A new method compresses deep neural networks by predicting and quantizing weights between layers.
problem Resource constraints in deep neural networks.
method Inter-Layer Weight Prediction (ILWP) and quantization based on Smoothly Varying Weight Hypothesis (SVWH).
result The method achieves higher weight compression rates at the same accuracy level.
Channel Pruning, widely used for accelerating Convolutional Neural Networks, is an NP-hard problem due to the inter-layer dependency of channel redundancy. Existing methods generally ignored the above dependency for computation simplicity. To solve the problem, under the Bayesian framework, we here propose a layer-wise…
A new method lifts training of input-convex neural networks to avoid dead weights and plateaued loss.
problem Training input-convex neural networks with non-negative weights.
method Introduces a hypernetwork that emits non-negative weights from a summary of the input batch, adding stochasticity to soften the loss landscape.
result The lift method achieves lower test loss than projected gradient descent and direct softplus reparametrization.
Recent work has studied the reasons for the remarkable performance of deep neural networks in image classification. We examine batch normalization on the one hand and the dynamical systems view of residual networks on the other hand. Our goal is in understanding the notions of stability and smoothness of the inter-laye…
DeepLight accelerates CTR predictions in ad serving by 46X.
problem Significantly increased serving delay and high memory usage for ad serving.
method Explicitly searching feature interactions, pruning layers, promoting sparsity.
result Accelerates model inference by 46X on Criteo dataset.
New model predicts financial connectedness via COVID-19 spread.
problem Predicting financial connectedness during COVID-19 spread.
method Semiparametric matrix regression model with Bayesian hierarchical mixture prior.
result Model captures heterogeneity in network responses to risk factors.
Using detailed statistical analyses of the size distribution of a universe of equity exchange-traded funds (ETFs), we discover a discrete hierarchy of sizes, which imprints a log-periodic structure on the probability distribution of ETF sizes that dominates the details of the asymptotic tail. This allows us to propose …
FOCA method prevents co-adaptation between feature extractor and classifier.
problem Co-adaptation between feature extractor and classifier degrades neural network performance.
method FOCA method uses randomly-generated, weak classifiers to optimize feature extractor without explicit co-adaptation.
result FOCA features form a point-like distribution within the same class under special conditions.
Sideways trains video models by overwriting activations as new frames arrive, potentially improving generalization.
problem Training deep video models synchronously slows down and requires storing activations, limiting parallelism.
method Sideways trains video models by overwriting activations as new frames arrive, breaking the precise correspondence between gradients and activations.
result Sideways training can converge and potentially generalize better than standard synchronized backpropagation.
We introduce a principled approach for unsupervised structure learning of deep neural networks. We propose a new interpretation for depth and inter-layer connectivity where conditional independencies in the input distribution are encoded hierarchically in the network structure. Thus, the depth of the network is determi…
TinyCNN accelerates CNN models on embedded FPGA with 15x speedup.
problem Limited memory on embedded FPGAs restricts CNN performance.
method Software and hardware design tool for FPGA resource-aware CNN accelerator.
result 3% accuracy loss with 15.75x speedup on image classification.
In this work, we introduce a deep-structured conditional random field (DS-CRF) model for the purpose of state-based object silhouette tracking. The proposed DS-CRF model consists of a series of state layers, where each state layer spatially characterizes the object silhouette at a particular point in time. The interact…
FF algorithm uses goodness as a likelihood-ratio test for scalar normalization.
problem Training each layer locally with scalar goodness.
method FF algorithm uses a likelihood-ratio test with squared goodness as the sufficient statistic.
result The FF algorithm generalizes to anisotropic and heavy-tailed populations.
A framework for multilayer networks predicts links without shared structures.
problem Link prediction in multilayer networks without shared structures.
method Bi-level model averaging with K-fold cross-validation. result Framework outperforms existing methods in predictive accuracy and robustness.
FF algorithm uses goodness as a measure of input quality, derived from likelihood-ratio tests.
problem Training each layer locally with a goodness measure.
method FF algorithm uses a likelihood-ratio test to define goodness, which is the sum of squared activations normalized between layers.
result The goodness measure is a sufficient statistic for a likelihood-ratio test, explaining the FF algorithm's performance.
The rapid development of high-throughput technologies has enabled the generation of data from biological or disease processes that span multiple layers, like genomic, proteomic or metabolomic data, and further pertain to multiple sources, like disease subtypes or experimental conditions. In this work, we propose a gene…
The study characterizes how deep learning models transform data.
problem Understanding how deep learning models transform data and their performance.
method Multivariate nonparametric estimator of class separation (HP statistic) to analyze layer-induced representations.
result Characterization of the distributional change to class separation induced at each layer of the model.
The analysis of temporal networks has a wide area of applications in a world of technological advances. An important aspect of temporal network analysis is the discovery of community structures. Real data networks are often very large and the communities are observed to have a hierarchical structure referred to as mult…
The increasing integration of world economies, which organize in complex multilayer networks of interactions, is one of the critical factors for the global propagation of economic crises. We adopt the network science approach to quantify shock propagation on the global trade-investment multiplex network. To this aim, w…
IC-Network improves CNNs by integrating elastic collision units.
problem Designing more effective basic units in neural networks.
method Developed IC layer and IC block units combining the IC structure with convolution operations.
result Significant performance improvements in existing CNNs, reducing top-1 error from 22.85% to 21.49% on imagenet.
Modern data analysis pipelines are becoming increasingly complex due to the presence of multi-view information sources. While graphs are effective in modeling complex relationships, in many scenarios a single graph is rarely sufficient to succinctly represent all interactions, and hence multi-layered graphs have become…
MLP residual networks implement a selective coarse-graining procedure governed by the spectral structure of the input distribution.
problem Understanding the coarse-graining procedure in MLP residual networks
method Analyzing a pure MLP residual stack on synthetic Markov chain sequences
result MLP residual networks implement a selective coarse-graining procedure governed by the spectral structure of the input distribution
A new loss function α-loss bridges log-loss and 0-1 loss for binary classification.
problem Improving binary classification performance using a tunable loss function.
method Introducing α-loss, proving its margin-based form and classification-calibration, and providing an upper bound on empirical risk. result Empirical and expected risk difference upper bound for logistic regression-based classification.
Introduces Fitzpatrick losses, tighter than Fenchel-Young losses.
problem Improving loss functions for machine learning.
method Introduces Fitzpatrick losses based on the Fitzpatrick function.
result Fitzpatrick losses are tighter than Fenchel-Young losses.
We study losses for binary classification and class probability estimation and extend the understanding of them from margin losses to general composite losses which are the composition of a proper loss with a link function. We characterise when margin losses can be proper composite losses, explicitly show how to determ…
We present the Tamed Cross Entropy (TCE) loss function, a robust derivative of the standard Cross Entropy (CE) loss used in deep learning for classification tasks. However, unlike other robust losses, the TCE loss is designed to exhibit the same training properties than the CE loss in noiseless scenarios. Therefore, th…
Unified surrogate loss framework for multi-label learning with strong consistency guarantees.
problem Improving consistency and accounting for label correlations in multi-label learning.
method Introducing multi-label logistic loss and extending it to comprehensive multi-label comp-sum losses, proving strong consistency guarantees for any multi-label loss.
result Unified surrogate loss framework benefiting from strong consistency guarantees for any multi-label loss.
This paper introduces new loss functions for balanced multi-class classification.
problem Balancing class imbalance in multi-class classification.
method Introduces two new surrogate loss families: GLA and GCA.
result GCA losses offer stronger theoretical guarantees in imbalanced settings.
This work broadens calibeating to various proper losses using Bregman divergence.
problem Calibration for a wide range of proper losses.
method Regret minimization and Bregman divergence approach.
result U-calibration results for a family of Tsallis losses with logarithmic regret and dimension independence.
This work generalizes calibeating for a broader range of proper losses using Bregman divergence.
problem Calibration for a wide range of proper losses beyond Brier and log loss.
method Regret minimization based on Bregman divergence for a family of proper losses.
result U-calibration results for a family of Tsallis losses with logarithmic regret and dimension independence.
Proposes squentropy loss for improved classification accuracy and model calibration.
problem Theoretical and empirical evidence for cross-entropy loss is lacking.
method Introduces squentropy loss as the sum of cross-entropy and average square loss over incorrect classes.
result Squentropy loss outperforms cross-entropy and rescaled square losses in classification accuracy and model calibration.
New loss function calibrates WW-hinge loss for multiclass SVM.
problem WW-hinge loss not calibrated with 0-1 loss.
method Introduced ordered partition loss and proved WW-hinge loss is calibrated.
result WW-hinge loss is calibrated with ordered partition loss.
The study analyzes a model for aggregate losses with dependent and overdispersed inter-losses times.
problem Analyzing aggregate loss models with dependent and overdispersed inter-losses times.
method The study uses a two-state Markovian arrival process (MAP2) and a Markov renewal process to model the inter-losses times. Severities are modeled using a heavy-tailed, double-Pareto Lognormal distribution. The model is estimated via direct maximization of the likelihood function.
result The model with dependence and overdispersion in inter-losses times leads to higher capital charges compared to a Poisson process.
Logitron combines Perceptron and logistic loss for improved classification.
problem Non-convex and non-smooth zero-one loss function in classification models.
method Introduces a Perceptron-augmented convex classification framework with an extended logistic loss function.
result Hinge-Logitron outperforms logistic regression and SVM in classification accuracy.
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.
Two new algorithms improve performance in adversarial bandits with unbounded losses.
problem Adversarial Multi-Armed Bandits with unbounded losses.
method Developed UMAB-NN and UMAB-G for non-negative and general unbounded losses respectively.
result UMAB-NN achieves the first adaptive and scale-free regret bound for non-negative unbounded losses.
Paper explores connections between loss functions and consistency in binary classification and regression.
problem Consistency in binary classification and regression applications.
method Characterization of conformable loss functions and derivation of a new Huber-type loss function.
result Margin-based loss functions are equivalent to loss functions of squared standardized logistic regression residuals.
Paper introduces a new topological loss for better convergence.
problem Optimizing topological losses for model's desired topological behavior.
method Introduces a new regularized topology-aware loss function.
result Guarantees efficient optimization of the new loss function.
Novel loss functions improve decision tree learning from noisy data.
problem Training decision trees with noisy labels.
method Introducing distribution losses and a new negative exponential loss.
result The negative exponential loss leads to efficient and robust decision tree learning.
Theoretical analysis of cross-entropy loss functions and their robustness.
problem Guarantees for using cross-entropy as a surrogate loss function.
method Theoretical analysis of a broad family of loss functions, including cross-entropy.
result First H-consistency bounds for comp-sum losses and smooth adversarial comp-sum losses. A new loss function α-loss improves classification robustness and calibration.
problem Improving classification robustness and calibration in machine learning.
method Introduces a tunable loss function α-loss, parameterized by α, and analyzes its theoretical and practical properties. result The α-loss function can improve model robustness to label flips and sensitivity to imbalanced classes. This paper improves operational risk modeling by selecting better loss severity distributions.
problem Inconsistent regulatory capital calculations due to changing loss severity distribution families.
method Presented truncation probability estimates and a consistent quantile scoring function for selection criteria. Also, recommended collecting loss frequencies below the minimum reporting threshold.
result More stable regulatory capital calculations through better selection of loss severity distributions.
This paper improves loss functions for deep learning with noisy labels.
problem Training deep neural networks with noisy labels.
method The paper introduces a normalization technique to make any loss function robust to noisy labels and proposes a framework called Active Passive Loss (APL) to combine robust loss functions.
result The proposed APL framework consistently outperforms state-of-the-art methods, especially under high noise rates.
Introduces MWLD to measure loss inequality across groups.
problem Machine learning's focus on average loss can lead to large group loss discrepancies.
method Defines MWLD, relates it to fairness and robustness, and provides estimation methods.
result MWLD can be estimated efficiently under certain weighting functions and reduces loss variance without significant accuracy loss.
We study cross-country GDP losses due to financial crises in terms of frequency (number of loss events per period) and severity (loss per occurrence). We perform the Loss Distribution Approach (LDA) to estimate a multi-country aggregate GDP loss probability density function and the percentiles associated to extreme eve…
The paper proves deep learning can be robust with certain loss functions.
problem The robustness of deep learning models under flawed data.
method Empirical-risk minimization with unbounded, Lipschitz-continuous loss functions.
result These loss functions provide efficient prediction under minimal data assumptions.
The paper explores transferability of adversarial examples between convex and 01 loss models, finding non-transferability due to different decision boundaries caused by outliers.
problem Transferability of adversarial examples between convex and 01 loss models.
method Empirical study of transferability between linear 01 loss and convex (hinge) loss models, and between neural networks with different activation functions.
result Adversarial examples are non-transferable between convex and 01 loss models due to different decision boundaries caused by outliers.
Symmetrizes loss functions to improve neural network robustness against noisy labels.
problem Designing robust loss functions for noisy labels in neural networks.
method Symmetrization of multi-class loss functions, focusing on cross-entropy and unhinged loss.
result The multi-class unhinged loss is the unique convex symmetric loss under suitable assumptions.