Google Inc.

Baidu USA LLC

Sercan Omer Arik

Tomas Jon Pfister

Gregory Diamos

Han Zhang

Andrew Gibiansky

John Miller

Jonathan Raiman

Wei Ping

Kainan Peng

Jinsung Yoon

Yanqi Zhou

Adam Coates

Zizhao Zhang

Andrew Ng

Mike Chrzanowski

Mohammad Shoeybi

Shubhahrata Sengupta

Chen Xing

Ryan Prenger

Haonan Yu

Christopher Fougner

Hanjun Dai

Hee Woo Jun

Hairong Liu

Ruoxi Sun

Jitong Chen

Sharan Narang

Joel Hestness

Markus Kliegl

Mingfei Gao

Ajay Kannan

Rewon Child

Eric Undersander

Siavash Ebrahimi

Xingchen Wan

Aspects of the disclosure are directed to automatically selecting examples in a prompt for an LLM to demonstrate how to perform tasks. Aspects of the disclosure can select and build a set of examples from LLM zero-shot outputs via predetermined criteria that can combine consistency, diversity, and repetition. In the zero-shot setting for three different LLMs, using only LLM predictions, aspects of the disclosure can improve performance up to 15% compared to zero-shot baselines and can match or exceed few-shot base-lines for a range of reasoning tasks.

Self-improving LLMs through consistency-based self-generated demonstrations

A method for active learning includes obtaining a set of unlabeled training samples and for each unlabeled training sample, perturbing the unlabeled training sample to generate an augmented training sample. The method includes generating, using a machine learning model, a predicted label for both samples and determining an inconsistency value for the unlabeled training sample that represents variance between the predicted labels for the unlabeled and augmented training samples. The method includes sorting the unlabeled training samples based on the inconsistency values and obtaining, for a threshold number of samples selected from the sorted unlabeled training samples, a ground truth label. The method includes selecting a current set of labeled training samples including each selected unlabeled training samples paired with the corresponding ground truth label. The method includes training, using the current set and a proper subset of unlabeled training samples, the machine learning model.

Active learning via a sample consistency assessment

A method for training a locally interpretable model includes obtaining a set of training samples and training a black-box model using the set of training samples. The method also includes generating, using the trained black-box model and the set of training samples, a set of auxiliary training samples and training a baseline interpretable model using the set of auxiliary training samples. The method also includes training, using the set of auxiliary training samples and baseline interpretable model, an instance-wise weight estimator model. For each auxiliary training sample in the set of auxiliary training samples, the method also includes determining, using the trained instance-wise weight estimator model, a selection probability for the auxiliary training sample. The method also includes selecting, based on the selection probabilities, a subset of auxiliary training samples and training the locally interpretable model using the subset of auxiliary training samples.

Reinforcement learning based locally interpretable models

A method includes obtaining a batch of training samples. For each particular training sample in the batch of training samples, the method includes generating, using a data value estimator model and the particular training sample, a corresponding predicted value of the particular training sample when used to train a machine learning model. The method includes selecting, based on the corresponding predicted values, a subset of the batch of training samples. For each particular training sample in the subset of the batch of training samples, the method includes determining, using the machine learning model and the particular training sample, a corresponding prediction performance measurement. The method includes adjusting one or more estimator parameter values of the data value estimator model based on the corresponding prediction performance measurements.

Data valuation using reinforcement learning

A method includes receiving a training data set including a plurality of training data subsets. From two or more training data subsets in the training data set, the method includes selecting a support set of training examples and a query set of training examples. The method includes determining, using the classification model, a centroid value for each respective class. For each training example in the query set of training examples, the method includes generating, using the classification model, a query encoding, determining a class distance measure, determining a ground-truth distance, and updating parameters of the classification model. For each training example in the query set of training examples identified as being misclassified, the method further includes generating a standard deviation value, sampling a new query, and updating parameters of the confidence model based on the new query encoding.

Distance-based learning confidence model

A method of interpreting tabular data includes receiving, at a deep tabular data learning network (TabNet) executing on data processing hardware, a set of features. For each of multiple sequential processing steps, the method also includes: selecting, using a sparse mask of the TabNet, a subset of relevant features of the set of features; processing using a feature transformer of the TabNet, the subset of relevant features to generate a decision step output and information for a next processing step in the multiple sequential processing steps; and providing the information to the next processing step. The method also includes determining a final decision output by aggregating the decision step outputs generated for the multiple sequential processing steps.

Interpretable tabular data learning using sequential sparse attention

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for processing an input data element to generate a prediction output that characterizes the input data element. In one aspect, a method comprises: determining a respective attention weight between an input data element and each of a plurality of reference data elements; processing each of the reference data elements using the encoder neural network to generate a respective value embedding of each reference data element; determining a combined value embedding of the reference data elements based on (i) the respective value embedding of each reference data element, and (ii) the respective attention weight between the input data element and each reference data element; and processing the combined value embedding of the reference data elements using a prediction neural network to generate the prediction output that characterizes the input data element.

Attention-based prototypical learning

A method includes obtaining a set of training samples. During each of a plurality of training iterations, the method also includes sampling a batch of training samples from the set of training samples. The method includes, for each training sample in the batch of training samples, determining, using a data value estimator, a selection probability. The selection probability for the training sample is based on estimator parameter values of the data value estimator. The method also includes selecting, based on the selection probabilities of each training sample, a subset of training samples from the batch of training samples, and determining, using a predictor model with the subset of training samples, performance measurements. The method also includes adjusting model parameter values of the predictor model based on the performance measurements, and updating the estimator parameter values of the data value estimator based on the performance measurements.

Neural Architecture Search (NAS) is a laborious process. Prior work on automated NAS targets mainly on improving accuracy but lacked consideration of computational resource use. Presented herein are embodiments of a Resource-Efficient Neural Architect (RENA), an efficient resource-constrained NAS using reinforcement learning with network embedding. RENA embodiments use a policy network to process the network embeddings to generate new configurations. Example demonstrates of RENA embodiments on image recognition and keyword spotting (KWS) problems are also presented herein. RENA embodiments can find novel architectures that achieve high performance even with tight resource constraints. For the CIFAR10 dataset, the tested embodiment achieved 2.95% test error when compute intensity is greater than 100 FLOPs/byte, and 3.87% test error when model size was less than 3M parameters. For the Google Speech Commands Dataset, the tested RENA embodiment achieved the state-of-the-art accuracy without resource constraints, and it outperformed the optimized architectures with tight resource constraints.

Resource-efficient neural architects

Embodiments of a production-quality text-to-speech (TTS) system constructed from deep neural networks are described. System embodiments comprise five major building blocks: a segmentation model for locating phoneme boundaries, a grapheme-to-phoneme conversion model, a phoneme duration prediction model, a fundamental frequency prediction model, and an audio synthesis model. For embodiments of the segmentation model, phoneme boundary detection was performed with deep neural networks using Connectionist Temporal Classification (CTC) loss. For embodiments of the audio synthesis model, a variant of WaveNet was created that requires fewer parameters and trains faster than the original. By using a neural network for each component, system embodiments are simpler and more flexible than traditional TTS systems, where each component requires laborious feature engineering and extensive domain expertise. Inference with system embodiments may be performed faster than real time.

Real-time neural text-to-speech

Described herein are systems and methods for augmenting neural speech synthesis networks with low-dimensional trainable speaker embeddings in order to generate speech from different voices from a single model. As a starting point for multi-speaker experiments, improved single-speaker model embodiments, which may be referred to generally as Deep Voice 2 embodiments, were developed, as well as a post-processing neural vocoder for Tacotron (a neural character-to-spectrogram model). New techniques for multi-speaker speech synthesis were performed for both Deep Voice 2 and Tacotron embodiments on two multi-speaker TTS datasets&#x2014;showing that neural text-to-speech systems can learn hundreds of unique voices from twenty-five minutes of audio per speaker.

Multi-speaker neural text-to-speech

A method for jointly training a classification model and a confidence model. The method includes receiving a training data set including a plurality of training data subsets. From two or more training data subsets in the training data set, the method includes selecting a support set of training examples and a query set of training examples. The method includes determining, using the classification model, a centroid value for each respective class. For each training example in the query set of training examples, the method includes generating, using the classification model, a query encoding, determining a class distance measure, determining a ground-truth distance, and updating parameters of the classification model. For each training example in the query set of training examples identified as being misclassified, the method further includes generating a standard deviation value, sampling a new query, and updating parameters of the confidence model based on the new query encoding.

For the problem of waveform synthesis from spectrograms, presented herein are embodiments of an efficient neural network architecture, based on transposed convolutions to achieve a high compute intensity and fast inference. In one or more embodiments, for training of the convolutional vocoder architecture, losses are used that are related to perceptual audio quality, as well as a GAN framework to guide with a critic that discerns unrealistic waveforms. While yielding a high-quality audio, embodiments of the model can achieve more than 500 times faster than real-time audio synthesis. Multi-head convolutional neural network (MCNN) embodiments for waveform synthesis from spectrograms are also disclosed. MCNN embodiments enable significantly better utilization of modern multi-core processors than commonly-used iterative algorithms like Griffin-Lim and yield very fast (more than 300&#xd7; real-time) waveform synthesis. Embodiments herein yield high-quality speech synthesis, without any iterative algorithms or autoregression in computations.

Spectrogram to waveform synthesis using convolutional networks

Voice cloning is a highly desired capability for personalized speech interfaces. Neural network-based speech synthesis has been shown to generate high quality speech for a large number of speakers. Neural voice cloning systems that take a few audio samples as input are presented herein. Two approaches, speaker adaptation and speaker encoding, are disclosed. Speaker adaptation embodiments are based on fine-tuning a multi-speaker generative model with a few cloning samples. Speaker encoding embodiments are based on training a separate model to directly infer a new speaker embedding from cloning audios, which is used in or with a multi-speaker generative model. Both approaches achieve good performance in terms of naturalness of the speech and its similarity to original speaker&#x2014;even with very few cloning audios.

Growing community of inventors

San Francisco, CA, United States of America

Sercan Omer Arik