Explanations make AI systems more transparent and also justify their predictions. The top-ranked Visual Question Answering (VQA) systems are ensembles of multiple systems; however, there has been no work on generating explanations for such ensembles. In this paper, we propose different methods for ensembling visual explanations for VQA using the localization maps of the component systems. Our crowd-sourced human evaluation indicates that our ensemble visual explanation is superior to each of the individual system’s visual explanation, although the results vary depending on the individual system that the ensemble is compared against as well as the number of individual systems that agree with the ensemble model’s answer. Overall, our ensemble explanation is better 63% of the time when compared to any individual system’s explanation. Our algorithm is also efficient and scales linearly in the number of component systems in the ensemble.
ML ID: 359
We test whether distributional models can do one-shot learning of definitional properties from text only. Using Bayesian models, we find that first learning overarching structure in the known data, regularities in textual contexts and in properties, helps one-shot learning, and that individual context items can be highly informative. Our experiments show that our model can learn properties from a single exposure when given an informative utterance.
ML ID: 354
Generating computer code from natural language descriptions has been a long- standing problem. Prior work in this domain has restricted itself to generating code in one shot from a single description. To overcome this limitation, we propose a system that can engage users in a dialog to clarify their intent until it has all the information to produce correct code. To evaluate the efficacy of dialog in code generation, we focus on synthesizing conditional statements in the form of IFTTT recipes.
ML ID: 353
We explore techniques to maximize the effectiveness of discourse information in the task of authorship attribution. We present a novel method to embed discourse features in a Convolutional Neural Network text classifier, which achieves a state-of-the-art result by a significant margin. We empirically investigate several featurization methods to understand the conditions under which discourse features contribute non-trivial performance gains, and analyze discourse embeddings.
ML ID: 352
Speech is a natural channel for human-computer interaction in robotics and consumer applications. Natural language understanding pipelines that start with speech can have trouble recovering from speech recognition errors. Black-box automatic speech recognition (ASR) systems, built for general purpose use, are unable to take advantage of in-domain language models that could otherwise ameliorate these errors. In this work, we present a method for re-ranking black-box ASR hypotheses using an in-domain language model and semantic parser trained for a particular task. Our re-ranking method significantly improves both transcription accuracy and semantic understanding over a state-of-the-art ASR’s vanilla output.
ML ID: 351
Active learning identifies data points from a pool of unlabeled examples whose labels, if made available, are most likely to improve the predictions of a supervised model. Most research on active learning assumes that an agent has access to the entire pool of unlabeled data and can ask for labels of any data points during an initial training phase. However, when incorporated in a larger task, an agent may only be able to query some subset of the unlabeled pool. An agent can also opportunistically query for labels that may be useful in the future, even if they are not immediately relevant. In this paper, we demonstrate that this type of opportunistic active learning can improve performance in grounding natural language descriptions of everyday objects---an important skill for home and office robots. We find, with a real robot in an object identification setting, that inquisitive behavior---asking users important questions about the meanings of words that may be off-topic for the current dialog---leads to identifying the correct object more often over time.
ML ID: 350
For most people, watching a brief video and describing what happened (in words) is an easy task. For machines, extracting meaning from video pixels and generating a sentence description is a very complex problem. The goal of this thesis is to develop models that can automatically generate natural language descriptions for events in videos. It presents several approaches to automatic video description by building on recent advances in “deep” machine learning. The techniques presented in this thesis view the task of video description akin to machine translation, treating the video domain as a source “language” and uses deep neural net architectures to “translate” videos to text. Specifically, I develop video captioning techniques using a unified deep neural network with both convolutional and recurrent structure, modeling the temporal elements in videos and language with deep recurrent neural networks. In my initial approach, I adapt a model that can learn from paired images and captions to transfer knowledge from this auxiliary task to generate descriptions for short video clips. Next, I present an end-to-end deep network that can jointly model a sequence of video frames and a sequence of words. To further improve grammaticality and descriptive quality, I also propose methods to integrate linguistic knowledge from plain text corpora. Additionally, I show that such linguistic knowledge can help describe novel objects unseen in paired image/video-caption data. Finally, moving beyond short video clips, I present methods to process longer multi-activity videos, specifically to jointly segment and describe coherent event sequences in full-length movies.
ML ID: 349
When humans encode information into natural language, they do so with the clear assumption that the reader will be able to seamlessly make inferences based on world knowledge. For example, given the sentence ``Mrs. Dalloway said she would buy the flowers herself,'' one can make a number of probable inferences based on event co-occurrences: she bought flowers, she went to a store, she took the flowers home, and so on.Observing this, it is clear that many different useful natural language end-tasks could benefit from models of events as they typically co-occur (so-called script models). Robust question-answering systems must be able to infer highly-probable implicit events from what is explicitly stated in a text, as must robust information-extraction systems that map from unstructured text to formal assertions about relations expressed in the text. Coreference resolution systems, semantic role labeling, and even syntactic parsing systems could, in principle, benefit from event co-occurrence models.
To this end, we present a number of contributions related to statistical event co-occurrence models. First, we investigate a method of incorporating multiple entities into events in a count-based co-occurrence model. We find that modeling multiple entities interacting across events allows for improved empirical performance on the task of modeling sequences of events in documents.
Second, we give a method of applying Recurrent Neural Network sequence models to the task of predicting held-out predicate-argument structures from documents. This model allows us to easily incorporate entity noun information, and can allow for more complex, higher-arity events than a count-based co-occurrence model. We find the neural model improves performance considerably over the count-based co-occurrence model.
Third, we investigate the performance of a sequence-to-sequence encoder-decoder neural model on the task of predicting held-out predicate-argument events from text. This model does not explicitly model any external syntactic information, and does not require a parser. We find the text-level model to be competitive in predictive performance with an event level model directly mediated by an external syntactic analysis.
Finally, motivated by this result, we investigate incorporating features derived from these models into a baseline noun coreference resolution system. We find that, while our additional features do not appreciably improve top-level performance, we can nonetheless provide empirical improvement on a number of restricted classes of difficult coreference decisions.
ML ID: 348
Generating computer code from natural language descriptions has been a longstanding problem in computational linguistics. Prior work in this domain has restricted itself to generating code in one shot from a single description. To overcome this limitation, we propose a system that can engage users in a dialog to clarify their intent until it is confident that it has all the information to produce correct and complete code. Further, we demonstrate how the dialog conversations can be leveraged for continuous improvement of the dialog system. To evaluate the efficacy of dialog in code generation, we focus on synthesizing conditional statements in the form of IFTTT recipes. IFTTT (if-this-then-that) is a web-service that provides event-driven automation, enabling control of smart devices and web-applications based on user-defined events.
ML ID: 347
We present results on using explanations as auxiliary features to improve stacked ensembles for Visual Question Answering (VQA). VQA is a challenging task that requires systems to jointly reason about natural language and vision. We present results applying a recent ensembling approach to VQA, Stacking with Auxiliary Features (SWAF), which learns to combine the results of multiple systems. We propose using features based on explanations to improve SWAF. Using explanations we are able to improve ensembling of three recent VQA systems.
ML ID: 346
Multi-modal grounded language learning connects language predicates to physical properties of objects in the world. Sensing with multiple modalities, such as audio, haptics, and visual colors and shapes while performing interaction behaviors like lifting, dropping, and looking on objects enables a robot to ground non-visual predicates like "empty" as well as visual predicates like "red". Previous work has established that grounding in multi-modal space improves performance on object retrieval from human descriptions. In this work, we gather behavior annotations from humans and demonstrate that these improve language grounding performance by allowing a system to focus on relevant behaviors for words like "white" or "half-full" that can be understood by looking or lifting, respectively. We also explore adding modality annotations (whether to focus on audio or haptics when performing a behavior), which improves performance, and sharing information between linguistically related predicates (if "green" is a color, "white" is a color), which improves grounding recall but at the cost of precision.
ML ID: 345
A word in natural language can be polysemous, having multiple meanings, as well as synonymous, meaning the same thing as other words. Word sense induction attempts to find the senses of polysemous words. Synonymy detection attempts to find when two words are interchangeable. We combine these tasks, first inducing word senses and then detecting similar senses to form word-sense synonym sets (synsets) in an unsupervised fashion. Given pairs of images and text with noun phrase labels, we perform synset induction to produce collections of underlying concepts described by one or more noun phrases. We find that considering multi-modal features from both visual and textual context yields better induced synsets than using either context alone. Human evaluations show that our unsupervised, multi-modally induced synsets are comparable in quality to annotation-assisted ImageNet synsets, achieving about 84% of ImageNet synsets' approval.
ML ID: 344
Ensembling methods are well known for improving prediction accuracy. However, they are limited in the sense that they cannot effectively discriminate among component models. In this paper, we propose stacking with auxiliary features that learns to fuse additional relevant information from multiple component systems as well as input instances to improve performance. We use two types of auxiliary features -- instance features and provenance features. The instance features enable the stacker to discriminate across input instances and the provenance features enable the stacker to discriminate across component systems. When combined together, our algorithm learns to rely on systems that not just agree on an output but also the provenance of this output in conjunction with the properties of the input instance. We demonstrate the success of our approach on three very different and challenging natural language and vision problems: Slot Filling, Entity Discovery and Linking, and ImageNet Object Detection. We obtain new state-of-the-art results on the first two tasks and significant improvements on the ImageNet task, thus verifying the power and generality of our approach.
ML ID: 343
Natural language understanding and dialog management are two integral components of interactive dialog systems. Previous research has used machine learning techniques to individually optimize these components, with different forms of direct and indirect supervision. We present an approach to integrate the learning of both a dialog strategy using reinforcement learning, and a semantic parser for robust natural language understanding, using only natural dialog interaction for supervision. Experimental results on a simulated task of robot instruction demonstrate that joint learning of both components improves dialog performance over learning either of these components alone.
ML ID: 342
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.
ML ID: 341
Recent progress in both AI and robotics have enabled the development of general purpose robot platforms that are capable of executing a wide variety of complex, temporally extended service tasks in open environments. This article introduces a novel, custom-designed multi-robot platform for research on AI, robotics, and especially human–robot interaction for service robots. Called BWIBots, the robots were designed as a part of the Building-Wide Intelligence (BWI) project at the University of Texas at Austin. The article begins with a description of, and justification for, the hardware and software design decisions underlying the BWIBots, with the aim of informing the design of such platforms in the future. It then proceeds to present an overview of various research contributions that have enabled the BWIBots to better (a) execute action sequences to complete user requests, (b) efficiently ask questions to resolve user requests, (c) understand human commands given in natural language, and (d) understand human intention from afar. The article concludes with a look forward towards future research opportunities and applications enabled by the BWIBot platform.
