Improved robustness in ASR systems with speaker adaptation.
problem Improving robustness in automatic speech recognition systems.
method Weighted-Simple-Add method for adding weighted speaker information vectors to the conformer-based acoustic model.
result Achieved 11% relative improvement in WER on Switchboard 300h Hub5'00 dataset.
GNMR controls runtime stability in low-precision language model training.
problem Efficient low-precision training faces numerical risks at specific operators.
method GNMR compares gradient norms to historical means, applying bounded recovery actions.
result GNMR preserves high-fidelity quality with sparse, budgeted recovery.
Modified RV-coefficient reveals how training affects neural network representations.
problem Understanding how training affects intermediate representations in convolutional neural networks.
method Experimented with modified RV-coefficient (RV2) to compare activation patterns in deep networks trained on varying amounts of data and layers.
result RV2 successfully recovered expected similarity patterns and provided interpretable similarity matrices.
The EM training algorithm of the classical i-vector extractor is often incorrectly described as a maximum-likelihood method. The i-vector model is however intractable: the likelihood itself and the hidden-variable posteriors needed for the EM algorithm cannot be computed in closed form. We show here that the classical …
Transformers use ReLUs to approximate softmax efficiently.
problem Analyzing resource usage in softmax transformer models.
method Translating ReLU approximation results to softmax attention mechanisms.
result Economic resource bounds for softmax attention mechanisms.
New method trains deep networks robustly without adaptive methods.
problem Training deep networks with robustness and efficiency.
method Scale invariant architecture + SGD + weight decay + gradient clipping.
result SGD can achieve similar performance to adaptive methods like Adam.
Improves model accuracy for neural nets in stochastic dynamics with partial prior knowledge.
problem Stability and accuracy in neural nets modeling stochastic dynamics with many parameters.
method Three steps: probabilistic weights, partial knowledge incorporation, and PAC-Bayesian training.
result Improved model fit with partial and noisy prior knowledge.
Increased public interest in healthy lifestyles has motivated the study of algorithms that encourage people to follow a healthy diet. Applying collaborative filtering to build recommendation systems in domains where only implicit feedback is available is also a rapidly growing research area. In this report we combine t…
New graph foundation models respect symmetries for broader applicability.
problem Tailored graph machine learning architectures limit broader applicability.
method Investigates symmetries for label and feature permutations, proving network universal approximator.
result Universal approximator on multisets respecting node and feature permutations.
Proposes EVE for efficient exploration in reinforcement learning.
problem Efficient exploration in reinforcement learning.
method EVE: a recipe for posterior over parameters, facilitating efficient exploration.
result Competitive performance on benchmarks, efficient exploration confirmed.
Reinforcement learning after next-token prediction aids in learning from diverse sequence lengths.
problem Learning from sequences of varying lengths and complexity.
method Introducing a framework to study reinforcement learning with autoregressive transformers, focusing on next-token prediction and mixture distributions of short and long sequences.
result Reinforcement learning after next-token prediction enables autoregressive transformers to generalize from long sequences, even when they are rare.
Simplified Variational Bayes for easier inference.
problem Complex derivation of Variational Bayes.
method 3-step recipe to identify posterior form and directly write updates.
result Easier, faster, shorter derivation of Variational Bayes.
Dynamic model improves static economics by incorporating time effects.
problem Static economics overlooks time-dependent phenomena, limiting model accuracy.
method Signals-based approach to reinterpret microeconomic theory, using utility function.
result Dynamic models provide better comparisons with empirical observations.
We give a simple explicit construction of pseudo-Anosov mapping classes using an improvement of the homological criterion of Casson-Bleiler.
Improved graph attention model for noisy graphs.
problem Understanding and improving graph attention in noisy graphs.
method Proposes SuperGAT, a self-supervised graph attention network.
result SuperGAT learns more expressive attention by encoding edges.
We present a theoretical analysis of Gaussian-binary restricted Boltzmann machines (GRBMs) from the perspective of density models. The key aspect of this analysis is to show that GRBMs can be formulated as a constrained mixture of Gaussians, which gives a much better insight into the model's capabilities and limitation…
Unified view of generative models using GFlowNet framework.
problem Diverse deep generative models with varied training and inference methods.
method Integrates GFlowNet framework to unify training and inference.
result Unified training and inference algorithms for generative models.
A new approach to protein language models combines latent space prediction with masked language modeling.
problem Improving protein language models by predicting amino acid identities at masked positions.
method A variant of masked language modeling that predicts latent targets only at masked positions, retaining the MLM cross-entropy.
result The new approach outperforms pure masked language modeling on 11 out of 16 downstream tasks.
A hybrid ASR system using conformer architecture improves word-error-rate and training speed.
problem Improving word-error-rate and training efficiency for hybrid ASR systems.
method Used conformer architecture, applied time downsampling, and transposed convolutions.
result Conformer-based hybrid model achieves competitive results and significantly outperforms BLSTM-based hybrid model.
This paper tackles multilingual speech processing by optimizing conflicting objectives hierarchically.
problem Training models for multilingual, multi-task speech processing is hampered by conflicting objectives.
method Investigates three multi-objective MSP formulations and introduces a lightweight layer-selection mechanism.
result A bi-level recipe outperforms standard flat optimization in state-of-the-art MSP models.
FedShuffle improves local updates in FL, especially with data imbalance.
problem Data imbalance in FL leads to different clients performing different numbers of local updates.
method FedShuffle incorporates random reshuffling, data imbalance, and client sampling.
result FedShuffle improves upon FL methods that assume homogeneous updates in heterogeneous setups.
In recent work on both generative and discriminative score to log-likelihood-ratio calibration, it was shown that linear transforms give good accuracy only for a limited range of operating points. Moreover, these methods required tailoring of the calibration training objective functions in order to target the desired r…
A new method for diffusion generative models improves sample quality and speed.
problem Improving the efficiency and quality of diffusion generative models.
method Developed a complete recipe for forward diffusion processes in SGMs, introducing PSLD.
result PSLD achieves superior sample quality and speed compared to existing methods.
We give a recipe for constructing families of distinct knots that have identical Khovanov homology and give examples of pairs of prime knots, as well as infinite families, with this property.
Twisted SL2C local systems on surfaces of finite type appear often in geometry and physics. Most of them arise geometrically as local systems of charts for pleated hyperbolic structures. Bonahon and Thurston's "shear-bend coordinates" parameterize these local systems of charts. On a surface …
We propose a theoretical framework for thinking about score normalization, which confirms that normalization is not needed under (admittedly fragile) ideal conditions. If, however, these conditions are not met, e.g. under data-set shift between training and runtime, our theory reveals dependencies between scores that c…
LEMON uses pre-trained models to scale neural networks efficiently.
problem Efficiency in scaling deep neural networks, especially Transformers, which are resource-intensive to train from scratch.
method LEMON initializes scaled models using pre-trained weights and optimizes learning rates.
result Significant reduction in training time and computational costs for Vision Transformers and BERT.
Geometrically computes superpotentials for certain 4D N=2 theories.
problem Computing effective twisted superpotentials for 4D N=2 theories.
method Spectral networks and abelianization to compute generating functions of brane opers.
result Geometric recipe for computing effective twisted superpotentials.
As a vast number of ingredients exist in the culinary world, there are countless food ingredient pairings, but only a small number of pairings have been adopted by chefs and studied by food researchers. In this work, we propose KitcheNette which is a model that predicts food ingredient pairing scores and recommends opt…
The paper improves generalization bounds using interpolation between various divergences.
problem Improving generalization bounds in machine learning.
method Derives new PAC-Bayes generalization bounds based on (f,Γ)-divergence and interpolates between various divergences. result Connects derived bounds to earlier statistical learning results and provides practical training objectives.
Optimizing data movement significantly improves transformer training efficiency.
problem Data movement is a major bottleneck in training transformers.
method Developed a recipe to globally optimize data movement in transformers.
result Achieved up to 1.30x performance improvement over state-of-the-art frameworks.
Many recent Markov chain Monte Carlo (MCMC) samplers leverage continuous dynamics to define a transition kernel that efficiently explores a target distribution. In tandem, a focus has been on devising scalable variants that subsample the data and use stochastic gradients in place of full-data gradients in the dynamic s…
Recent advances in conditional image generation tasks, such as image-to-image translation and image inpainting, are largely accounted to the success of conditional GAN models, which are often optimized by the joint use of the GAN loss with the reconstruction loss. However, we reveal that this training recipe shared by …
Paper proposes a new method for training diffusion models using Markov operators.
problem Training efficiency and accuracy in diffusion models.
method Operator-informed score matching using spectral decomposition of Markov operators.
result Improved score matching for both low and high-dimensional distributions.
New method for bidirectional generative modeling using adversarial gradient estimation.
problem Bidirectional generative modeling with various f-divergences. method Adversarial gradient estimation for f-divergence optimization. result Similar algorithms for different f-divergences with varying scaling. A geometric theory explains loss functions for robust representation learning.
problem Treats robustness, domain adaptation, and sensor drift as separate literatures.
method Estimates covariance Sigma_task and uses it to pin Jacobian penalties.
result Proves optimality and necessity of range coverage for penalty matrices.
Estimates model performance from compute budget for distillation.
problem Risk mitigation in large-scale distillation.
method Distillation scaling law based on compute budget allocation.
result Maximizes student performance with compute-optimal allocation.
COMET is a single-pass MapReduce algorithm for learning on large-scale data. It builds multiple random forest ensembles on distributed blocks of data and merges them into a mega-ensemble. This approach is appropriate when learning from massive-scale data that is too large to fit on a single machine. To get the best acc…
CNNs encode data augmentation transformations, especially in early layers.
problem Whether neural network features encode data augmentation transformations.
method Systematic approach using pre-trained vision models to predict augmentation transformations.
result Neural network features encode data augmentation transformations, especially in early layers.
Improved hybrid acoustic model using interleaved self-attention and convolution.
problem Limited application of transformer in hybrid acoustic models.
method Proposed a model structure with interleaved self-attention and 1D convolution.
result Competitive recognition results on Librispeech dataset.
Improves retrieval accuracy for hierarchical documents, especially for distant matches.
problem Limited expressive power of dual encoder models in hierarchical retrieval.
method Proves feasibility of DEs for HR, introduces pretrain-finetune recipe to improve long-distance retrieval.
result Pretrain-finetune boosts recall on long-distance pairs from 19% to 76%.
We describe Sockeye (version 1.12), an open-source sequence-to-sequence toolkit for Neural Machine Translation (NMT). Sockeye is a production-ready framework for training and applying models as well as an experimental platform for researchers. Written in Python and built on MXNet, the toolkit offers scalable training a…
Model generates label-dependent paraphrases for NLP tasks.
problem Generating semantically different paraphrases for NLP tasks.
method Deep variational model with label-dependent generation.
result Model improves generative power of paraphrasing models.
Zero-shot anomaly detection method using batch normalization.
problem Adapting anomaly detectors to new normal data distributions without training data.
method Adaptive Centered Representations (ACR) with batch normalization.
result First zero-shot AD results for tabular data and image data.
BASIS improves LLM reasoning by sharing batchwise rollout info, reducing MSE by 69%.
problem Improving large language model reasoning with limited rollouts and batch information.
method BASIS samples only one rollout per prompt but uses batch information to improve value function estimation.
result BASIS reduces MSE in value function estimation by 69% compared to REINFORCE++.
XRM discovers environments without human annotations for OOD methods.
problem Costly and biased manual annotations limit OOD methods.
method XRM trains twin networks to mimic mistakes, eliminating hyper-parameters.
result XRM achieves oracle worst-group-accuracy for OOD methods.
While Reinforcement Learning (RL) approaches lead to significant achievements in a variety of areas in recent history, natural language tasks remained mostly unaffected, due to the compositional and combinatorial nature that makes them notoriously hard to optimize. With the emerging field of Text-Based Games (TBGs), re…
Programmatic Motion Concepts learn human actions from paired videos.
problem Learning motion concepts from paired video and action sequences.
method Semi-supervised learning architecture for hierarchical motion representation.
result Outperforms established baselines, especially in small data settings.