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.

Recent research has made the surprising finding that state-of-the-art deep learning models sometimes fail to generalize to small variations of the input. Adversarial training has been shown to be an effective approach to overcome this problem. However, its application has been limited to enforcing invariance to analytically defined transformations like lp-norm bounded perturbations. Such perturbations do not necessarily cover plausible real-world variations that preserve the semantics of the input (such as a change in lighting conditions). In this paper, we propose a novel approach to express and formalize robustness to these kinds of real-world transformations of the input. The two key ideas underlying our formulation are (1) leveraging disentangled representations of the input to define different factors of variations, and (2) generating new input images by adversarially composing the representations of different images. We use a StyleGAN model to demonstrate the efficacy of this framework. Specifically, we leverage the disentangled latent representations computed by a StyleGAN model to generate perturbations of an image that are similar to real-world variations (like adding make-up, or changing the skin-tone of a person) and train models to be invariant to these perturbations. Extensive experiments show that our method improves generalization and reduces the effect of spurious correlations (reducing the error rate of a "smile" detector by 21% for example).

In this paper we propose to augment a modern neuralnetwork architecture with an attention model inspired by human perception. Specifically, we adversarially train and analyze a neural model incorporating a human inspired, visual attention component that is guided by a recurrent top-down sequential process. Our experimental evaluation uncovers several notable findings about the robustness and behavior of this new model. First, introducing attention to the model significantly improves adversarial robustness resulting in state-of-the-art ImageNet accuracies under a wide range of random targeted attack strengths. Second, we show that by varying the number of attention steps (glances/fixations) for which the model is unrolled, we are able to make its defense capabilities stronger, even in light of stronger attacks - resulting in a “computational race” between the attacker and the defender. Finally, we show that some of the adversarial examples generated by attacking our model are quite different from conventional adversarial examples - they contain global, salient and spatially coherent structures coming from the target class that would be recognizable even to a human, and work by distracting the attention of the model away from the main object in the original image.

Neural architectures able to generate variable-length outputs are extremely effective for applications like Machine Translation and Image Captioning. In this paper, we study the vulnerability of these models to attacks aimed at changing the output-size that can have undesirable consequences including increased computation and inducing faults in downstream modules that expect outputs of a certain length. We show the existence and construction of such attacks with two key contributions. First, to overcome the difficulties of discrete search space and the non-differentiable adversarial objective function, we develop an easy-to-compute differentiable proxy objective that can be used with gradient-based algorithms to find output-lengthening inputs. Second, we develop a verification approach to formally prove that the network cannot produce outputs greater than a certain length. Experimental results on Machine Translation and Image Captioning models show that our adversarial output-lengthening approach can produce outputs that are 50 times longer than the input, while our verification approach can, given a model and input domain, prove that the output length is below a certain size.

Po-Sen Huang, Robert Stanforth, Johannes Welbl, Chris Dyer, Dani Yogatama, Sven Gowal, Krishnamurthy Dvijotham, Pushmeet Kohli. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Recent work has uncovered the interesting (and somewhat surprising) finding that training models to be invariant to adversarial perturbations requires substantially larger datasets than those required for standard classification. This result is a key hurdle in the deployment of robust machine learning models in many real world applications where labeled data is expensive. Our main insight is that unlabeled data can be a competitive alternative to labeled data for training adversarially robust models. Theoretically, we show that in a simple statistical setting, the sample complexity for learning an adversarially robust model from unlabeled data matches the fully supervised case up to constant factors. On standard datasets like CIFAR- 10, a simple Unsupervised Adversarial Training (UAT) approach using unlabeled data improves robust accuracy by 21.7% over using 4K supervised examples alone, and captures over 95% of the improvement from the same number of labeled examples. Finally, we report an improvement of 4% over the previous state-of-the- art on CIFAR-10 against the strongest known attack by using additional unlabeled data from the uncurated 80 Million Tiny Images dataset. This demonstrates that our finding extends as well to the more realistic case where unlabeled data is also uncurated, therefore opening a new avenue for improving adversarial training.

This paper considers the problem of efficient exploration of unseen environments, a key challenge in AI. We propose a learning to explore' framework where we learn a policy from a distribution of environments. At test time, presented with an unseen environment from the same distribution, the policy aims to generalize the exploration strategy to visit the maximum number of unique states in a limited number of steps. We particularly focus on environments with graph-structured state-spaces that are encountered in many important real-world applications like software testing and map building. We formulate this task as a reinforcement learning problem where theexploration' agent is rewarded for transitioning to previously unseen environment states and employ a graph-structured memory to encode the agent's past trajectory. Experimental results demonstrate that our approach is extremely effective for exploration of spatial maps; and when applied on the challenging problems of coverage-guided software-testing of domain-specific programs and real-world mobile applications, it outperforms methods that have been hand-engineered by human experts.

Antoine Bosselut, Asli Celikyilmaz, Xiaodong He, Jianfeng Gao, Po-Sen Huang, Yejin Choi. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Learning to walk over a graph towards a target node for a given query and a source node is an important problem in applications such as knowledge base completion (KBC). It can be formulated as a reinforcement learning (RL) problem with a known state transition model. To overcome the challenge of sparse rewards, we develop a graph-walking agent called M-Walk, which consists of a deep recurrent neural network (RNN) and Monte Carlo Tree Search (MCTS). The RNN encodes the state (i.e., history of the walked path) and maps it separately to a policy and Q-values. In order to effectively train the agent from sparse rewards, we combine MCTS with the neural policy to generate trajectories yielding more positive rewards. From these trajectories, the network is improved in an off-policy manner using Q-learning, which modifies the RNN policy via parameter sharing. Our proposed RL algorithm repeatedly applies this policy-improvement step to learn the model. At test time, MCTS is combined with the neural policy to predict the target node. Experimental results on several graph-walking benchmarks show that M-Walk is able to learn better policies than other RL-based methods, which are mainly based on policy gradients. M-Walk also outperforms traditional KBC baselines.

We develop a technique for transfer learning in machine comprehension (MC) using a novel two-stage synthesis network (SynNet).Given a high-performing MC model in one domain, our technique aims to answer questions about documents in another domain, where we use no labeled data of question-answer pairs.Using the proposed SynNet with a pretrained model on the SQuAD dataset, we achieve an F1 measure of 46.6% on the challenging NewsQA dataset, approaching performance of in-domain models (F1 measure of 50.0%) and outperforming the out-ofdomain baseline by 7.6%, without use of provided annotations. 1

Segmental structure is a common pattern in many types of sequences such as phrases in human languages. In this paper, we present a probabilistic model for sequences via their segmentations. The probability of a segmented sequence is calculated as the product of the probabilities of all its segments, where each segment is modeled using existing tools such as recurrent neural networks. Since the segmentation of a sequence is usually unknown in advance, we sum over all valid segmentations to obtain the final probability for the sequence. An efficient dynamic programming algorithm is developed for forward and backward computations without resorting to any approximation. We demonstrate our approach on text segmentation and speech recognition tasks. In addition to quantitative results, we also show that our approach can discover meaningful segments in their respective application contexts.

Teaching a computer to read and answer general questions pertaining to a document is a challenging yet unsolved problem. In this paper, we describe a novel neural network architecture called the Reasoning Network (ReasoNet) for machine comprehension tasks. ReasoNets make use of multiple turns to effectively exploit and then reason over the relation among queries, documents, and answers. Different from previous approaches using a fixed number of turns during inference, ReasoNets introduce a termination state to relax this constraint on the reasoning depth. With the use of reinforcement learning, ReasoNets can dynamically determine whether to continue the comprehension process after digesting intermediate results, or to terminate reading when it concludes that existing information is adequate to produce an answer. ReasoNets achieve superior performance in machine comprehension datasets, including unstructured CNN and Daily Mail datasets, the Stanford SQuAD dataset, and a structured Graph Reachability dataset.

Monaural source separation is important for many real world applications. It is challenging because, with only a single channel of information available, without any constraints, an infinite number of solutions are possible. In this paper, we explore joint optimization of masking functions and deep recurrent neural networks for monaural source separation tasks, including speech separation, singing voice separation, and speech denoising. The joint optimization of the deep recurrent neural networks with an extra masking layer enforces a reconstruction constraint. Moreover, we explore a discriminative criterion for training neural networks to further enhance the separation performance. We evaluate the proposed system on the TSP, MIR-1K, and TIMIT datasets for speech separation, singing voice separation, and speech denoising tasks, respectively. Our approaches achieve 2.30-4.98 dB SDR gain compared to NMF models in the speech separation task, 2.30-2.48 dB GNSDR gain and 4.32-5.42 dB GSIR gain compared to existing models in the singing voice separation task, and outperform NMF and DNN baselines in the speech denoising task.

Despite their theoretical appeal and grounding in tractable convex optimization techniques, kernel methods are often not the first choice for large-scale speech applications due to their significant memory requirements and computational expense. In recent years, randomized approximate feature maps have emerged as an elegant mechanism to scale-up kernel methods. Still, in practice, a large number of random features is required to obtain acceptable accuracy in predictive tasks. In this paper, we develop two algorithmic schemes to address this computational bottleneck in the context of kernel ridge regression. The first scheme is a specialized distributed block coordinate descent procedure that avoids the explicit materialization of the feature space data matrix, while the second scheme gains efficiency by combining multiple weak random feature models in an ensemble learning framework. We demonstrate that these schemes enable kernel methods to match the performance of state of the art Deep Neural Networks on TIMIT for speech recognition and classification tasks. In particular, we obtain the best classification error rates reported on TIMIT using kernel methods.

Monaural source separation is useful for many real-world applications though it is a challenging problem. In this paper, we study deep learning for monaural speech separation. We propose the joint optimization of the deep learning models (deep neural networks and recurrent neural networks) with an extra masking layer, which enforces a reconstruction constraint. Moreover, we explore a discriminative training criterion for the neural networks to further enhance the separation performance. We evaluate our approaches using the TIMIT speech corpus for a monaural speech separation task. Our proposed models achieve about 3.8∼4.9 dB SIR gain compared to NMF models, while maintaining better SDRs and SARs.

Latent semantic models, such as LSA, intend to map a query to its relevant documents at the semantic level where keyword-based matching often fails. In this study we strive to develop a series of new latent semantic models with a deep structure that project queries and documents into a common low-dimensional space where the relevance of a document given a query is readily computed as the distance between them. The proposed deep structured semantic models are discriminatively trained by maximizing the conditional likelihood of the clicked documents given a query using the clickthrough data. To make our models applicable to large-scale Web search applications, we also use a technique called word hashing, which is shown to effectively scale up our semantic models to handle large vocabularies which are common in such tasks. The new models are evaluated on a Web document ranking task using a real-world data set. Results show that our best model significantly outperforms other latent semantic models, which were considered state-of-the-art in the performance prior to the work presented in this paper.

The recently developed deep learning architecture, a kernel version of the deep convex network (K-DCN), is improved to address the scalability problem when the training and testing samples become very large. We have developed a solution based on the use of random Fourier features, which possess the strong theoretical property of approximating the Gaussian kernel while rendering efficient computation in both training and evaluation of the K-DCN with large training samples. We empirically demonstrate that just like the conventional K-DCN exploiting rigorous Gaussian kernels, the use of random Fourier features also enables successful stacking of kernel modules to form a deep architecture. Our evaluation experiments on phone recognition and speech understanding tasks both show the computational efficiency of the K-DCN which makes use of random features. With sufficient depth in the K-DCN, the phone recognition accuracy and slot-filling accuracy are shown to be comparable or slightly higher than the K-DCN with Gaussian kernels while significant computational saving has been achieved.

Confidence classifiers for automatic speech recognition (ASR) provide a quantitative representation for the reliability of ASR decoding. In this paper, we improve the ASR confidence measure performance for an utterance using two distinct approaches: (1) to define and incorporate additional predictors in the confidence classifier including those based on the word identity and on the aggregated words, and (2) to train the confidence classifier built on deep learning architectures including the deep neural network (DNN) and the kernel deep convex network (K-DCN). Our experiments show that adding the new predictors to our multi-layer perceptron (MLP)-based baseline classifier provides 38.6% relative reduction in the correct-reject rate as our measure of the classifier performance. Further, replacing the MLP with the DNN and K-DCN provides an additional 14.5% and 47.5% in the relative performance gain, respectively.

Separating singing voices from music accompaniment is an important task in many applications, such as music information retrieval, lyric recognition and alignment. Music accompaniment can be assumed to be in a low-rank subspace, because of its repetition structure; on the other hand, singing voices can be regarded as relatively sparse within songs. In this paper, based on this assumption, we propose using robust principal component analysis for singing-voice separation from music accompaniment. Moreover, we examine the separation result by using a binary time-frequency masking method. Evaluations on the MIR-1K dataset show that this method can achieve around 1~1.4 dB higher GNSDR compared with two state-of-the-art approaches without using prior training or requiring particular features.

This paper presents a system that uses symbolic representations of audio concepts as words for the descriptions of audio tracks, that enable it to go beyond the state of the art, which is audio event classification of a small number of audio classes in constrained settings, to large-scale classification in the wild. These audio words might be less meaningful for an annotator but they are descriptive for computer algorithms. We devise a random-forest vocabulary learning method with an audio word weighting scheme based on TF-IDF and TD-IDD, so as to combine the computational simplicity and accurate multi-class classification of the random forest with the data-driven discriminative power of the TF-IDF/TD-IDD methods. The proposed random forest clustering with text-retrieval methods significantly outperforms two state-of-the-art methods on the dry-run set and the full set of the TRECVID MED 2010 dataset.

In recent years, designing the coding and pooling structures in layered networks has been shown to be a useful method for learning high-level feature representations for visual data. Yet, such learning structures have not been extensively studied for audio signals. In this paper, we investigate the different pooling strategies based on the sparse coding scheme and propose a temporal pyramid pooling method to extract discriminative and shift-invariant feature representations. We demonstrate the superiority of our new feature representation over traditional features on the acoustic event classification task. Index Terms: sparse coding, pooling, acoustic event classification 1.

Acoustic event detection (AED) aims to identify both timestamps and types of multiple events and has been found to be very challenging. The cues for these events often times exist in both audio and vision, but not necessarily in a synchronized fashion. We study improving the detection and classification of the events using cues from both modalities. We propose optical flow based spatial pyramid histograms as a generalizable visual representation that does not require training on labeled video data. Hidden Markov models (HMMs) are used for audio-only modeling, and multi-stream HMMs or coupled HMMs (CHMM) are used for audio-visual joint modeling. To allow the flexibility of audio-visual state asynchrony, we explore effective CHMM training via HMM state-space mapping, parameter tying and different initialization schemes. The proposed methods successfully improve acoustic event classification and detection on a multimedia meeting room dataset containing eleven types of general non-speech events without using extra data resource other than the video stream accompanying the audio observations. Our systems perform favorably compared to previously reported systems leveraging ad-hoc visual cue detectors and localization information obtained from multiple microphones.

Recently, audio concepts emerged as a useful building block in multimodal video retrieval systems. Information like "this file contains laughter", "this file contains engine sounds" or "this file contains slow music" can significantly improve purely visual based retrieval. The weak point of current approaches to audio concept detection is that they heavily rely on human annotators. In most approaches, audio material is manually inspected to identify relevant concepts. Then instances that contain examples of relevant concepts are selected -- again manually -- and used to train concept detectors. This approach comes with two major disadvantages: (1) it leads to rather abstract audio concepts that hardly cover the audio domain at hand and (2) the way human annotators identify audio concepts likely differs from the way a computer algorithm clusters audio data -- introducing additional noise in training data. This paper explores whether unsupervized audio segementation systems can be used to identify useful audio concepts by analyzing training data automatically and whether these audio concepts can be used for multimedia document classification and retrieval. A modified version of the ICSI (International Computer Science Institute) speaker diarization system finds segments in an audio track that have similar perceptual properties and groups these segments. This article provides an in-depth analysis on the statistic properties of similar acoustic segments identified by the diarization system in a predefined document set and the theoretical fitness of this approach to discern one document class from another.