Subhashini Venugopalan

Joel Shor, Subhashini Venugopalan

Recent advances in self-supervision have dramatically improved the quality of speech representations. However, deployment of state-of-the-art embedding models on devices has been restricted due to their limited public availability and large resource footprint. Our work addresses these issues by publicly releasing a collection of paralinguistic speech models that are small and near state-of-the-art performance. Our approach is based on knowledge distillation, and our models are distilled on public data only. We explore different architectures and thoroughly evaluate our models on the Non-Semantic Speech (NOSS) benchmark. Our largest distilled model is less than 15% the size of the original model (314MB vs 2.2GB), achieves over 96% the accuracy on 6 of 7 tasks, and is trained on 6.5% the data. The smallest model is 1% in size (22MB) and achieves over 90% the accuracy on 6 of 7 tasks. Our models outperform the open source Wav2Vec 2.0 model on 6 of 7 tasks, and our smallest model outperforms the open source Wav2Vec 2.0 on both emotion recognition tasks despite being 7% the size.

Mukund Varma T, Peihao Wang, Xuxi Chen et al.

We present Generalizable NeRF Transformer (GNT), a transformer-based architecture that reconstructs Neural Radiance Fields (NeRFs) and learns to renders novel views on the fly from source views. While prior works on NeRFs optimize a scene representation by inverting a handcrafted rendering equation, GNT achieves neural representation and rendering that generalizes across scenes using transformers at two stages. (1) The view transformer leverages multi-view geometry as an inductive bias for attention-based scene representation, and predicts coordinate-aligned features by aggregating information from epipolar lines on the neighboring views. (2) The ray transformer renders novel views using attention to decode the features from the view transformer along the sampled points during ray marching. Our experiments demonstrate that when optimized on a single scene, GNT can successfully reconstruct NeRF without an explicit rendering formula due to the learned ray renderer. When trained on multiple scenes, GNT consistently achieves state-of-the-art performance when transferring to unseen scenes and outperform all other methods by ~10% on average. Our analysis of the learned attention maps to infer depth and occlusion indicate that attention enables learning a physically-grounded rendering. Our results show the promise of transformers as a universal modeling tool for graphics. Please refer to our project page for video results: https://vita-group.github.io/GNT/.

Shanqing Cai, Subhashini Venugopalan, Katrin Tomanek et al.

Motivated by the need for accelerating text entry in augmentative and alternative communication (AAC) for people with severe motor impairments, we propose a paradigm in which phrases are abbreviated aggressively as primarily word-initial letters. Our approach is to expand the abbreviations into full-phrase options by leveraging conversation context with the power of pretrained large language models (LLMs). Through zero-shot, few-shot, and fine-tuning experiments on four public conversation datasets, we show that for replies to the initial turn of a dialog, an LLM with 64B parameters is able to exactly expand over 70% of phrases with abbreviation length up to 10, leading to an effective keystroke saving rate of up to about 77% on these exact expansions. Including a small amount of context in the form of a single conversation turn more than doubles abbreviation expansion accuracies compared to having no context, an effect that is more pronounced for longer phrases. Additionally, the robustness of models against typo noise can be enhanced through fine-tuning on noisy data.

Shraman Pramanick, Rama Chellappa, Subhashini Venugopalan

Seeking answers to questions within long scientific research articles is a crucial area of study that aids readers in quickly addressing their inquiries. However, existing question-answering (QA) datasets based on scientific papers are limited in scale and focus solely on textual content. We introduce SPIQA (Scientific Paper Image Question Answering), the first large-scale QA dataset specifically designed to interpret complex figures and tables within the context of scientific research articles across various domains of computer science. Leveraging the breadth of expertise and ability of multimodal large language models (MLLMs) to understand figures, we employ automatic and manual curation to create the dataset. We craft an information-seeking task on interleaved images and text that involves multiple images covering plots, charts, tables, schematic diagrams, and result visualizations. SPIQA comprises 270K questions divided into training, validation, and three different evaluation splits. Through extensive experiments with 12 prominent foundational models, we evaluate the ability of current multimodal systems to comprehend the nuanced aspects of research articles. Additionally, we propose a Chain-of-Thought (CoT) evaluation strategy with in-context retrieval that allows fine-grained, step-by-step assessment and improves model performance. We further explore the upper bounds of performance enhancement with additional textual information, highlighting its promising potential for future research and the dataset's impact on revolutionizing how we interact with scientific literature.

Jimmy Tobin, Qisheng Li, Subhashini Venugopalan et al.

Word Error Rate (WER) is the primary metric used to assess automatic speech recognition (ASR) model quality. It has been shown that ASR models tend to have much higher WER on speakers with speech impairments than typical English speakers. It is hard to determine if models can be be useful at such high error rates. This study investigates the use of BERTScore, an evaluation metric for text generation, to provide a more informative measure of ASR model quality and usefulness. Both BERTScore and WER were compared to prediction errors manually annotated by Speech Language Pathologists for error type and assessment. BERTScore was found to be more correlated with human assessment of error type and assessment. BERTScore was specifically more robust to orthographic changes (contraction and normalization errors) where meaning was preserved. Furthermore, BERTScore was a better fit of error assessment than WER, as measured using an ordinal logistic regression and the Akaike's Information Criterion (AIC). Overall, our findings suggest that BERTScore can complement WER when assessing ASR model performance from a practical perspective, especially for accessibility applications where models are useful even at lower accuracy than for typical speech.

Haoyu Guo, Maria Tikhanovskaya, Paul Raccuglia et al.

Large Language Models (LLMs) show great promise as a powerful tool for scientific literature exploration. However, their effectiveness in providing scientifically accurate and comprehensive answers to complex questions within specialized domains remains an active area of research. Using the field of high-temperature cuprates as an exemplar, we evaluate the ability of LLM systems to understand the literature at the level of an expert. We construct an expert-curated database of 1,726 scientific papers that covers the history of the field, and a set of 67 expert-formulated questions that probe deep understanding of the literature. We then evaluate six different LLM-based systems for answering these questions, including both commercially available closed models and a custom retrieval-augmented generation (RAG) system capable of retrieving images alongside text. Experts then evaluate the answers of these systems against a rubric that assesses balanced perspectives, factual comprehensiveness, succinctness, and evidentiary support. Among the six systems two using RAG on curated literature outperformed existing closed models across key metrics, particularly in providing comprehensive and well-supported answers. We discuss promising aspects of LLM performances as well as critical short-comings of all the models. The set of expert-formulated questions and the rubric will be valuable for assessing expert level performance of LLM based reasoning systems.

Joel Shor, Ruyue Agnes Bi, Subhashini Venugopalan et al.

Automatic Speech Recognition (ASR) in medical contexts has the potential to save time, cut costs, increase report accuracy, and reduce physician burnout. However, the healthcare industry has been slower to adopt this technology, in part due to the importance of avoiding medically-relevant transcription mistakes. In this work, we present the Clinical BERTScore (CBERTScore), an ASR metric that penalizes clinically-relevant mistakes more than others. We demonstrate that this metric more closely aligns with clinician preferences on medical sentences as compared to other metrics (WER, BLUE, METEOR, etc), sometimes by wide margins. We collect a benchmark of 18 clinician preferences on 149 realistic medical sentences called the Clinician Transcript Preference benchmark (CTP) and make it publicly available for the community to further develop clinically-aware ASR metrics. To our knowledge, this is the first public dataset of its kind. We demonstrate that CBERTScore more closely matches what clinicians prefer.

Hao Cui, Zahra Shamsi, Gowoon Cheon et al.

Scientific problem-solving involves synthesizing information while applying expert knowledge. We introduce CURIE, a scientific long-Context Understanding,Reasoning and Information Extraction benchmark to measure the potential of Large Language Models (LLMs) in scientific problem-solving and assisting scientists in realistic workflows. This benchmark introduces ten challenging tasks with a total of 580 problems and solution pairs curated by experts in six disciplines - materials science, condensed matter physics, quantum computing, geospatial analysis, biodiversity, and proteins - covering both experimental and theoretical work-flows in science. We evaluate a range of closed and open LLMs on tasks in CURIE which requires domain expertise, comprehension of long in-context information,and multi-step reasoning. While Gemini Flash 2.0 and Claude-3 show consistent high comprehension across domains, the popular GPT-4o and command-R+ fail dramatically on protein sequencing tasks. With the best performance at 32% there is much room for improvement for all models. We hope that insights gained from CURIE can guide the future development of LLMs in sciences. Evaluation code and data are in https://github.com/google/curie

Nayantara Mudur, Hao Cui, Subhashini Venugopalan et al.

Building precise simulations of the real world and invoking numerical solvers to answer quantitative problems is an essential requirement in engineering and science. We present FEABench, a benchmark to evaluate the ability of large language models (LLMs) and LLM agents to simulate and solve physics, mathematics and engineering problems using finite element analysis (FEA). We introduce a comprehensive evaluation scheme to investigate the ability of LLMs to solve these problems end-to-end by reasoning over natural language problem descriptions and operating COMSOL Multiphysics$^\circledR$, an FEA software, to compute the answers. We additionally design a language model agent equipped with the ability to interact with the software through its Application Programming Interface (API), examine its outputs and use tools to improve its solutions over multiple iterations. Our best performing strategy generates executable API calls 88% of the time. LLMs that can successfully interact with and operate FEA software to solve problems such as those in our benchmark would push the frontiers of automation in engineering. Acquiring this capability would augment LLMs' reasoning skills with the precision of numerical solvers and advance the development of autonomous systems that can tackle complex problems in the real world. The code is available at https://github.com/google/feabench

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.

Piyush Tiwary, Utkarsh Ahuja, Depanshu Sani et al.

Agricultural landscape segmentation in the Global South is challenging as it is characterized by fragmented plots, high intra-class variance, and a scarcity of labeled training data. Recent advances in segmentation have been made by Multimodal Large Language Models (MLLMs). However, current approaches encounter critical context length bottlenecks and a domain alignment gap in understanding satellite features. We address these limitations through MAgSeg, a novel, decoder-free MLLM segmentation approach. MAgSeg is an architecturally efficient approach that enables standard MLLMs to perform segmentation of complex smallholder agricultural landscapes from high-resolution satellite imagery, without requiring auxiliary vision decoders. We introduce a novel instruction tuning data format designed to enable scalable fine-tuning and post-training on high resolution satellite imagery, which enables MAgSeg to learn from the global context of the image while generating text tokens for only a patch within the image. Extensive evaluations on datasets spanning three countries in the Global South demonstrate that MAgSeg significantly outperforms state-of-the-art MLLM baselines, offering a scalable solution to map smallholder agricultural environments.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Abdoulaye Diack, Perry Nelson, Kwaku Agbesi et al.

The advancement of speech technology has predominantly favored high-resource languages, creating a significant digital divide for speakers of most Sub-Saharan African languages. To address this gap, we introduce WAXAL, a large-scale, openly accessible speech dataset for 21 languages representing over 100 million speakers. The collection consists of two main components: an Automated Speech Recognition (ASR) dataset containing approximately 1,250 hours of transcribed, natural speech from a diverse range of speakers, and a Text-to-Speech (TTS) dataset with over 180 hours of high-quality, single-speaker recordings reading phonetically balanced scripts. This paper details our methodology for data collection, annotation, and quality control, which involved partnerships with four African academic and community organizations. We provide a detailed statistical overview of the dataset and discuss its potential limitations and ethical considerations. The WAXAL datasets are released at https://huggingface.co/datasets/google/WaxalNLP under the permissive CC-BY-4.0 license to catalyze research, enable the development of inclusive technologies, and serve as a vital resource for the digital preservation of these languages.

Mina Lee, Katy Ilonka Gero, John Joon Young Chung et al. · allen-ai, deepmind

In our era of rapid technological advancement, the research landscape for writing assistants has become increasingly fragmented across various research communities. We seek to address this challenge by proposing a design space as a structured way to examine and explore the multidimensional space of intelligent and interactive writing assistants. Through a large community collaboration, we explore five aspects of writing assistants: task, user, technology, interaction, and ecosystem. Within each aspect, we define dimensions (i.e., fundamental components of an aspect) and codes (i.e., potential options for each dimension) by systematically reviewing 115 papers. Our design space aims to offer researchers and designers a practical tool to navigate, comprehend, and compare the various possibilities of writing assistants, and aid in the envisioning and design of new writing assistants.

Subham Sekhar Sahoo, Subhashini Venugopalan, Li Li et al.

Symbolic techniques based on Satisfiability Modulo Theory (SMT) solvers have been proposed for analyzing and verifying neural network properties, but their usage has been fairly limited owing to their poor scalability with larger networks. In this work, we propose a technique for combining gradient-based methods with symbolic techniques to scale such analyses and demonstrate its application for model explanation. In particular, we apply this technique to identify minimal regions in an input that are most relevant for a neural network's prediction. Our approach uses gradient information (based on Integrated Gradients) to focus on a subset of neurons in the first layer, which allows our technique to scale to large networks. The corresponding SMT constraints encode the minimal input mask discovery problem such that after masking the input, the activations of the selected neurons are still above a threshold. After solving for the minimal masks, our approach scores the mask regions to generate a relative ordering of the features within the mask. This produces a saliency map which explains "where a model is looking" when making a prediction. We evaluate our technique on three datasets - MNIST, ImageNet, and Beer Reviews, and demonstrate both quantitatively and qualitatively that the regions generated by our approach are sparser and achieve higher saliency scores compared to the gradient-based methods alone. Code and examples are at - https://github.com/google-research/google-research/tree/master/smug_saliency

Haining Pan, Nayantara Mudur, Will Taranto et al.

Large language models (LLMs) have demonstrated an unprecedented ability to perform complex tasks in multiple domains, including mathematical and scientific reasoning. We demonstrate that with carefully designed prompts, LLMs can accurately carry out key calculations in research papers in theoretical physics. We focus on a broadly used approximation method in quantum physics: the Hartree-Fock method, requiring an analytic multi-step calculation deriving approximate Hamiltonian and corresponding self-consistency equations. To carry out the calculations using LLMs, we design multi-step prompt templates that break down the analytic calculation into standardized steps with placeholders for problem-specific information. We evaluate GPT-4's performance in executing the calculation for 15 research papers from the past decade, demonstrating that, with correction of intermediate steps, it can correctly derive the final Hartree-Fock Hamiltonian in 13 cases and makes minor errors in 2 cases. Aggregating across all research papers, we find an average score of 87.5 (out of 100) on the execution of individual calculation steps. Overall, the requisite skill for doing these calculations is at the graduate level in quantum condensed matter theory. We further use LLMs to mitigate the two primary bottlenecks in this evaluation process: (i) extracting information from papers to fill in templates and (ii) automatic scoring of the calculation steps, demonstrating good results in both cases. The strong performance is the first step for developing algorithms that automatically explore theoretical hypotheses at an unprecedented scale.

Eser Aygün, Anastasiya Belyaeva, Gheorghe Comanici et al.

The cycle of scientific discovery is frequently bottlenecked by the slow, manual creation of software to support computational experiments. To address this, we present an AI system that creates expert-level scientific software whose goal is to maximize a quality metric. The system uses a Large Language Model (LLM) and Tree Search (TS) to systematically improve the quality metric and intelligently navigate the large space of possible solutions. The system achieves expert-level results when it explores and integrates complex research ideas from external sources. The effectiveness of tree search is demonstrated across a wide range of benchmarks. In bioinformatics, it discovered 40 novel methods for single-cell data analysis that outperformed the top human-developed methods on a public leaderboard. In epidemiology, it generated 14 models that outperformed the CDC ensemble and all other individual models for forecasting COVID-19 hospitalizations. Our method also produced state-of-the-art software for geospatial analysis, neural activity prediction in zebrafish, time series forecasting and numerical solution of integrals. By devising and implementing novel solutions to diverse tasks, the system represents a significant step towards accelerating scientific progress.

Chirag Nagpal, Subhashini Venugopalan, Jimmy Tobin et al.

We introduce a large language model (LLM) capable of processing speech inputs and show that tuning it further with reinforcement learning on human preference (RLHF) enables it to adapt better to disordered speech than traditional fine-tuning. Our method replaces low-frequency text tokens in an LLM's vocabulary with audio tokens and enables the model to recognize speech by fine-tuning it on speech with transcripts. We then use RL with rewards based on syntactic and semantic accuracy measures generalizing the LLM further to recognize disordered speech. While the resulting LLM does not outperform existing systems for speech recognition, we find that tuning with reinforcement learning using custom rewards leads to substantially better performance than supervised fine-tuning of the language model, specifically when adapting to speech in a different setting. This presents a compelling alternative tuning strategy for speech recognition using large language models.

Daniel J. Liebling, Malcolm Kane, Madeleine Grunde-Mclaughlin et al.

We present components of an AI-assisted academic writing system including citation recommendation and introduction writing. The system recommends citations by considering the user's current document context to provide relevant suggestions. It generates introductions in a structured fashion, situating the contributions of the research relative to prior work. We demonstrate the effectiveness of the components through quantitative evaluations. Finally, the paper presents qualitative research exploring how researchers incorporate citations into their writing workflows. Our findings indicate that there is demand for precise AI-assisted writing systems and simple, effective methods for meeting those needs.

Katrin Tomanek, Shanqing Cai, Subhashini Venugopalan

Abbreviation expansion is a strategy used to speed up communication by limiting the amount of typing and using a language model to suggest expansions. Here we look at personalizing a Large Language Model's (LLM) suggestions based on prior conversations to enhance the relevance of predictions, particularly when the user data is small (~1000 samples). Specifically, we compare fine-tuning, prompt-tuning, and retrieval augmented generation of expanded text suggestions for abbreviated inputs. Our case study with a deployed 8B parameter LLM on a real user living with ALS, and experiments on movie character personalization indicates that (1) customization may be necessary in some scenarios and prompt-tuning generalizes well to those, (2) fine-tuning on in-domain data (with as few as 600 samples) still shows some gains, however (3) retrieval augmented few-shot selection also outperforms fine-tuning. (4) Parameter efficient tuning allows for efficient and scalable personalization. For prompt-tuning, we also find that initializing the learned "soft-prompts" to user relevant concept tokens leads to higher accuracy than random initialization.

Katy Blumer, Subhashini Venugopalan, Michael P. Brenner et al.

We analyze a dataset of retinal images using linear probes: linear regression models trained on some "target" task, using embeddings from a deep convolutional (CNN) model trained on some "source" task as input. We use this method across all possible pairings of 93 tasks in the UK Biobank dataset of retinal images, leading to ~164k different models. We analyze the performance of these linear probes by source and target task and by layer depth. We observe that representations from the middle layers of the network are more generalizable. We find that some target tasks are easily predicted irrespective of the source task, and that some other target tasks are more accurately predicted from correlated source tasks than from embeddings trained on the same task.

Subhashini Venugopalan, Joel Shor, Manoj Plakal et al.

Automatic classification of disordered speech can provide an objective tool for identifying the presence and severity of speech impairment. Classification approaches can also help identify hard-to-recognize speech samples to teach ASR systems about the variable manifestations of impaired speech. Here, we develop and compare different deep learning techniques to classify the intelligibility of disordered speech on selected phrases. We collected samples from a diverse set of 661 speakers with a variety of self-reported disorders speaking 29 words or phrases, which were rated by speech-language pathologists for their overall intelligibility using a five-point Likert scale. We then evaluated classifiers developed using 3 approaches: (1) a convolutional neural network (CNN) trained for the task, (2) classifiers trained on non-semantic speech representations from CNNs that used an unsupervised objective [1], and (3) classifiers trained on the acoustic (encoder) embeddings from an ASR system trained on typical speech [2]. We found that the ASR encoder's embeddings considerably outperform the other two on detecting and classifying disordered speech. Further analysis shows that the ASR embeddings cluster speech by the spoken phrase, while the non-semantic embeddings cluster speech by speaker. Also, longer phrases are more indicative of intelligibility deficits than single words.

Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci et al.

Integrated Gradients (IG) is a commonly used feature attribution method for deep neural networks. While IG has many desirable properties, the method often produces spurious/noisy pixel attributions in regions that are not related to the predicted class when applied to visual models. While this has been previously noted, most existing solutions are aimed at addressing the symptoms by explicitly reducing the noise in the resulting attributions. In this work, we show that one of the causes of the problem is the accumulation of noise along the IG path. To minimize the effect of this source of noise, we propose adapting the attribution path itself -- conditioning the path not just on the image but also on the model being explained. We introduce Adaptive Path Methods (APMs) as a generalization of path methods, and Guided IG as a specific instance of an APM. Empirically, Guided IG creates saliency maps better aligned with the model's prediction and the input image that is being explained. We show through qualitative and quantitative experiments that Guided IG outperforms other, related methods in nearly every experiment.

Ashish Bora, Siva Balasubramanian, Boris Babenko et al.

Diabetic retinopathy (DR) screening is instrumental in preventing blindness, but faces a scaling challenge as the number of diabetic patients rises. Risk stratification for the development of DR may help optimize screening intervals to reduce costs while improving vision-related outcomes. We created and validated two versions of a deep learning system (DLS) to predict the development of mild-or-worse ("Mild+") DR in diabetic patients undergoing DR screening. The two versions used either three-fields or a single field of color fundus photographs (CFPs) as input. The training set was derived from 575,431 eyes, of which 28,899 had known 2-year outcome, and the remaining were used to augment the training process via multi-task learning. Validation was performed on both an internal validation set (set A; 7,976 eyes; 3,678 with known outcome) and an external validation set (set B; 4,762 eyes; 2,345 with known outcome). For predicting 2-year development of DR, the 3-field DLS had an area under the receiver operating characteristic curve (AUC) of 0.79 (95%CI, 0.78-0.81) on validation set A. On validation set B (which contained only a single field), the 1-field DLS's AUC was 0.70 (95%CI, 0.67-0.74). The DLS was prognostic even after adjusting for available risk factors (p<0.001). When added to the risk factors, the 3-field DLS improved the AUC from 0.72 (95%CI, 0.68-0.76) to 0.81 (95%CI, 0.77-0.84) in validation set A, and the 1-field DLS improved the AUC from 0.62 (95%CI, 0.58-0.66) to 0.71 (95%CI, 0.68-0.75) in validation set B. The DLSs in this study identified prognostic information for DR development from CFPs. This information is independent of and more informative than the available risk factors.

Arunachalam Narayanaswamy, Subhashini Venugopalan, Dale R. Webster et al.

Model explanation techniques play a critical role in understanding the source of a model's performance and making its decisions transparent. Here we investigate if explanation techniques can also be used as a mechanism for scientific discovery. We make three contributions: first, we propose a framework to convert predictions from explanation techniques to a mechanism of discovery. Second, we show how generative models in combination with black-box predictors can be used to generate hypotheses (without human priors) that can be critically examined. Third, with these techniques we study classification models for retinal images predicting Diabetic Macular Edema (DME), where recent work showed that a CNN trained on these images is likely learning novel features in the image. We demonstrate that the proposed framework is able to explain the underlying scientific mechanism, thus bridging the gap between the model's performance and human understanding.

Shawn Xu, Subhashini Venugopalan, Mukund Sundararajan

We study the attribution problem [28] for deep networks applied to perception tasks. For vision tasks, attribution techniques attribute the prediction of a network to the pixels of the input image. We propose a new technique called \emph{Blur Integrated Gradients}. This technique has several advantages over other methods. First, it can tell at what scale a network recognizes an object. It produces scores in the scale/frequency dimension, that we find captures interesting phenomena. Second, it satisfies the scale-space axioms [14], which imply that it employs perturbations that are free of artifact. We therefore produce explanations that are cleaner and consistent with the operation of deep networks. Third, it eliminates the need for a 'baseline' parameter for Integrated Gradients [31] for perception tasks. This is desirable because the choice of baseline has a significant effect on the explanations. We compare the proposed technique against previous techniques and demonstrate application on three tasks: ImageNet object recognition, Diabetic Retinopathy prediction, and AudioSet audio event identification.

Subhashini Venugopalan, Arunachalam Narayanaswamy, Samuel Yang et al.

Confounding variables are a well known source of nuisance in biomedical studies. They present an even greater challenge when we combine them with black-box machine learning techniques that operate on raw data. This work presents two case studies. In one, we discovered biases arising from systematic errors in the data generation process. In the other, we found a spurious source of signal unrelated to the prediction task at hand. In both cases, our prediction models performed well but under careful examination hidden confounders and biases were revealed. These are cautionary tales on the limits of using machine learning techniques on raw data from scientific experiments.

Avinash Varadarajan, Pinal Bavishi, Paisan Raumviboonsuk et al.

Diabetic eye disease is one of the fastest growing causes of preventable blindness. With the advent of anti-VEGF (vascular endothelial growth factor) therapies, it has become increasingly important to detect center-involved diabetic macular edema (ci-DME). However, center-involved diabetic macular edema is diagnosed using optical coherence tomography (OCT), which is not generally available at screening sites because of cost and workflow constraints. Instead, screening programs rely on the detection of hard exudates in color fundus photographs as a proxy for DME, often resulting in high false positive or false negative calls. To improve the accuracy of DME screening, we trained a deep learning model to use color fundus photographs to predict ci-DME. Our model had an ROC-AUC of 0.89 (95% CI: 0.87-0.91), which corresponds to a sensitivity of 85% at a specificity of 80%. In comparison, three retinal specialists had similar sensitivities (82-85%), but only half the specificity (45-50%, p<0.001 for each comparison with model). The positive predictive value (PPV) of the model was 61% (95% CI: 56-66%), approximately double the 36-38% by the retinal specialists. In addition to predicting ci-DME, our model was able to detect the presence of intraretinal fluid with an AUC of 0.81 (95% CI: 0.81-0.86) and subretinal fluid with an AUC of 0.88 (95% CI: 0.85-0.91). The ability of deep learning algorithms to make clinically relevant predictions that generally require sophisticated 3D-imaging equipment from simple 2D images has broad relevance to many other applications in medical imaging.

Yun Liu, Krishna Gadepalli, Mohammad Norouzi et al.

Each year, the treatment decisions for more than 230,000 breast cancer patients in the U.S. hinge on whether the cancer has metastasized away from the breast. Metastasis detection is currently performed by pathologists reviewing large expanses of biological tissues. This process is labor intensive and error-prone. We present a framework to automatically detect and localize tumors as small as 100 x 100 pixels in gigapixel microscopy images sized 100,000 x 100,000 pixels. Our method leverages a convolutional neural network (CNN) architecture and obtains state-of-the-art results on the Camelyon16 dataset in the challenging lesion-level tumor detection task. At 8 false positives per image, we detect 92.4% of the tumors, relative to 82.7% by the previous best automated approach. For comparison, a human pathologist attempting exhaustive search achieved 73.2% sensitivity. We achieve image-level AUC scores above 97% on both the Camelyon16 test set and an independent set of 110 slides. In addition, we discover that two slides in the Camelyon16 training set were erroneously labeled normal. Our approach could considerably reduce false negative rates in metastasis detection.

Ronghang Hu, Marcus Rohrbach, Subhashini Venugopalan et al.

Image segmentation from referring expressions is a joint vision and language modeling task, where the input is an image and a textual expression describing a particular region in the image; and the goal is to localize and segment the specific image region based on the given expression. One major difficulty to train such language-based image segmentation systems is the lack of datasets with joint vision and text annotations. Although existing vision datasets such as MS COCO provide image captions, there are few datasets with region-level textual annotations for images, and these are often smaller in scale. In this paper, we explore how existing large scale vision-only and text-only datasets can be utilized to train models for image segmentation from referring expressions. We propose a method to address this problem, and show in experiments that our method can help this joint vision and language modeling task with vision-only and text-only data and outperforms previous results.

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, 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.

Huijuan Xu, Subhashini Venugopalan, Vasili Ramanishka et al.

Generating natural language descriptions for in-the-wild videos is a challenging task. Most state-of-the-art methods for solving this problem borrow existing deep convolutional neural network (CNN) architectures (AlexNet, GoogLeNet) to extract a visual representation of the input video. However, these deep CNN architectures are designed for single-label centered-positioned object classification. While they generate strong semantic features, they have no inherent structure allowing them to detect multiple objects of different sizes and locations in the frame. Our paper tries to solve this problem by integrating the base CNN into several fully convolutional neural networks (FCNs) to form a multi-scale network that handles multiple receptive field sizes in the original image. FCNs, previously applied to image segmentation, can generate class heat-maps efficiently compared to sliding window mechanisms, and can easily handle multiple scales. To further handle the ambiguity over multiple objects and locations, we incorporate the Multiple Instance Learning mechanism (MIL) to consider objects in different positions and at different scales simultaneously. We integrate our multi-scale multi-instance architecture with a sequence-to-sequence recurrent neural network to generate sentence descriptions based on the visual representation. Ours is the first end-to-end trainable architecture that is capable of multi-scale region processing. Evaluation on a Youtube video dataset shows the advantage of our approach compared to the original single-scale whole frame CNN model. Our flexible and efficient architecture can potentially be extended to support other video processing tasks.

Subhashini Venugopalan, Marcus Rohrbach, Jeff Donahue et al.

Real-world videos often have complex dynamics; and methods for generating open-domain video descriptions should be sensitive to temporal structure and allow both input (sequence of frames) and output (sequence of words) of variable length. To approach this problem, we propose a novel end-to-end sequence-to-sequence model to generate captions for videos. For this we exploit recurrent neural networks, specifically LSTMs, which have demonstrated state-of-the-art performance in image caption generation. Our LSTM model is trained on video-sentence pairs and learns to associate a sequence of video frames to a sequence of words in order to generate a description of the event in the video clip. Our model naturally is able to learn the temporal structure of the sequence of frames as well as the sequence model of the generated sentences, i.e. a language model. We evaluate several variants of our model that exploit different visual features on a standard set of YouTube videos and two movie description datasets (M-VAD and MPII-MD).

Subhashini Venugopalan, Huijuan Xu, Jeff Donahue et al.

Solving the visual symbol grounding problem has long been a goal of artificial intelligence. The field appears to be advancing closer to this goal with recent breakthroughs in deep learning for natural language grounding in static images. In this paper, we propose to translate videos directly to sentences using a unified deep neural network with both convolutional and recurrent structure. Described video datasets are scarce, and most existing methods have been applied to toy domains with a small vocabulary of possible words. By transferring knowledge from 1.2M+ images with category labels and 100,000+ images with captions, our method is able to create sentence descriptions of open-domain videos with large vocabularies. We compare our approach with recent work using language generation metrics, subject, verb, and object prediction accuracy, and a human evaluation.

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.