidence provided by Sparrow supports the sampled response 78% of the time. Sparrow is preferred more often than baselines while being more resilient to adversarial probing by humans, violating our rules only 8% of the time when probed. Finally, we conduct extensive analyses showing that though our model learns to follow our rules it can exhibit distributional biases.

Shravan Nayak, Kanishk Jain, Rabiul Awal et al. · eth-zurich

Foundation models and vision-language pre-training have notably advanced Vision Language Models (VLMs), enabling multimodal processing of visual and linguistic data. However, their performance has been typically assessed on general scene understanding - recognizing objects, attributes, and actions - rather than cultural comprehension. This study introduces CulturalVQA, a visual question-answering benchmark aimed at assessing VLM's geo-diverse cultural understanding. We curate a collection of 2,378 image-question pairs with 1-5 answers per question representing cultures from 11 countries across 5 continents. The questions probe understanding of various facets of culture such as clothing, food, drinks, rituals, and traditions. Benchmarking VLMs on CulturalVQA, including GPT-4V and Gemini, reveals disparity in their level of cultural understanding across regions, with strong cultural understanding capabilities for North America while significantly lower performance for Africa. We observe disparity in their performance across cultural facets too, with clothing, rituals, and traditions seeing higher performances than food and drink. These disparities help us identify areas where VLMs lack cultural understanding and demonstrate the potential of CulturalVQA as a comprehensive evaluation set for gauging VLM progress in understanding diverse cultures.

Laura Weidinger, Maribeth Rauh, Nahema Marchal et al.

Generative AI systems produce a range of risks. To ensure the safety of generative AI systems, these risks must be evaluated. In this paper, we make two main contributions toward establishing such evaluations. First, we propose a three-layered framework that takes a structured, sociotechnical approach to evaluating these risks. This framework encompasses capability evaluations, which are the main current approach to safety evaluation. It then reaches further by building on system safety principles, particularly the insight that context determines whether a given capability may cause harm. To account for relevant context, our framework adds human interaction and systemic impacts as additional layers of evaluation. Second, we survey the current state of safety evaluation of generative AI systems and create a repository of existing evaluations. Three salient evaluation gaps emerge from this analysis. We propose ways forward to closing these gaps, outlining practical steps as well as roles and responsibilities for different actors. Sociotechnical safety evaluation is a tractable approach to the robust and comprehensive safety evaluation of generative AI systems.

Maribeth Rauh, John Mellor, Jonathan Uesato et al.

Large language models produce human-like text that drive a growing number of applications. However, recent literature and, increasingly, real world observations, have demonstrated that these models can generate language that is toxic, biased, untruthful or otherwise harmful. Though work to evaluate language model harms is under way, translating foresight about which harms may arise into rigorous benchmarks is not straightforward. To facilitate this translation, we outline six ways of characterizing harmful text which merit explicit consideration when designing new benchmarks. We then use these characteristics as a lens to identify trends and gaps in existing benchmarks. Finally, we apply them in a case study of the Perspective API, a toxicity classifier that is widely used in harm benchmarks. Our characteristics provide one piece of the bridge that translates between foresight and effective evaluation.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Laura Weidinger, Joslyn Barnhart, Jenny Brennan et al.

Safety and responsibility evaluations of advanced AI models are a critical but developing field of research and practice. In the development of Google DeepMind's advanced AI models, we innovated on and applied a broad set of approaches to safety evaluation. In this report, we summarise and share elements of our evolving approach as well as lessons learned for a broad audience. Key lessons learned include: First, theoretical underpinnings and frameworks are invaluable to organise the breadth of risk domains, modalities, forms, metrics, and goals. Second, theory and practice of safety evaluation development each benefit from collaboration to clarify goals, methods and challenges, and facilitate the transfer of insights between different stakeholders and disciplines. Third, similar key methods, lessons, and institutions apply across the range of concerns in responsibility and safety - including established and emerging harms. For this reason it is important that a wide range of actors working on safety evaluation and safety research communities work together to develop, refine and implement novel evaluation approaches and best practices, rather than operating in silos. The report concludes with outlining the clear need to rapidly advance the science of evaluations, to integrate new evaluations into the development and governance of AI, to establish scientifically-grounded norms and standards, and to promote a robust evaluation ecosystem.

Shravan Nayak, Mehar Bhatia, Xiaofeng Zhang et al. · eth-zurich

The increasing ubiquity of text-to-image (T2I) models as tools for visual content generation raises concerns about their ability to accurately represent diverse cultural contexts -- where missed cues can stereotype communities and undermine usability. In this work, we present the first study to systematically quantify the alignment of T2I models and evaluation metrics with respect to both explicit (stated) as well as implicit (unstated, implied by the prompt's cultural context) cultural expectations. To this end, we introduce CulturalFrames, a novel benchmark designed for rigorous human evaluation of cultural representation in visual generations. Spanning 10 countries and 5 socio-cultural domains, CulturalFrames comprises 983 prompts, 3637 corresponding images generated by 4 state-of-the-art T2I models, and over 10k detailed human annotations. We find that across models and countries, cultural expectations are missed an average of 44% of the time. Among these failures, explicit expectations are missed at a surprisingly high average rate of 68%, while implicit expectation failures are also significant, averaging 49%. Furthermore, we show that existing T2I evaluation metrics correlate poorly with human judgments of cultural alignment, irrespective of their internal reasoning. Collectively, our findings expose critical gaps, provide a concrete testbed, and outline actionable directions for developing culturally informed T2I models and metrics that improve global usability.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Emanuele Bugliarello, Aida Nematzadeh, Lisa Anne Hendricks

Recent work in vision-and-language pretraining has investigated supervised signals from object detection data to learn better, fine-grained multimodal representations. In this work, we take a step further and explore how we can tap into supervision from small-scale visual relation data. In particular, we propose two pretraining approaches to contextualise visual entities in a multimodal setup. With verbalised scene graphs, we transform visual relation triplets into structured captions, and treat them as additional image descriptions. With masked relation prediction, we further encourage relating entities from image regions with visually masked contexts. When applied to strong baselines pretrained on large amounts of Web data, zero-shot evaluations on both coarse-grained and fine-grained tasks show the efficacy of our methods in learning multimodal representations from weakly-supervised relations data.

Emanuele Bugliarello, Laurent Sartran, Aishwarya Agrawal et al.

While pretraining on large-scale image-text data from the Web has facilitated rapid progress on many vision-and-language (V&L) tasks, recent work has demonstrated that pretrained models lack "fine-grained" understanding, such as the ability to recognise relationships, verbs, and numbers in images. This has resulted in an increased interest in the community to either develop new benchmarks or models for such capabilities. To better understand and quantify progress in this direction, we investigate four competitive V&L models on four fine-grained benchmarks. Through our analysis, we find that X-VLM (Zeng et al., 2022) consistently outperforms other baselines, and that modelling innovations can impact performance more than scaling Web data, which even degrades performance sometimes. Through a deeper investigation of X-VLM, we highlight the importance of both novel losses and rich data sources for learning fine-grained skills. Finally, we inspect training dynamics, and discover that for some tasks, performance peaks early in training or significantly fluctuates, never converging.

Jack W. Rae, Sebastian Borgeaud, Trevor Cai et al.

Language modelling provides a step towards intelligent communication systems by harnessing large repositories of written human knowledge to better predict and understand the world. In this paper, we present an analysis of Transformer-based language model performance across a wide range of model scales -- from models with tens of millions of parameters up to a 280 billion parameter model called Gopher. These models are evaluated on 152 diverse tasks, achieving state-of-the-art performance across the majority. Gains from scale are largest in areas such as reading comprehension, fact-checking, and the identification of toxic language, but logical and mathematical reasoning see less benefit. We provide a holistic analysis of the training dataset and model's behaviour, covering the intersection of model scale with bias and toxicity. Finally we discuss the application of language models to AI safety and the mitigation of downstream harms.

Laura Weidinger, John Mellor, Maribeth Rauh et al.

This paper aims to help structure the risk landscape associated with large-scale Language Models (LMs). In order to foster advances in responsible innovation, an in-depth understanding of the potential risks posed by these models is needed. A wide range of established and anticipated risks are analysed in detail, drawing on multidisciplinary expertise and literature from computer science, linguistics, and social sciences. We outline six specific risk areas: I. Discrimination, Exclusion and Toxicity, II. Information Hazards, III. Misinformation Harms, V. Malicious Uses, V. Human-Computer Interaction Harms, VI. Automation, Access, and Environmental Harms. The first area concerns the perpetuation of stereotypes, unfair discrimination, exclusionary norms, toxic language, and lower performance by social group for LMs. The second focuses on risks from private data leaks or LMs correctly inferring sensitive information. The third addresses risks arising from poor, false or misleading information including in sensitive domains, and knock-on risks such as the erosion of trust in shared information. The fourth considers risks from actors who try to use LMs to cause harm. The fifth focuses on risks specific to LLMs used to underpin conversational agents that interact with human users, including unsafe use, manipulation or deception. The sixth discusses the risk of environmental harm, job automation, and other challenges that may have a disparate effect on different social groups or communities. In total, we review 21 risks in-depth. We discuss the points of origin of different risks and point to potential mitigation approaches. Lastly, we discuss organisational responsibilities in implementing mitigations, and the role of collaboration and participation. We highlight directions for further research, particularly on expanding the toolkit for assessing and evaluating the outlined risks in LMs.

Johannes Welbl, Amelia Glaese, Jonathan Uesato et al.

Large language models (LM) generate remarkably fluent text and can be efficiently adapted across NLP tasks. Measuring and guaranteeing the quality of generated text in terms of safety is imperative for deploying LMs in the real world; to this end, prior work often relies on automatic evaluation of LM toxicity. We critically discuss this approach, evaluate several toxicity mitigation strategies with respect to both automatic and human evaluation, and analyze consequences of toxicity mitigation in terms of model bias and LM quality. We demonstrate that while basic intervention strategies can effectively optimize previously established automatic metrics on the RealToxicityPrompts dataset, this comes at the cost of reduced LM coverage for both texts about, and dialects of, marginalized groups. Additionally, we find that human raters often disagree with high automatic toxicity scores after strong toxicity reduction interventions -- highlighting further the nuances involved in careful evaluation of LM toxicity.

Lisa Anne Hendricks, Aida Nematzadeh

Multimodal image-language transformers have achieved impressive results on a variety of tasks that rely on fine-tuning (e.g., visual question answering and image retrieval). We are interested in shedding light on the quality of their pretrained representations -- in particular, if these models can distinguish different types of verbs or if they rely solely on nouns in a given sentence. To do so, we collect a dataset of image-sentence pairs (in English) consisting of 421 verbs that are either visual or commonly found in the pretraining data (i.e., the Conceptual Captions dataset). We use this dataset to evaluate pretrained image-language transformers and find that they fail more in situations that require verb understanding compared to other parts of speech. We also investigate what category of verbs are particularly challenging.

Lisa Anne Hendricks, John Mellor, Rosalia Schneider et al.

Recently multimodal transformer models have gained popularity because their performance on language and vision tasks suggest they learn rich visual-linguistic representations. Focusing on zero-shot image retrieval tasks, we study three important factors which can impact the quality of learned representations: pretraining data, the attention mechanism, and loss functions. By pretraining models on six datasets, we observe that dataset noise and language similarity to our downstream task are important indicators of model performance. Through architectural analysis, we learn that models with a multimodal attention mechanism can outperform deeper models with modality specific attention mechanisms. Finally, we show that successful contrastive losses used in the self-supervised learning literature do not yield similar performance gains when used in multimodal transformers

Viktoriia Sharmanska, Lisa Anne Hendricks, Trevor Darrell et al.

Computer vision algorithms, e.g. for face recognition, favour groups of individuals that are better represented in the training data. This happens because of the generalization that classifiers have to make. It is simpler to fit the majority groups as this fit is more important to overall error. We propose to create a balanced training dataset, consisting of the original dataset plus new data points in which the group memberships are intervened, minorities become majorities and vice versa. We show that current generative adversarial networks are a powerful tool for learning these data points, called contrastive examples. We experiment with the equalized odds bias measure on tabular data as well as image data (CelebA and Diversity in Faces datasets). Contrastive examples allow us to expose correlations between group membership and other seemingly neutral features. Whenever a causal graph is available, we can put those contrastive examples in the perspective of counterfactuals.

Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns et al.

Despite continuously improving performance, contemporary image captioning models are prone to "hallucinating" objects that are not actually in a scene. One problem is that standard metrics only measure similarity to ground truth captions and may not fully capture image relevance. In this work, we propose a new image relevance metric to evaluate current models with veridical visual labels and assess their rate of object hallucination. We analyze how captioning model architectures and learning objectives contribute to object hallucination, explore when hallucination is likely due to image misclassification or language priors, and assess how well current sentence metrics capture object hallucination. We investigate these questions on the standard image captioning benchmark, MSCOCO, using a diverse set of models. Our analysis yields several interesting findings, including that models which score best on standard sentence metrics do not always have lower hallucination and that models which hallucinate more tend to make errors driven by language priors.

Lisa Anne Hendricks, Oliver Wang, Eli Shechtman et al.

Localizing moments in a longer video via natural language queries is a new, challenging task at the intersection of language and video understanding. Though moment localization with natural language is similar to other language and vision tasks like natural language object retrieval in images, moment localization offers an interesting opportunity to model temporal dependencies and reasoning in text. We propose a new model that explicitly reasons about different temporal segments in a video, and shows that temporal context is important for localizing phrases which include temporal language. To benchmark whether our model, and other recent video localization models, can effectively reason about temporal language, we collect the novel TEMPOral reasoning in video and language (TEMPO) dataset. Our dataset consists of two parts: a dataset with real videos and template sentences (TEMPO - Template Language) which allows for controlled studies on temporal language, and a human language dataset which consists of temporal sentences annotated by humans (TEMPO - Human Language).

Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell et al.

Existing visual explanation generating agents learn to fluently justify a class prediction. However, they may mention visual attributes which reflect a strong class prior, although the evidence may not actually be in the image. This is particularly concerning as ultimately such agents fail in building trust with human users. To overcome this limitation, we propose a phrase-critic model to refine generated candidate explanations augmented with flipped phrases which we use as negative examples while training. At inference time, our phrase-critic model takes an image and a candidate explanation as input and outputs a score indicating how well the candidate explanation is grounded in the image. Our explainable AI agent is capable of providing counter arguments for an alternative prediction, i.e. counterfactuals, along with explanations that justify the correct classification decisions. Our model improves the textual explanation quality of fine-grained classification decisions on the CUB dataset by mentioning phrases that are grounded in the image. Moreover, on the FOIL tasks, our agent detects when there is a mistake in the sentence, grounds the incorrect phrase and corrects it significantly better than other models.

Lisa Anne Hendricks, Kaylee Burns, Kate Saenko et al.

Most machine learning methods are known to capture and exploit biases of the training data. While some biases are beneficial for learning, others are harmful. Specifically, image captioning models tend to exaggerate biases present in training data. This can lead to incorrect captions in domains where unbiased captions are desired, or required, due to over reliance on the learned prior and image context. We investigate generation of gender specific caption words (e.g. man, woman) based on the person's appearance or the image context. We introduce a new Equalizer model that ensures equal gender probability when gender evidence is occluded in a scene and confident predictions when gender evidence is present. The resulting model is forced to look at a person rather than use contextual cues to make a gender specific prediction. The losses that comprise our model, the Appearance Confusion Loss and the Confident Loss, are general, and can be added to any description model in order to mitigate impacts of unwanted bias in a description dataset. Our proposed model has lower error than prior work when describing images with people and mentioning their gender and more closely matches the ground truth ratio of sentences including women to sentences including men.

Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell et al.

Natural language explanations of deep neural network decisions provide an intuitive way for a AI agent to articulate a reasoning process. Current textual explanations learn to discuss class discriminative features in an image. However, it is also helpful to understand which attributes might change a classification decision if present in an image (e.g., "This is not a Scarlet Tanager because it does not have black wings.") We call such textual explanations counterfactual explanations, and propose an intuitive method to generate counterfactual explanations by inspecting which evidence in an input is missing, but might contribute to a different classification decision if present in the image. To demonstrate our method we consider a fine-grained image classification task in which we take as input an image and a counterfactual class and output text which explains why the image does not belong to a counterfactual class. We then analyze our generated counterfactual explanations both qualitatively and quantitatively using proposed automatic metrics.

Kaylee Burns, Lisa Anne Hendricks, Kate Saenko et al.

Most machine learning methods are known to capture and exploit biases of the training data. While some biases are beneficial for learning, others are harmful. Specifically, image captioning models tend to exaggerate biases present in training data (e.g., if a word is present in 60% of training sentences, it might be predicted in 70% of sentences at test time). This can lead to incorrect captions in domains where unbiased captions are desired, or required, due to over-reliance on the learned prior and image context. In this work we investigate generation of gender-specific caption words (e.g. man, woman) based on the person's appearance or the image context. We introduce a new Equalizer model that ensures equal gender probability when gender evidence is occluded in a scene and confident predictions when gender evidence is present. The resulting model is forced to look at a person rather than use contextual cues to make a gender-specific predictions. The losses that comprise our model, the Appearance Confusion Loss and the Confident Loss, are general, and can be added to any description model in order to mitigate impacts of unwanted bias in a description dataset. Our proposed model has lower error than prior work when describing images with people and mentioning their gender and more closely matches the ground truth ratio of sentences including women to sentences including men. We also show that unlike other approaches, our model is indeed more often looking at people when predicting their gender.

Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata et al.

Deep models that are both effective and explainable are desirable in many settings; prior explainable models have been unimodal, offering either image-based visualization of attention weights or text-based generation of post-hoc justifications. We propose a multimodal approach to explanation, and argue that the two modalities provide complementary explanatory strengths. We collect two new datasets to define and evaluate this task, and propose a novel model which can provide joint textual rationale generation and attention visualization. Our datasets define visual and textual justifications of a classification decision for activity recognition tasks (ACT-X) and for visual question answering tasks (VQA-X). We quantitatively show that training with the textual explanations not only yields better textual justification models, but also better localizes the evidence that supports the decision. We also qualitatively show cases where visual explanation is more insightful than textual explanation, and vice versa, supporting our thesis that multimodal explanation models offer significant benefits over unimodal approaches.

Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata et al.

Deep models are the defacto standard in visual decision problems due to their impressive performance on a wide array of visual tasks. On the other hand, their opaqueness has led to a surge of interest in explainable systems. In this work, we emphasize the importance of model explanation in various forms such as visual pointing and textual justification. The lack of data with justification annotations is one of the bottlenecks of generating multimodal explanations. Thus, we propose two large-scale datasets with annotations that visually and textually justify a classification decision for various activities, i.e. ACT-X, and for question answering, i.e. VQA-X. We also introduce a multimodal methodology for generating visual and textual explanations simultaneously. We quantitatively show that training with the textual explanations not only yields better textual justification models, but also models that better localize the evidence that support their decision.

Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell et al.

Existing models which generate textual explanations enforce task relevance through a discriminative term loss function, but such mechanisms only weakly constrain mentioned object parts to actually be present in the image. In this paper, a new model is proposed for generating explanations by utilizing localized grounding of constituent phrases in generated explanations to ensure image relevance. Specifically, we introduce a phrase-critic model to refine (re-score/re-rank) generated candidate explanations and employ a relative-attribute inspired ranking loss using "flipped" phrases as negative examples for training. At test time, our phrase-critic model takes an image and a candidate explanation as input and outputs a score indicating how well the candidate explanation is grounded in the image.

Lisa Anne Hendricks, Oliver Wang, Eli Shechtman et al.

We consider retrieving a specific temporal segment, or moment, from a video given a natural language text description. Methods designed to retrieve whole video clips with natural language determine what occurs in a video but not when. To address this issue, we propose the Moment Context Network (MCN) which effectively localizes natural language queries in videos by integrating local and global video features over time. A key obstacle to training our MCN model is that current video datasets do not include pairs of localized video segments and referring expressions, or text descriptions which uniquely identify a corresponding moment. Therefore, we collect the Distinct Describable Moments (DiDeMo) dataset which consists of over 10,000 unedited, personal videos in diverse visual settings with pairs of localized video segments and referring expressions. We demonstrate that MCN outperforms several baseline methods and believe that our initial results together with the release of DiDeMo will inspire further research on localizing video moments with natural language.

Rakshith Shetty, Marcus Rohrbach, Lisa Anne Hendricks et al.

While strong progress has been made in image captioning over the last years, machine and human captions are still quite distinct. A closer look reveals that this is due to the deficiencies in the generated word distribution, vocabulary size, and strong bias in the generators towards frequent captions. Furthermore, humans -- rightfully so -- generate multiple, diverse captions, due to the inherent ambiguity in the captioning task which is not considered in today's systems. To address these challenges, we change the training objective of the caption generator from reproducing groundtruth captions to generating a set of captions that is indistinguishable from human generated captions. Instead of handcrafting such a learning target, we employ adversarial training in combination with an approximate Gumbel sampler to implicitly match the generated distribution to the human one. While our method achieves comparable performance to the state-of-the-art in terms of the correctness of the captions, we generate a set of diverse captions, that are significantly less biased and match the word statistics better in several aspects.

Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata et al.

Deep models are the defacto standard in visual decision models due to their impressive performance on a wide array of visual tasks. However, they are frequently seen as opaque and are unable to explain their decisions. In contrast, humans can justify their decisions with natural language and point to the evidence in the visual world which led to their decisions. We postulate that deep models can do this as well and propose our Pointing and Justification (PJ-X) model which can justify its decision with a sentence and point to the evidence by introspecting its decision and explanation process using an attention mechanism. Unfortunately there is no dataset available with reference explanations for visual decision making. We thus collect two datasets in two domains where it is interesting and challenging to explain decisions. First, we extend the visual question answering task to not only provide an answer but also a natural language explanation for the answer. Second, we focus on explaining human activities which is traditionally more challenging than object classification. We extensively evaluate our PJ-X model, both on the justification and pointing tasks, by comparing it to prior models and ablations using both automatic and human evaluations.

Subhashini Venugopalan, Lisa Anne Hendricks, Marcus Rohrbach et al.

Recent captioning models are limited in their ability to scale and describe concepts unseen in paired image-text corpora. We propose the Novel Object Captioner (NOC), a deep visual semantic captioning model that can describe a large number of object categories not present in existing image-caption datasets. Our model takes advantage of external sources -- labeled images from object recognition datasets, and semantic knowledge extracted from unannotated text. We propose minimizing a joint objective which can learn from these diverse data sources and leverage distributional semantic embeddings, enabling the model to generalize and describe novel objects outside of image-caption datasets. We demonstrate that our model exploits semantic information to generate captions for hundreds of object categories in the ImageNet object recognition dataset that are not observed in MSCOCO image-caption training data, as well as many categories that are observed very rarely. Both automatic evaluations and human judgements show that our model considerably outperforms prior work in being able to describe many more categories of objects.

Subhashini Venugopalan, Lisa Anne Hendricks, Raymond Mooney et al.

This paper investigates how linguistic knowledge mined from large text corpora can aid the generation of natural language descriptions of videos. Specifically, we integrate both a neural language model and distributional semantics trained on large text corpora into a recent LSTM-based architecture for video description. We evaluate our approach on a collection of Youtube videos as well as two large movie description datasets showing significant improvements in grammaticality while modestly improving descriptive quality.

Lisa Anne Hendricks, Zeynep Akata, Marcus Rohrbach et al.

Clearly explaining a rationale for a classification decision to an end-user can be as important as the decision itself. Existing approaches for deep visual recognition are generally opaque and do not output any justification text; contemporary vision-language models can describe image content but fail to take into account class-discriminative image aspects which justify visual predictions. We propose a new model that focuses on the discriminating properties of the visible object, jointly predicts a class label, and explains why the predicted label is appropriate for the image. We propose a novel loss function based on sampling and reinforcement learning that learns to generate sentences that realize a global sentence property, such as class specificity. Our results on a fine-grained bird species classification dataset show that our model is able to generate explanations which are not only consistent with an image but also more discriminative than descriptions produced by existing captioning methods.

Yang Gao, Lisa Anne Hendricks, Katherine J. Kuchenbecker et al.

Robots which interact with the physical world will benefit from a fine-grained tactile understanding of objects and surfaces. Additionally, for certain tasks, robots may need to know the haptic properties of an object before touching it. To enable better tactile understanding for robots, we propose a method of classifying surfaces with haptic adjectives (e.g., compressible or smooth) from both visual and physical interaction data. Humans typically combine visual predictions and feedback from physical interactions to accurately predict haptic properties and interact with the world. Inspired by this cognitive pattern, we propose and explore a purely visual haptic prediction model. Purely visual models enable a robot to "feel" without physical interaction. Furthermore, we demonstrate that using both visual and physical interaction signals together yields more accurate haptic classification. Our models take advantage of recent advances in deep neural networks by employing a unified approach to learning features for physical interaction and visual observations. Even though we employ little domain specific knowledge, our model still achieves better results than methods based on hand-designed features.

Lisa Anne Hendricks, Subhashini Venugopalan, Marcus Rohrbach et al.

While recent deep neural network models have achieved promising results on the image captioning task, they rely largely on the availability of corpora with paired image and sentence captions to describe objects in context. In this work, we propose the Deep Compositional Captioner (DCC) to address the task of generating descriptions of novel objects which are not present in paired image-sentence datasets. Our method achieves this by leveraging large object recognition datasets and external text corpora and by transferring knowledge between semantically similar concepts. Current deep caption models can only describe objects contained in paired image-sentence corpora, despite the fact that they are pre-trained with large object recognition datasets, namely ImageNet. In contrast, our model can compose sentences that describe novel objects and their interactions with other objects. We demonstrate our model's ability to describe novel concepts by empirically evaluating its performance on MSCOCO and show qualitative results on ImageNet images of objects for which no paired image-caption data exist. Further, we extend our approach to generate descriptions of objects in video clips. Our results show that DCC has distinct advantages over existing image and video captioning approaches for generating descriptions of new objects in context.

Chelsea Finn, Lisa Anne Hendricks, Trevor Darrell

Recently, nested dropout was proposed as a method for ordering representation units in autoencoders by their information content, without diminishing reconstruction cost. However, it has only been applied to training fully-connected autoencoders in an unsupervised setting. We explore the impact of nested dropout on the convolutional layers in a CNN trained by backpropagation, investigating whether nested dropout can provide a simple and systematic way to determine the optimal representation size with respect to the desired accuracy and desired task and data complexity.

Jeff Donahue, Lisa Anne Hendricks, Marcus Rohrbach et al.

Models based on deep convolutional networks have dominated recent image interpretation tasks; we investigate whether models which are also recurrent, or "temporally deep", are effective for tasks involving sequences, visual and otherwise. We develop a novel recurrent convolutional architecture suitable for large-scale visual learning which is end-to-end trainable, and demonstrate the value of these models on benchmark video recognition tasks, image description and retrieval problems, and video narration challenges. In contrast to current models which assume a fixed spatio-temporal receptive field or simple temporal averaging for sequential processing, recurrent convolutional models are "doubly deep"' in that they can be compositional in spatial and temporal "layers". Such models may have advantages when target concepts are complex and/or training data are limited. Learning long-term dependencies is possible when nonlinearities are incorporated into the network state updates. Long-term RNN models are appealing in that they directly can map variable-length inputs (e.g., video frames) to variable length outputs (e.g., natural language text) and can model complex temporal dynamics; yet they can be optimized with backpropagation. Our recurrent long-term models are directly connected to modern visual convnet models and can be jointly trained to simultaneously learn temporal dynamics and convolutional perceptual representations. Our results show such models have distinct advantages over state-of-the-art models for recognition or generation which are separately defined and/or optimized.