We present a novel usage of Transformers to make image classification interpretable. Unlike mainstream classifiers that wait until the last fully connected layer to incorporate class information to make predictions, we investigate a proactive approach, asking each class to search for itself in an image. We realize this idea via a Transformer encoder-decoder inspired by DEtection TRansformer (DETR). We learn ''class-specific'' queries (one for each class) as input to the decoder, enabling each class to localize its patterns in an image via cross-attention. We name our approach INterpretable TRansformer (INTR), which is fairly easy to implement and exhibits several compelling properties. We show that INTR intrinsically encourages each class to attend distinctively; the cross-attention weights thus provide a faithful interpretation of the prediction. Interestingly, via ''multi-head'' cross-attention, INTR could identify different ''attributes'' of a class, making it particularly suitable for fine-grained classification and analysis, which we demonstrate on eight datasets. Our code and pre-trained models are publicly accessible at the Imageomics Institute GitHub site: https://github.com/Imageomics/INTR.

Neural contextual biasing for end-to-end neural ASR transducers has shown significant improvements in the recognition of named entities, such as contact names or device names. However, it comes with the cost of increased compute, as the biasing layers (which are usually based on cross-attention) add complexity to the neural transducers. In this paper, we propose gated contextual biasing models that can estimate at runtime when contextual biasing is needed and can toggle it on or off. That way, contextual biasing does not run on every audio frame, but only on the frames where it can be helpful for correct ASR recognition. We show that our gated contextual biasing models can maintain all the performance improvements of contextual biasing while offering significant compute-cost saving, as the contextual biasing needs to be executed for fewer than 15% of the audio frames.

Personalization in multi-turn dialogs has been a long standing challenge for end-to-end automatic speech recognition (E2E ASR) models. Recent work on contextual adapters has tackled rare word recognition using user catalogs. This adaptation, however, does not incorporate an important cue, the dialog act, which is available in a multi-turn dialog scenario. In this work, we propose a dialog act guided contextual adapter network. Specifically, it leverages dialog acts to select the most relevant user catalogs and creates queries based on both – the audio as well as the semantic relationship between the carrier phrase and user catalogs to better guide the contextual biasing. On industrial voice assistant datasets, our model outperforms both the baselines - dialog act encoder-only model, and the contextual adaptation, leading to the most improvement over the no-context model: 58% average relative word error rate reduction (WERR) in the multi-turn dialog scenario, in comparison to the prior-art contextual adapter, which has achieved 39% WERR over the no-context model.

We present dual-attention neural biasing, an architecture designed to boost Wake Words (WW) recognition and improve inference time latency on speech recognition tasks. This architecture enables a dynamic switch for its runtime compute paths by exploiting WW spotting to select which branch of its attention networks to execute for an input audio frame. With this approach, we effectively improve WW spotting accuracy while saving runtime compute cost as defined by floating point operations (FLOPs). Using an in-house de-identified dataset, we demonstrate that the proposed dual-attention network can reduce the compute cost by 90% for WW audio frames, with only 1% increase in the number of parameters. This architecture improves WW F1 score by 16% relative and improves generic rare word error rate by 3% relative compared to the baselines.

End-to-end Spoken Language Understanding (E2E SLU) has attracted increasing interest due to its advantages of joint optimization and low latency when compared to traditionally cascaded pipelines. Existing E2E SLU models usually follow a two-stage configuration where an Automatic Speech Recognition (ASR) network first predicts a transcript which is then passed to a Natural Language Understanding (NLU) module through an interface to infer semantic labels, such as intent and slot tags. This design, however, does not consider the NLU posterior while making transcript predictions, nor correct the NLU prediction error immediately by considering the previously predicted word-pieces. In addition, the NLU model in the two-stage system is not streamable, as it must wait for the audio segments to complete processing, which ultimately impacts the latency of the SLU system. In this work, we propose a streamable multi-task semantic transducer model to address these considerations. Our proposed architecture predicts ASR and NLU labels auto-regressively and uses a semantic decoder to ingest both previously predicted word-pieces and slot tags while aggregating them through a fusion network. Using an industry scale SLU and a public FSC dataset, we show the proposed model outperforms the two-stage E2E SLU model for both ASR and NLU metrics.

Personal rare word recognition in end-to-end Automatic Speech Recognition (E2E ASR) models is a challenge due to the lack of training data. A standard way to address this issue is with shallow fusion methods at inference time. However, due to their dependence on external language models and the deterministic approach to weight boosting, their performance is limited. In this paper, we propose training neural contextual adapters for personalization in neural transducer based ASR models. Our approach can not only bias towards user-defined words, but also has the flexibility to work with pretrained ASR models. Using an in-house dataset, we demonstrate that contextual adapters can be applied to any general purpose pretrained ASR model to improve personalization. Our method outperforms shallow fusion, while retaining functionality of the pretrained models by not altering any of the model weights. We further show that the adapter style training is superior to full-fine-tuning of the ASR models on datasets with user-defined content.

We present a streaming, Transformer-based end-to-end automatic speech recognition (ASR) architecture which achieves efficient neural inference through compute cost amortization. Our architecture creates sparse computation pathways dynamically at inference time, resulting in selective use of compute resources throughout decoding, enabling significant reductions in compute with minimal impact on accuracy. The fully differentiable architecture is trained end-to-end with an accompanying lightweight arbitrator mechanism operating at the frame-level to make dynamic decisions on each input while a tunable loss function is used to regularize the overall level of compute against predictive performance. We report empirical results from experiments using the compute amortized Transformer-Transducer (T-T) model conducted on LibriSpeech data. Our best model can achieve a 60% compute cost reduction with only a 3% relative word error rate (WER) increase.

End-to-end (E2E) automatic speech recognition (ASR) systems often have difficulty recognizing uncommon words, that appear infrequently in the training data. One promising method, to improve the recognition accuracy on such rare words, is to latch onto personalized/contextual information at inference. In this work, we present a novel context-aware transformer transducer (CATT) network that improves the state-of-the-art transformer-based ASR system by taking advantage of such contextual signals. Specifically, we propose a multi-head attention-based context-biasing network, which is jointly trained with the rest of the ASR sub-networks. We explore different techniques to encode contextual data and to create the final attention context vectors. We also leverage both BLSTM and pretrained BERT based models to encode contextual data and guide the network training. Using an in-house far-field dataset, we show that CATT, using a BERT based context encoder, improves the word error rate of the baseline transformer transducer and outperforms an existing deep contextual model by 24.2% and 19.4% respectively.

Transformers are powerful neural architectures that allow integrating different modalities using attention mechanisms. In this paper, we leverage the neural transformer architectures for multi-channel speech recognition systems, where the spectral and spatial information collected from different microphones are integrated using attention layers. Our multi-channel transformer network mainly consists of three parts: channel-wise self attention layers (CSA), cross-channel attention layers (CCA), and multi-channel encoder-decoder attention layers (EDA). The CSA and CCA layers encode the contextual relationship "within" and "between" channels and across time, respectively. The channel-attended outputs from CSA and CCA are then fed into the EDA layers to help decode the next token given the preceding ones. The experiments show that in a far-field in-house dataset, our method outperforms the baseline single-channel transformer, as well as the super-directive and neural beamformers cascaded with the transformers.

In order to achieve high accuracy for machine learning (ML) applications, it is essential to employ models with a large number of parameters. Certain applications, such as Automatic Speech Recognition (ASR), however, require real-time interactions with users, hence compelling the model to have as low latency as possible. Deploying large scale ML applications thus necessitates model quantization and compression, especially when running ML models on resource constrained devices. For example, by forcing some of the model weight values into zero, it is possible to apply zero-weight compression, which reduces both the model size and model reading time from the memory. In the literature, such methods are referred to as sparse pruning. The fundamental questions are when and which weights should be forced to zero, i.e. be pruned. In this work, we propose a compressed sensing based pruning (CSP) approach to effectively address those questions. By reformulating sparse pruning as a sparsity inducing and compression-error reduction dual problem, we introduce the classic compressed sensing process into the ML model training process. Using ASR task as an example, we show that CSP consistently outperforms existing approaches in the literature.

Multi-channel inputs offer several advantages over singlechannel, to improve the robustness of on-device speech recognition systems.Recent work on multi-channel transformer, has proposed a way to incorporate such inputs into end-to-end ASR for improved accuracy.However, this approach is characterized by a high computational complexity, which prevents it from being deployed in on-device systems.In this paper, we present a novel speech recognition model, Multi-Channel Transformer Transducer (MCTT), which features end-to-end multi-channel training, low computation cost, and low latency so that it is suitable for streaming decoding in on-device speech recognition.In a far-field in-house dataset, our MCTT outperforms stagewise multi-channel models with transformer-transducer up to 6.01% relative WER improvement (WERR).In addition, MCTT outperforms the multi-channel transformer up to 11.62% WERR, and is 15.8 times faster in terms of inference speed.We further show that we can improve the computational cost of MCTT by constraining the future and previous context in attention computations.

We propose a method designed to push the frontiers of unconstrained face recognition in the wild with an emphasis on extreme out-of-plane pose variations. Existing methods either expect a single model to learn pose invariance by training on massive amounts of data or else normalize images by aligning faces to a single frontal pose. Contrary to these, our method is designed to explicitly tackle pose variations. Our proposed Pose-Aware Models (PAM) process a face image using several pose-specific, deep convolutional neural networks (CNN). 3D rendering is used to synthesize multiple face poses from input images to both train these models and to provide additional robustness to pose variations at test time. Our paper presents an extensive analysis of the IARPA Janus Benchmark A (IJB-A), evaluating the effects that landmark detection accuracy, CNN layer selection, and pose model selection all have on the performance of the recognition pipeline. It further provides comparative evaluations on IJB-A and the PIPA dataset. These tests show that our approach outperforms existing methods, even surprisingly matching the accuracy of methods that were specifically fine-tuned to the target dataset. Parts of this work previously appeared in [1] and [2].

We describe a deep learning based method for estimating 3D facial expression coefficients. Unlike previous work, our process does not relay on facial landmark detection methods as a proxy step. Recent methods have shown that a CNN can be trained to regress accurate and discriminative 3D morphable model (3DMM) representations, directly from image intensities. By foregoing landmark detection, these methods were able to estimate shapes for occluded faces appearing in unprecedented viewing conditions. We build on those methods by showing that facial expressions can also be estimated by a robust, deep, landmark-free approach. Our ExpNet CNN is applied directly to the intensities of a face image and regresses a 29D vector of 3D expression coefficients. We propose a unique method for collecting data to train our network, leveraging on the robustness of deep networks to training label noise. We further offer a novel means of evaluating the accuracy of estimated expression coefficients: by measuring how well they capture facial emotions on the CK+ and EmotiW-17 emotion recognition benchmarks. We show that our ExpNet produces expression coefficients which better discriminate between facial emotions than those obtained using state of the art, facial landmark detectors. Moreover, this advantage grows as image scales drop, demonstrating that our ExpNet is more robust to scale changes than landmark detectors. Finally, our ExpNet is orders of magnitude faster than its alternatives.

A key to successful face recognition is accurate and reliable face alignment using automatically-detected facial landmarks. Given this strong dependency between face recognition and facial landmark detection, robust face recognition requires knowledge of when the facial landmark detection algorithm succeeds and when it fails. Facial landmark confidence represents this measure of success. In this paper, we propose two methods to measure landmark detection confidence: local confidence based on local predictors of each facial landmark, and global confidence based on a 3D rendered face model. A score fusion approach is also introduced to integrate these two confidences effectively. We evaluate both confidence metrics on two datasets for face recognition: JANUS CS2 and IJB-A datasets. Our experiments show up to 9% improvements when face recognition algorithm integrates the local-global confidence metrics.

We aim at improving the object recognition with few training data in the target domain by leveraging abundant auxiliary data in the source domain. The major issue obstructing knowledge transfer from source to target is the limited correlation between the two domains. Transferring irrelevant information from the source domain usually leads to performance degradation in the target domain. To address this issue, we propose a transfer learning framework with the two key components, such as discriminative source data reconstruction and dual-domain boosting. The former correlates the two domains via reconstructing source data by target data in a discriminative manner. The latter discovers and delivers only knowledge shared by the target data and the reconstructed source data. Hence, it facilitates recognition in the target. The promising experimental results on three benchmarks of object recognition demonstrate the effectiveness of our approach.

We present an approach MSIL-CRF that incorporates multiple instance learning (MIL) into conditional random fields (CRFs). It can generalize CRFs to work on training data with uncertain labels by the principle of MIL. In this work, it is applied to saving manual efforts on annotating training data for semantic segmentation. Specifically, we consider the setting in which the training dataset for semantic segmentation is a mixture of a few object segments and an abundant set of objects' bounding boxes. Our goal is to infer the unknown object segments enclosed by the bounding boxes so that they can serve as training data for semantic segmentation. To this end, we generate multiple segment hypotheses for each bounding box with the assumption that at least one hypothesis is close to the ground truth. By treating a bounding box as a bag with its segment hypotheses as structured instances, MSIL-CRF selects the most likely segment hypotheses by leveraging the knowledge derived from both the labeled and uncertain training data. The experimental results on the Pascal VOC segmentation task demonstrate that MSIL-CRF can provide effective alternatives to manually labeled segments for semantic segmentation.

According to the observation that an image under the canonical light has a nearly neutral chromaticity distribution, an efficient three-step color correction approach is proposed: First, a preliminary color-corrected image is generated by linearly shifting the chromaticity distribution of an input image towards the neutral point. Then, a least square approximation formulation is presented to derive the illuminant color from the preliminary color-corrected image. Finally, based on the derived illuminant color and the Von Kries model, a refined color-corrected image can be obtained. From the experimental results on the widely-used SFU image dataset, the proposed approach could achieve comparable or even better performances against other well-known methods. In addition, several web images are examined to further demonstrate the effectiveness of our approach.

In this paper, an improved color constancy approach is presented by considering the disadvantage of the well-known max-RGB algorithm: Only the unreliable maximum intensities are taken for illuminant estimation. To solve this problem, we propose the chromaticity neutralization process, which selects the representative pixels for robust illuminant computing. Moreover, to deal with the color deviation cases under multiple illuminant sources, we further present a procedure by first softly clustering the image pixels into several groups; the estimated illuminants from these groups are then combined based on the clustering weights to specifically determine the illuminant of each pixel. The experiments on two databases and web images demonstrate the effectiveness of our approach.