Zero loss is achievable in overparametrized DL networks under specific conditions.
problem Achieving zero loss in overparametrized deep learning networks.
method Determine sufficient conditions for zero loss attainability and present an explicit construction of zero loss minimizers.
result Explicit minimizers for zero loss in overparametrized DL networks are constructed without gradient descent.
Overparameterized deep networks have the capacity to memorize training data with zero \emph{training error}. Even after memorization, the \emph{training loss} continues to approach zero, making the model overconfident and the test performance degraded. Since existing regularizers do not directly aim to avoid zero train…
Gradient descent struggles to achieve zero loss in deep learning models due to non-generic data distributions.
problem Achieving zero loss minimizers in deep learning networks.
method Analysis of gradient descent algorithm in deep learning, focusing on underparametrized networks.
result Zero loss minimization cannot be achieved generically in deep learning networks.
Deep ResNets can achieve zero loss with enough layers and weights.
problem Finding parameters in deep ResNets that fit training data perfectly.
method Mean-field analysis and gradient descent flow to a PDE.
result Gradient descent for ResNet training converges to a zero-loss solution in the large-NN limit.
Uncertainty sampling, a popular active learning algorithm, is used to reduce the amount of data required to learn a classifier, but it has been observed in practice to converge to different parameters depending on the initialization and sometimes to even better parameters than standard training on all the data. In this…
Zero-shot contrastive loss improves text-guided image style transfer without extra training.
problem Stochastic nature of diffusion models leads to trade-offs between style transformation and content preservation.
method Proposes a zero-shot contrastive loss for diffusion models that doesn't require additional fine-tuning or auxiliary networks.
result Method outperforms existing methods while preserving content and requiring no additional training.
Boosted CVaR Classification improves tail performance in classification tasks.
problem Maximizing tail performance in classification tasks.
method Proposed Boosted CVaR Classification framework using randomized classifiers and LPBoost algorithm.
result Minimizing CVaR loss over randomized classifiers leads to better tail performance.
Introduces a differentiable approximation to the zero-one loss.
problem Incompatibility of zero-one loss with gradient-based optimization.
method Smooth projection onto hypersimplex through constrained optimization.
result Achieves significant improvements in generalization under large-batch training.
Neural networks and linear systems linked, revealing training loss and kernel limitations.
problem Exploring the training loss and limitations of neural networks and their kernels.
method Drawing connections between neural networks and under-determined linear systems, providing lower bounds, and analyzing gradient descent.
result Zero training loss achievable for neural networks under certain conditions, but not for ReLU kernels.
Improved GAN training stability through tunable classification losses.
problem Training instabilities in GANs.
method Reformulated GAN value function using class probability estimation (CPE) losses, defined (αD,αG)-GANs. result Tuning (αD,αG) can alleviate training instabilities. Paper establishes a universal growth rate for smooth surrogate losses in classification.
problem Analyzing growth rates of consistency bounds for various surrogate losses.
method Proves square-root growth rate for smooth margin-based losses; extends to multi-class classification.
result Demonstrates a universal square-root growth rate for smooth comp-sum and constrained losses.
Adversarial training achieves optimal test error for shallow networks.
problem Achieving optimal adversarial test error for general data distributions.
method Applying new Rademacher complexity bounds and properties of optimal adversarial predictors.
result Adversarial training can achieve optimal adversarial test error for general data distributions.
The paper studies the loss landscape of regularized deep matrix factorization, revealing unique and sharp minimizers.
problem Understanding the loss landscape and minimizers of regularized deep matrix factorization problems.
method Theoretical analysis of ℓ2-regularized deep matrix factorization/deep linear network training problems with squared-error loss. result The unique end-to-end minimizer exists for all target matrices except for a set of Lebesgue measure zero.
In this paper we apply a compressibility loss that enables learning highly compressible neural network weights. The loss was previously proposed as a measure of negated sparsity of a signal, yet in this paper we show that minimizing this loss also enforces the non-zero parts of the signal to have very low entropy, thus…
We estimate generic statistical properties of a structural credit risk model by considering an ensemble of correlation matrices. This ensemble is set up by Random Matrix Theory. We demonstrate analytically that the presence of correlations severely limits the effect of diversification in a credit portfolio if the corre…
Latent factor models for Recommender Systems with implicit feedback typically treat unobserved user-item interactions (i.e. missing information) as negative feedback. This is frequently done either through negative sampling (point--wise loss) or with a ranking loss function (pair-- or list--wise estimation). Since a ze…
Wide neural networks converge linearly to zero loss with feature learning.
problem Optimizing wide neural networks with feature learning guarantees.
method Gradient flow analysis for wide shallow and multi-layer NNs.
result Training loss converges linearly to zero for wide NNs under GF, demonstrating feature learning and better generalization.
We consider binary classification problems with positive definite kernels and square loss, and study the convergence rates of stochastic gradient methods. We show that while the excess testing loss (squared loss) converges slowly to zero as the number of observations (and thus iterations) goes to infinity, the testing …
Gradient descent with logistic loss can make two-layer networks interpolate binary classification data.
problem Training two-layer networks for binary classification.
method Gradient descent with logistic loss applied to two-layer networks.
result Gradient descent can drive training loss to zero under certain conditions.
This paper explains how low-precision arithmetic causes loss spikes in deep learning models.
problem Loss spikes during long-term training of deep neural networks.
method Analyzes the impact of floating-point precision limits on gradient updates and feature means.
result Numerical Feature Inflation (NFI) explains loss spikes and rapid parameter norm growth.
We analyze the loss landscape and expressiveness of practical deep convolutional neural networks (CNNs) with shared weights and max pooling layers. We show that such CNNs produce linearly independent features at a "wide" layer which has more neurons than the number of training samples. This condition holds e.g. for the…
In finance, durations between successive transactions are usually modeled by the autoregressive conditional duration model based on a continuous distribution omitting zero values. Zero or close-to-zero durations can be caused by either split transactions or independent transactions. We propose a discrete model allowing…
Study of estimation errors in surrogate loss minimizers, providing stronger guarantees than existing methods.
problem Estimation errors in surrogate loss minimizers for various hypothesis sets.
method Detailed study of H-consistency estimation error bounds, proving general theorems for distribution-dependent and independent settings. result Explicit bounds for zero-one and adversarial losses, showing enhancements under distributional assumptions.
A main puzzle of deep neural networks (DNNs) revolves around the apparent absence of "overfitting", defined in this paper as follows: the expected error does not get worse when increasing the number of neurons or of iterations of gradient descent. This is surprising because of the large capacity demonstrated by DNNs to…
Gradient descent finds halfspaces with low error for agnostic learning.
problem Agnostic learning of linear halfspaces with convex surrogates.
method Gradient descent on convex surrogates for zero-one loss.
result Gradient descent finds halfspaces with error O(OPT1/2+ε) in poly time and sample complexity. We identify a class of over-parameterized deep neural networks with standard activation functions and cross-entropy loss which provably have no bad local valley, in the sense that from any point in parameter space there exists a continuous path on which the cross-entropy loss is non-increasing and gets arbitrarily clos…
A composite loss framework is proposed for low-rank modeling of data consisting of interesting and common values, such as excess zeros or missing values. The methodology is motivated by the generalized low-rank framework and the hurdle method which is commonly used to analyze zero-inflated counts. The model is demonstr…
TorchGAN is a PyTorch based framework for writing succinct and comprehensible code for training and evaluation of Generative Adversarial Networks. The framework's modular design allows effortless customization of the model architecture, loss functions, training paradigms, and evaluation metrics. The key features of Tor…
Class Normalization improves zero-shot learning models.
problem Improving zero-shot learning models in a continual setting.
method Class Normalization (CN) technique to address irregular loss surfaces and improve training efficiency.
result CN significantly outperforms state-of-the-art models on 4 standard ZSL datasets.
Recent work has noted that all bad local minima can be removed from neural network loss landscapes, by adding a single unit with a particular parameterization. We show that the core technique from these papers can be used to remove all bad local minima from any loss landscape, so long as the global minimum has a loss o…
Non-parallel many-to-many voice conversion, as well as zero-shot voice conversion, remain under-explored areas. Deep style transfer algorithms, such as generative adversarial networks (GAN) and conditional variational autoencoder (CVAE), are being applied as new solutions in this field. However, GAN training is sophist…
Gradient descent with logistic loss can interpolate deep networks with smoothed ReLU activations under certain conditions.
problem Conditions for gradient descent to drive logistic loss to zero in deep networks with smoothed ReLU activations.
method Gradient descent applied to fixed-width deep networks with smoothed ReLU approximations (e.g., Swish, Huberized ReLU).
result Gradient descent can drive logistic loss to zero under specific conditions, providing bounds on convergence rate.
We address the problem of learning fine-grained cross-modal representations. We propose an instance-based deep metric learning approach in joint visual and textual space. The key novelty of this paper is that it shows that using per-image semantic supervision leads to substantial improvement in zero-shot performance ov…
Generalization and reliability of multilingual translation often highly depend on the amount of available parallel data for each language pair of interest. In this paper, we focus on zero-shot generalization---a challenging setup that tests models on translation directions they have not been optimized for at training t…
We introduce the isoperimetric loss as a regularization criterion for learning the map from a visual representation to a semantic embedding, to be used to transfer knowledge to unknown classes in a zero-shot learning setting. We use a pre-trained deep neural network model as a visual representation of image data, a Wor…
It is widely conjectured that the reason that training algorithms for neural networks are successful because all local minima lead to similar performance, for example, see (LeCun et al., 2015, Choromanska et al., 2015, Dauphin et al., 2014). Performance is typically measured in terms of two metrics: training performanc…
Unified framework connects different neural network models.
problem Understanding the geometry of neural network loss landscapes.
method Unified framework capturing four symmetry classes.
result First discovery of low- and zero-barrier linear interpolation paths.
LMC loss barrier decreases to zero with large network width.
problem Understanding the LMC phenomenon in neural networks.
method Fine-grained analysis of LMC for two-layer ReLU networks.
result LMC loss barrier decreases to zero at a rate of O(m^-1/2) for large network width.
We consider strategies of investments into options and diffusion market model. It is shown that there exists a correct proportion between "put" and "call" in the portfolio such that the average gain is almost always positive for a generic Black and Scholes model. This gain is zero if and only if the market price of ris…
One-pass SGD dynamics in overparameterized quadratic networks show slow escape from poor solutions.
problem Slow escape from poor generalization solutions in overparameterized neural networks.
method Analysis of one-pass SGD dynamics using ordinary differential equations for overlap matrices.
result Overparameterization only modestly accelerates escape from poor solutions.
New method for high-fidelity shape representations from raw data.
problem Creating accurate shape representations from raw data.
method A simple loss function encouraging neural network to vanish on input point cloud and have unit norm gradient.
result Our method produces high-fidelity, smooth, and natural zero level set surfaces.
Paper introduces arctan pinball loss for XGBoost quantile regression.
problem Efficiently predicting multiple quantiles with XGBoost.
method Smooth approximation of pinball loss for XGBoost, using arctan pinball loss.
result Arctan pinball loss reduces quantile crossings and improves efficiency.
In this work, we addressed the issue of applying a stochastic classifier and a local, fuzzy confusion matrix under the framework of multi-label classification. We proposed a novel solution to the problem of correcting label pairwise ensembles. The main step of the correction procedure is to compute classifier- specific…
The paper proves inequalities for zero mode equations on spin manifolds.
problem Sharp inequalities for zero mode equations on spin manifolds.
method Simple proof and classification of equality cases.
result Equality holds if and only if the solution is a Killing spinor and the manifold is Sasaki-Einstein.
Symmetric losses improve classifier robustness from corrupted labels.
problem Improving classifier performance from corrupted labels.
method Symmetric losses that satisfy a certain condition.
result Symmetric losses enhance robust classification from corrupted labels.
This manuscript provides optimization guarantees, generalization bounds, and statistical consistency results for AdaBoost variants which replace the exponential loss with the logistic and similar losses (specifically, twice differentiable convex losses which are Lipschitz and tend to zero on one side). The heart of the…
We examine gradient descent on unregularized logistic regression problems, with homogeneous linear predictors on linearly separable datasets. We show the predictor converges to the direction of the max-margin (hard margin SVM) solution. The result also generalizes to other monotone decreasing loss functions with an inf…
This work improves diffusion models by estimating the optimal loss value for better training diagnostics.
problem The optimal loss value of diffusion models is unknown and not indicative of absolute data-fitting quality.
method Derive the optimal loss in closed form and develop effective estimators, including a stochastic variant.
result Unlocking the optimal loss as a metric for diagnosing training quality of diffusion models.