Paper proposes AXE loss for non-autoregressive machine translation, improving performance.
problem Challenges in training non-autoregressive models due to lack of autoregressive factors and cross entropy loss penalties.
method Proposes aligned cross entropy (AXE) loss function using a differentiable dynamic program for better word order alignment.
result AXE-based training improves performance on major WMT benchmarks and sets a new state of the art for non-autoregressive models.
Efficiently improves non-autoregressive sequence models for better translation performance.
problem Heavy inference latency and inconsistent output sentences in non-autoregressive models.
method Incorporates a structured inference module with an efficient CRF approximation and dynamic transition technique.
result Significantly better translation performance (BLEU score 26.80) compared to previous non-autoregressive models.
Non-autoregressive transformer improves speech recognition speed and accuracy.
problem Reducing inference computation cost in speech recognition.
method Two non-autoregressive transformer structures (A-CMLM and A-FMLM) trained with masked tokens and iterative prediction during inference.
result Non-autoregressive transformer can match autoregressive transformer performance with 7x speedup.
Proposes a non-autoregressive Transformer for time series forecasting.
problem Autoregressive errors and spatial-temporal dependencies in time series forecasting.
method Introduces a Non-Autoregressive Transformer with a learned temporal influence map.
result Demonstrates state-of-the-art performance on time series forecasting datasets.
Paper improves NART models for machine translation with faster inference.
problem Inference latency in NART models is high and translation accuracy is inferior.
method Leverages hints from hidden states and word alignments to train NART models.
result Significant improvement in translation accuracy and comparable inference speed to ART baselines.
Missing value imputation is a fundamental problem in spatiotemporal modeling, from motion tracking to the dynamics of physical systems. Deep autoregressive models suffer from error propagation which becomes catastrophic for imputing long-range sequences. In this paper, we take a non-autoregressive approach and propose …
Non-autoregressive method speeds up protein folding prediction 23 times.
problem Generating protein sequences with higher order interactions.
method Discrete diffusion conditioned on 3D structure using ProteinMPNN.
result 23 times speed up in inference without performance loss.
This work improves NAT translation accuracy through fine-tuning with curriculum learning.
problem Improving NAT translation accuracy while maintaining inference speed.
method Curriculum learning applied to fine-tuning of a pre-trained AT model to a NAT model.
result Significant improvement in translation accuracy (over 1 BLEU score) and speedup in inference.
We propose a conditional non-autoregressive neural sequence model based on iterative refinement. The proposed model is designed based on the principles of latent variable models and denoising autoencoders, and is generally applicable to any sequence generation task. We extensively evaluate the proposed model on machine…
Most machine translation systems generate text autoregressively from left to right. We, instead, use a masked language modeling objective to train a model to predict any subset of the target words, conditioned on both the input text and a partially masked target translation. This approach allows for efficient iterative…
Diffusion models generate music sequences without autoregressive loops.
problem Generating music sequences from symbolic data using diffusion models.
method Parameterize discrete symbolic data in continuous latent space, train diffusion model, generate sequences through reverse process.
result Strong unconditional generation and post-hoc conditional infilling compared to autoregressive models.
WaveGrad generates high-fidelity audio using gradient estimation.
problem Generating high-fidelity audio efficiently.
method Conditional model using score matching and diffusion models, iteratively refining a Gaussian white noise signal.
result WaveGrad can generate high-fidelity audio samples using as few as six iterations.
Normalizing flows are a powerful class of generative models for continuous random variables, showing both strong model flexibility and the potential for non-autoregressive generation. These benefits are also desired when modeling discrete random variables such as text, but directly applying normalizing flows to discret…
EventFlow forecasts event sequences without autoregression, improving accuracy.
problem Forecasting errors in autoregressive models for event sequences.
method EventFlow uses flow matching to learn joint distributions over event times directly.
result EventFlow reduces forecast error by 20%-53% compared to baselines.
New method for data assimilation using score-based models.
problem Bayesian inverse problem of identifying plausible state trajectories.
method Score-based data assimilation, learning a score-based generative model of state trajectories.
result Effective method for zero-shot observation scenarios.
Survey of graph learning methods for combinatorial optimization problems.
problem Efficient and effective analysis of graphs for combinatorial optimization problems.
method Two-stage framework: Graph representation learning followed by machine learning.
result Recent studies have shown promise in using machine learning to solve graph-based combinatorial optimization problems.
This paper proposes an EM approach to reduce inference latency in NAR sequence generation.
problem High inference latency in NAR models due to multi-modality in sequence generation.
method A unified EM framework that jointly optimizes AR and NAR models, with iterative refinement.
result The proposed approach achieves competitive performance with existing NAR models and significantly reduces inference latency.
ElasTST improves time-series forecasting across varying horizons.
problem Robust forecasting across different time horizons in varied industrial sectors.
method Elastic Time-Series Transformer (ElasTST) with non-autoregressive design, rotary position embedding, and multi-scale patching.
result ElasTST provides robust forecasts across varying horizons without retraining.
As a new neural machine translation approach, Non-Autoregressive machine Translation (NAT) has attracted attention recently due to its high efficiency in inference. However, the high efficiency has come at the cost of not capturing the sequential dependency on the target side of translation, which causes NAT to suffer …
Autoregressive state transitions, where predictions are conditioned on past predictions, are the predominant choice for both deterministic and stochastic sequential models. However, autoregressive feedback exposes the evolution of the hidden state trajectory to potential biases from well-known train-test discrepancies.…
This paper introduces a new learning-based approach for approximately solving the Travelling Salesman Problem on 2D Euclidean graphs. We use deep Graph Convolutional Networks to build efficient TSP graph representations and output tours in a non-autoregressive manner via highly parallelized beam search. Our approach ou…
Autoregressive feedback is considered a necessity for successful unconditional text generation using stochastic sequence models. However, such feedback is known to introduce systematic biases into the training process and it obscures a principle of generation: committing to global information and forgetting local nuanc…
This paper compares log-likelihood and BLEU scores for sequence generation tasks.
problem The discrepancy between density estimation and sequence generation performance.
method Comparing several density estimators on five machine translation tasks.
result The correlation between log-likelihood and BLEU varies depending on model families.
This work proposes an efficient autoregressive model for text generation.
problem The challenge of generating high-quality text with autoregressive models.
method Introduces a cascaded decoding approach using Markov transformers to achieve sub-linear parallel time generation.
result Shows competitive accuracy/speed tradeoff compared to existing methods on five machine translation datasets.
While normalizing flows have led to significant advances in modeling high-dimensional continuous distributions, their applicability to discrete distributions remains unknown. In this paper, we show that flows can in fact be extended to discrete events---and under a simple change-of-variables formula not requiring log-d…
Flow-based generative models are powerful exact likelihood models with efficient sampling and inference. Despite their computational efficiency, flow-based models generally have much worse density modeling performance compared to state-of-the-art autoregressive models. In this paper, we investigate and improve upon thr…
DiffWave generates high-fidelity audio waveforms efficiently.
problem Conditional and unconditional audio waveform generation.
method Non-autoregressive diffusion model using Markov chain synthesis.
result DiffWave produces high-quality audios in various tasks.
HCNAF models complex conditional distributions for probabilistic occupancy forecasting.
problem Modeling complex conditional probability density functions for occupancy forecasting.
method Hyper-Conditioned Neural Autoregressive Flow (HCNAF) combining AF and hyper-network.
result HCNAF achieves state-of-the-art performance in self-driving datasets.
Paper proposes energy objective for training normalizing flows without determinants.
problem Challenges in training normalizing flows due to Jacobian determinants.
method Introduces energy objective based on proper scoring rules, determinant-free.
result Energy objective supports novel model families and competitive performance.
Deep neural networks with discrete latent variables offer the promise of better symbolic reasoning, and learning abstractions that are more useful to new tasks. There has been a surge in interest in discrete latent variable models, however, despite several recent improvements, the training of discrete latent variable m…
WaveFit uses fixed-point iteration to create high-quality neural vocoders.
problem Creating high-quality neural vocoders with fast inference.
method Integrates GANs' adversarial training into a DDPM-like iterative framework based on fixed-point iteration.
result WaveFit synthesizes speech with naturalness comparable to human speech, and is significantly faster than existing methods.
PixelCNN models can achieve state-of-the-art results on CIFAR-10 with exact likelihood computation.
problem Dequantization gap in modeling discrete data like images.
method Introducing subset flows to allow exact computation of likelihoods for discrete data.
result PixelCNN models trained with exact likelihood computation achieve state-of-the-art results on CIFAR-10.
Diffusion LLMs can efficiently generate harmful prompts for adversarial testing.
problem Generating harmful prompts for adversarial testing is resource-intensive and costly.
method Transformed adversarial prompt optimization into an efficient inference task using pretrained Diffusion LLMs.
result Only a few conditional samples are required to generate harmful prompts with high reward.
This work proposes a new method to train models with deep latent hierarchies using Optimal Transport.
problem Training models with deep latent hierarchies using VAEs often leads to the 'latent variable collapse' issue.
method Proposes a novel approach based on Optimal Transport to train models with deep latent hierarchies.
result The method avoids the 'latent variable collapse' issue and provides better sample generations and latent representation.
We present the Insertion Transformer, an iterative, partially autoregressive model for sequence generation based on insertion operations. Unlike typical autoregressive models which rely on a fixed, often left-to-right ordering of the output, our approach accommodates arbitrary orderings by allowing for tokens to be ins…
End-to-end Text-to-speech (TTS) system can greatly improve the quality of synthesised speech. But it usually suffers form high time latency due to its auto-regressive structure. And the synthesised speech may also suffer from some error modes, e.g. repeated words, mispronunciations, and skipped words. In this paper, we…
Enhanced Sampling Scheme improves masked generative modeling.
problem Limitations of existing sampling schemes in masked non-autoregressive generative modeling.
method ESS consists of three stages: Naive Iterative Decoding, Critical Reverse Sampling, and Critical Resampling.
result ESS achieves significant performance gains in unconditional and class-conditional sampling.
Undirected neural sequence models such as BERT (Devlin et al., 2019) have received renewed interest due to their success on discriminative natural language understanding tasks such as question-answering and natural language inference. The problem of generating sequences directly from these models has received relativel…
NVAE improves VAE performance on large image datasets.
problem Improving variational autoencoder performance for large image datasets.
method Deep hierarchical VAE with depth-wise separable convolutions and batch normalization, residual parameterization of Normal distributions, and spectral regularization.
result NVAE achieves state-of-the-art results on MNIST, CIFAR-10, CelebA 64, and CelebA HQ datasets.
Latent Block-Diffusion Temporal Point Processes (LBDTPP) is a semi-autoregressive framework for generating asynchronous event sequences.
problem Generating asynchronous event sequences
method Latent Block-Diffusion Temporal Point Processes
result Outperforms state-of-the-art TPP baselines in both unconditional and conditional generation tasks
Unified theory for semi-implicit variational inference, bridging approximation and optimization.
problem Developing a statistical theory for semi-implicit variational inference.
method Unified theory combining approximation and optimization analyses.
result Unified theory characterizes SIVI's ability to recover target distributions and governs asymptotic behavior.
This research explores discrete diffusion models for natural language generation.
problem Challenges in applying diffusion models to discrete data, especially natural language.
method Investigates Discrete Denoising Diffusion Probabilistic Model (D3PM) and compares it with autoregressive models.
result Discrete diffusion models achieve better processing speed than autoregressive models.