Layer normalization placement affects training stability and warm-up stage necessity.
problem Training instability and the necessity of a learning rate warm-up stage in Transformers.
method Theoretical analysis and mean field theory to prove gradient behavior at initialization.
result Removing the warm-up stage for Pre-LN Transformers can achieve comparable results with less time and tuning.
Warm-up improves training by adapting learning rate based on curvature.
problem Improving training efficiency in deep learning models.
method Introducing a curvature condition to explain warm-up, and showing empirically that it leads to faster convergence.
result Adapting learning rate based on curvature condition naturally induces warm-up-like schedule, leading to faster convergence.
Paper eliminates warm-up phase for PO in linear MDPs, achieving optimal regret.
problem Costly warm-up phase in PO algorithms for linear MDPs.
method Simple contraction mechanism replaces warm-up phase.
result Achieves rate-optimal regret with improved dependence on problem parameters.
Active learning tackles cold start and imbalanced data issues.
problem Cold start problem and imbalanced data in ML models.
method ODAL approach with 3-stage AL policies.
result ODAL can quickly reach high performance models with limited labels.
DP-SGD can update fewer coordinates while maintaining privacy.
problem How to update fewer coordinates in DP-SGD without losing optimization signal.
method TP-TopK (Two-Phase TopK DP-SGD), a two-phase method for coordinate-sparse private training.
result Private training can update fewer coordinates without losing optimization signal, scaling noise with active dimension \(k\) instead of full dimension \(d\).
Click-through rate (CTR) prediction has been one of the most central problems in computational advertising. Lately, embedding techniques that produce low-dimensional representations of ad IDs drastically improve CTR prediction accuracies. However, such learning techniques are data demanding and work poorly on new ads w…
Paper optimizes MVE network convergence and regularization.
problem Optimizing Mean Variance Estimation networks for better performance.
method Presented two key insights: warm-up period for mean optimization and separate regularization of mean and variance.
result Warm-up period and separate regularization improve MVE network performance.
VCoTTA uses variational Bayesian methods to adapt models under continuous domain shifts.
problem Error accumulation in continual test-time adaptation.
method VCoTTA employs variational Bayesian techniques to update a Bayesian Neural Network (BNN) during testing, combining priors from source and teacher models.
result VCoTTA effectively mitigates error accumulation in CTTA, as shown by experimental results on three datasets.
Paper proposes EEIPU, a memoization-aware BO algorithm to reduce hyperparameter tuning costs.
problem High costs in GPU-days for training and fine-tuning language models.
method Memoization-aware Bayesian Optimization (EEIPU) algorithm in tandem with pipeline caching.
result EEIPU produces 103% more hyperparameter candidates and 108% more validation metric improvement.
New algorithm speeds up Lasso computation by proving faster convergence.
problem Lasso estimator's slow convergence rate due to ℓ1 penalty. method Homotopic approach using surrogate functions.
result Proves O([log(1/ε)]2) convergence rate for Lasso computation. Improves neural network performance by dynamically adjusting model weights based on source reliability.
problem Training neural networks on data from unreliable sources leads to poor performance.
method Dynamic re-weighting strategy using likelihood tempering to adjust model weights based on estimated source reliability.
result Significant improvement in model performance when trained on mixtures of reliable and unreliable data sources.
The paper shows how warming up the learning rate improves deep learning performance.
problem Improving deep learning performance through better handling of larger learning rates.
method Systematic experiments with SGD and Adam showing the benefits of warmup and different regimes of operation.
result Properly choosing ηextinit can eliminate the need for warmup and improve performance. Proposes efficient training method for deep thin networks.
problem Deploying deep learning models with accuracy and compactness.
method Three-stage method: widen, warm up, fine tune.
result Deep thin networks trained with method outperform standard deep networks.
AdaScale SGD adapts learning rates for large-batch training efficiently.
problem Adapting learning rates for large-batch training to balance speed-ups and model quality.
method Adaptive learning rate adaptation based on gradient variance.
result AdaScale achieves reliable speed-ups for a wide range of batch sizes without degrading model quality.
Optimal learning rate schedules derived for various tasks.
problem Inadequate learning rate schedules in practice compared to theory.
method Refined analysis of learning rate schedules for optimization algorithms.
result Derives new problem-adaptive learning rate schedules.
Despite the advances in the representational capacity of approximate distributions for variational inference, the optimization process can still limit the density that is ultimately learned. We demonstrate the drawbacks of biasing the true posterior to be unimodal, and introduce Annealed Variational Objectives (AVO) in…
A novel framework IMBoost improves outlier detection by leveraging the inlier memorization effect.
problem Challenges in unsupervised outlier detection, especially when inliers and outliers are not well-separated or form dense clusters.
method IMBoost framework that incorporates active learning to selectively acquire informative labels and explicitly reinforce the inlier memorization effect.
result IMBoost significantly outperforms state-of-the-art active outlier detection methods and requires less computational cost.
We provide a differential cocycle model for elliptic cohomology with complex coefficients and use analytic methods to construct a cocycle representative for the Witten class in this language. Our motivation stems from the conjectural connection between 2-dimensional field theories and elliptic cohomology originally due…
The paper provides examples of geometric transitions in low dimensions.
problem Exploring geometric transitions between different types of structures in low dimensions.
method Explicit examples and computations of transitions from hyperbolic to Euclidean, spherical, and Anti-de Sitter structures.
result Details of elementary computations and techniques are provided to explain geometric transitions.
We prove that in CAT(0) spaces a quasi-geodesic is Morse if and only if it is contracting. Specifically, in our main theorem we prove that for γ a quasi-geodesic in a CAT(0) space X, the following four statements are equivalent: (i) γ is Morse, (ii) γ is (b,c)--contracting, (iii), γ is strongly contracting, and…
This work analyzes the statistical properties of adaptive gradient methods.
problem Lack of understanding of the statistical properties of adaptive gradient methods.
method Theoretical analyses and experiments on the variance of update magnitudes.
result The variance of update magnitudes is an increasing and bounded function of time, not diverging.
A new framework for selecting base classes in multi-class classification boosts accuracy.
problem Selecting the base class in multi-class classification to improve accuracy.
method Introduces a unified framework with parameters (s,g,w) to search for the base class at each boosting iteration, improving computational efficiency. result Our framework can achieve better test accuracy than the exhaustive search strategy, providing a robust and reliable scheme.
Improves classifier performance in multi-stage selection processes.
problem Difficulty in training classifiers in multi-stage selection processes due to varying sample sizes and information.
method Multi-Stage Transfer Learning (MSGTL) approach that uses knowledge from simpler classifiers trained in early stages to improve later stages.
result MSGTL outperforms other transfer learning methods in real-world selection process data.
A new conformal prediction framework for two-stage models identifies stage-wise uncertainty.
problem Limited coverage guarantees and lack of modular structure understanding in existing conformal prediction methods.
method Decomposes prediction residuals into stage-specific components, calibrates parameters using FWER control, and adapts to non-stationary settings.
result Improves coverage and identifies stage-wise error contributions compared to standard conformal methods.
Variational Autoencoders are powerful models for unsupervised learning. However deep models with several layers of dependent stochastic variables are difficult to train which limits the improvements obtained using these highly expressive models. We propose a new inference model, the Ladder Variational Autoencoder, that…
To integrate strategic, tactical and operational decisions, the two-stage optimization has been widely used to guide dynamic decision making. In this paper, we study the two-stage stochastic programming for complex systems with unknown response estimated by simulation. We introduce the global-local metamodel assisted t…
Dual-stage sEMG classification improves gesture recognition accuracy.
problem Improving accuracy in hand gesture recognition from sEMG signals.
method Dual-stage classification approach: first stage groups similar activities, second stage classifies within groups.
result Dual-stage classification yields significantly higher accuracy than single-stage approach.
Improves classifier performance in multi-stage processes with adversarial autoencoders and multi-task learning.
problem Challenges in training classifiers due to varying sample sizes and information content across stages.
method Combines adversarial autoencoders, multi-task learning, and semi-supervised learning to address underfitting and overfitting.
result Our approach outperforms state-of-the-art methods across different domains.
SALSA automatically adjusts learning rates in stochastic gradient methods.
problem Automatic adjustment of learning rates in stochastic gradient methods.
method SALSA uses a line-search procedure to gradually increase the learning rate, then a statistical test to decrease it.
result SALSA matches the performance of best hand-tuned learning rate schedules in deep learning tasks.
New algorithm efficiently learns sparse staged trees.
problem Scalability issue in existing structural learning algorithms for staged trees.
method Searches over a space of models with limited dependencies.
result First scalable structural learning algorithm for staged trees.
This paper explores the following question: what kind of statistical guarantees can be given when doing variable selection in high-dimensional models? In particular, we look at the error rates and power of some multi-stage regression methods. In the first stage we fit a set of candidate models. In the second stage we s…
In this paper, we consider multi-stage stochastic optimization problems with convex objectives and conic constraints at each stage. We present a new stochastic first-order method, namely the dynamic stochastic approximation (DSA) algorithm, for solving these types of stochastic optimization problems. We show that DSA c…
New algorithms learn staged trees from incomplete data.
problem Handling missing data in staged tree learning.
method Characterized likelihood and pseudo-likelihoods for missing data, implemented structural expectation-maximization algorithm.
result It is feasible to account for different missingness patterns when learning staged trees.
In many classification systems, sensing modalities have different acquisition costs. It is often {\it unnecessary} to use every modality to classify a majority of examples. We study a multi-stage system in a prediction time cost reduction setting, where the full data is available for training, but for a test example, m…
Machine learning can improve 2SLS first stage predictions, but nonlinear methods often introduce bias.
problem Improving the first stage of 2SLS using machine learning.
method Decomposed bias into three components, investigated through simulation.
result Nonlinear machine learning methods can introduce substantial bias in second-stage estimates.
We propose a two-stage hybrid approach with neural networks as the new feature construction algorithms for bankcard response classifications. The hybrid model uses a very simple neural network structure as the new feature construction tool in the first stage, then the newly created features are used as the additional i…
We consider two stage estimation with a non-parametric first stage and a generalized method of moments second stage, in a simpler setting than (Chernozhukov et al. 2016). We give an alternative proof of the theorem given in (Chernozhukov et al. 2016) that orthogonal second stage moments, sample splitting and n1/4-…
CycleFQI tackles offline reinforcement learning for cyclic MDPs, mitigating state distribution mismatch.
problem Offline reinforcement learning for cyclic MDPs with heterogeneous dynamics and discount factors.
method CycleFQI decomposes the cyclic process into stage-wise sub-problems, using vector of stage-specific Q-functions.
result CycleFQI mitigates the curse of dimensionality and provides finite-sample suboptimality error bounds.
Two-stage framework detects multi-modal outliers.
problem Detecting outliers in multi-modal data structures.
method Combines global kernel PCA and local clustering stages.
result Significantly outperforms existing methods on challenging datasets.
The rise of algorithmic decision making led to active researches on how to define and guarantee fairness, mostly focusing on one-shot decision making. In several important applications such as hiring, however, decisions are made in multiple stage with additional information at each stage. In such cases, fairness issues…
A new dynamic learning-rate scheme for optimization.
problem Manual tuning of learning rates in optimization is time-consuming and error-prone.
method Locally optimal stepsize estimation for dynamic learning-rate scheduling.
result Our method achieves comparable performance with minimal tuning.
EML model tackles evolving features in online metric learning.
problem Challenges in applying metric learning to evolving features.
method Develops a new Evolving Metric Learning (EML) model for incremental and decremental features.
result EML model handles instance and feature evolutions simultaneously.
Many problems on signal processing reduce to nonparametric function estimation. We propose a new methodology, piecewise convex fitting (PCF), and give a two-stage adaptive estimate. In the first stage, the number and location of the change points is estimated using strong smoothing. In the second stage, a constrained s…
Hippo optimizes deep learning hyper-parameters by reducing redundant trials.
problem Redundant hyper-parameter trials in hyper-parameter optimization.
method Hippo breaks down hyper-parameter sequences into stages and executes them in a tree structure.
result Hippo reduces GPU-hours and training time significantly compared to existing methods.
The aim of this paper is to write explicit expression in terms of a given principal connection of the Lagrange-d'Alembert-Poincarè equations in several stages. This is obtained by using a reduced Lagrange-d'Alembert's Principle in several stages, extending methods introduced for the case of two stages by one of the aut…
Subspace clustering (SC) refers to the problem of clustering high-dimensional data into a union of low-dimensional subspaces. Based on spectral clustering, state-of-the-art approaches solve SC problem within a two-stage framework. In the first stage, data representation techniques are applied to draw an affinity matrix…
Deep reinforcement learning for high dimensional, hierarchical control tasks usually requires the use of complex neural networks as functional approximators, which can lead to inefficiency, instability and even divergence in the training process. Here, we introduce stacked deep Q learning (SDQL), a flexible modularized…
Generative adversarial networks (GAN) have recently been shown to be efficient for speech enhancement. However, most, if not all, existing speech enhancement GANs (SEGAN) make use of a single generator to perform one-stage enhancement mapping. In this work, we propose to use multiple generators that are chained to perf…