Filters








213 Hits in 3.3 sec

How does overparametrization affect performance on minority groups? [article]

Subha Maity, Saptarshi Roy, Songkai Xue, Mikhail Yurochkin, Yuekai Sun
2022 arXiv   pre-print
Recent empirical studies demonstrate encouraging results: (i) when groups are not known, overparameterized models trained with empirical risk minimization (ERM) perform better on minority groups; (ii)  ...  In a setting in which the regression functions for the majority and minority groups are different, we show that overparameterization always improves minority group performance.  ...  In this paper, we investigate how overparameterization affects the performance of ML models on minority groups in a regression setup.  ... 
arXiv:2206.03515v1 fatcat:wcpce3r2pbge5aglgghquey3zq

Removing Spurious Features can Hurt Accuracy and Affect Groups Disproportionately [article]

Fereshte Khani, Percy Liang
2020 arXiv   pre-print
We completely characterize how the removal of spurious features affects accuracy across different groups (more generally, test distributions) in noiseless overparameterized linear regression.  ...  The presence of spurious features interferes with the goal of obtaining robust models that perform well across many groups within the population.  ...  As a result, we were enabled to analyze the drop in accuracy in a more interpretable way and show how removing the spurious feature affects different groups.  ... 
arXiv:2012.04104v1 fatcat:k6sbbvkqo5a6teut56qss42jri

Label-Imbalanced and Group-Sensitive Classification under Overparameterization [article]

Ganesh Ramachandra Kini, Orestis Paraskevas, Samet Oymak, Christos Thrampoulidis
2021 arXiv   pre-print
In contrast to previous heuristics, we follow a principled analysis explaining how different loss adjustments affect margins.  ...  Importantly, our experiments on state-of-the-art datasets are fully consistent with our theoretical insights and confirm the superior performance of our algorithms.  ...  Our second experiment (Fig. 4 ) validates the theory of Sec. 3 by examining how these adjustments affect training. 1 evaluates LA/CDT/VS-losses on imbalanced instances of CIFAR-10/100.  ... 
arXiv:2103.01550v3 fatcat:x4nnoidpw5bk3jjdyxnvqntwyi

Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization [article]

Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, Percy Liang
2020 arXiv   pre-print
Instead, the poor worst-case performance arises from poor generalization on some groups.  ...  Overparameterized neural networks can be highly accurate on average on an i.i.d. test set yet consistently fail on atypical groups of the data (e.g., by learning spurious correlations that hold on average  ...  Unlike group DRO, upweighting the minority groups does not necessarily yield uniformly low training losses across groups in practice, as some groups might be easier to fit than others.  ... 
arXiv:1911.08731v2 fatcat:aannrfsuundw3ptcvqtwyr6eym

Certainty Groups: A Practical Approach to Distinguish Confidence Levels in Neural Networks

Lukas Lodes, Alexander Schiendorfer
2022 Proceedings of the European Conference of the Prognostics and Health Management Society (PHME)  
We present an approach to compute these certainty groups and demonstrate our approach on two datasets from a PHM setting.  ...  We therefore introduce the concept of certainty groups which distinguish the predictions of a neural network into the normal group and the certainty group.  ...  This was done to answer the question of how does the target accuracy for Certainty Groups affect their size.  ... 
doi:10.36001/phme.2022.v7i1.3331 fatcat:mx6ccibze5hnrgre6ngevn7gde

Training Efficiency and Robustness in Deep Learning [article]

Fartash Faghri
2021 arXiv   pre-print
In the context of learning visual-semantic embeddings, we find that prioritizing learning on more informative training data increases convergence speed and improves generalization performance on test data  ...  Deep Learning has revolutionized machine learning and artificial intelligence, achieving superhuman performance in several standard benchmarks.  ...  Random Features Models: How Does Overparametrization Affect the Variance?  ... 
arXiv:2112.01423v1 fatcat:3yqco7htnjdbng4hx2ilkrnkaq

Technical Challenges for Training Fair Neural Networks [article]

Valeriia Cherepanova and Vedant Nanda and Micah Goldblum and John P. Dickerson and Tom Goldstein
2021 arXiv   pre-print
We conduct our experiments on both facial recognition and automated medical diagnosis datasets using state-of-the-art architectures.  ...  While fairness constraints have been studied extensively for classical models, the effectiveness of methods for imposing fairness on deep neural networks is unclear.  ...  In this section, we take a closer look at how this affected disparities across another sensitive feature, age.  ... 
arXiv:2102.06764v1 fatcat:nunewctjxjd73as3muubliiknq

An Investigation of Why Overparameterization Exacerbates Spurious Correlations [article]

Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, Percy Liang
2020 arXiv   pre-print
We study why overparameterization -- increasing model size well beyond the point of zero training error -- can hurt test error on minority groups despite improving average test error when there are spurious  ...  Through simulations and experiments on two image datasets, we identify two key properties of the training data that drive this behavior: the proportions of majority versus minority groups, and the signal-to-noise  ...  How does the same reweighted logistic regression perform in the underparameterized regime? We focus on the setting where N = 0.  ... 
arXiv:2005.04345v3 fatcat:23wyihwb3bcdfpf26gnu2xdh64

Granular Motor State Monitoring of Free Living Parkinson's Disease Patients via Deep Learning [article]

Kamer A. Yuksel, Jann Goschenhofer, Hridya V. Varma, Urban Fietzek, Franz M.J. Pfister
2019 arXiv   pre-print
Parkinson's disease (PD) is the second most common neurodegenerative disease worldwide and affects around 1% of the (60+ years old) elderly population in industrial nations.  ...  is a significant variety in how these symptoms affects patient motions.  ...  FCN+ is a samller version that is 2/3 of FCN++'s size as it does contain the channel attention.  ... 
arXiv:1911.06913v2 fatcat:qwrv6cnyhvehlaut7vqfpuaksi

Pruning has a disparate impact on model accuracy [article]

Cuong Tran, Ferdinando Fioretto, Jung-Eun Kim, Rakshit Naidu
2022 arXiv   pre-print
The paper sheds light on the factors to cause such disparities, suggesting differences in gradient norms and distance to decision boundary across groups to be responsible for this critical issue.  ...  Notice how the accuracy of the majority group (White) tends to increase while that of the minority groups tends to decrease as the pruning ratio increases.  ...  It does so by connecting the maximum eigenvalues of the groups Hessians with their distance to decision boundary and the group accuracy. The following result sheds light on these observations.  ... 
arXiv:2205.13574v2 fatcat:aahvehbchjhxva3y3cjvqu7fsq

Mostly Harmless Direct Effects: A Comparison of Different Latent Markov Modeling Approaches

Roberto Di Mari, Zsuzsa Bakk
2017 Structural Equation Modeling  
We evaluate the performance of the most common estimators of Latent Markov (LM) models with covariates in the presence of direct effects of the covariates on the indicators of the LM model.  ...  We evaluate how the presence of direct effects influences the bias and efficiency of the three most common estimators of LM models, the one-step, two-step and three-step approaches.  ...  their tolerance toward minority groups on data from the 1976 and 1977 General Social Survey (GSS) using a naive three-step approach.  ... 
doi:10.1080/10705511.2017.1387860 fatcat:mnhhirne5vbwfgucnynl5sqp4q

Even-degree lateral variations in the Earth's mantle constrained by free oscillations and the free-air gravity anomaly

Miaki Ishii, Jeroen Tromp
2001 Geophysical Journal International  
The inversion puts weak constraints on even-degree topographic variations on the coremantle boundary, the 660 km discontinuity and dynamic free surface topography.  ...  Compared to earlier studies, the number of normal-mode constraints on lateral variations in the mantle has increased five-fold, and toroidal and cross-coupled modes complement the traditional spheroidal  ...  Backus-Gilbert resolution tests In contrast to the resolution test, which investigates how a given model is affected by the inversion, a Backus-Gilbert resolution test asks how an inverted model is related  ... 
doi:10.1111/j.1365-246x.2001.00385.x fatcat:csg3km7n35h6hlxhkjxyk6kknq

Even-degree lateral variations in the Earth's mantle constrained by free oscillations and the free-air gravity anomaly

Miaki Ishii, Jeroen Tromp
2001 Geophysical Journal International  
The inversion puts weak constraints on even-degree topographic variations on the coremantle boundary, the 660 km discontinuity and dynamic free surface topography.  ...  Compared to earlier studies, the number of normal-mode constraints on lateral variations in the mantle has increased five-fold, and toroidal and cross-coupled modes complement the traditional spheroidal  ...  Backus-Gilbert resolution tests In contrast to the resolution test, which investigates how a given model is affected by the inversion, a Backus-Gilbert resolution test asks how an inverted model is related  ... 
doi:10.1046/j.1365-246x.2001.00385.x fatcat:ox2cynxoerct3d5ehpjvitegbq

High Dimensional Forecasting via Interpretable Vector Autoregression [article]

William B. Nicholson, Ines Wilms, Jacob Bien, David S. Matteson
2020 arXiv   pre-print
The key modeling tool is a group lasso with nested groups which guarantees that the sparsity pattern of lag coefficients honors the VAR's ordered structure.  ...  Such an approach constrains the relationship between the components and impedes forecast performance.  ...  While this negatively affects their lag order selection performance, it helps for forecast performance as discussed in Section 5.1.  ... 
arXiv:1412.5250v4 fatcat:oxh4r7cadnhtxghufn4c3tmwuq

Insight Into Individual Differences in Emotion Dynamics With Clustering

Anja F. Ernst, Marieke E. Timmerman, Bertus F. Jeronimus, Casper J. Albers
2019 Assessment (Odessa, Fla.)  
To enable comparisons and generalizations of dynamics across groups of individuals, one needs sophisticated tools that express the essential similarities and differences.  ...  We evaluate the performance of the method and compare it with a nonprobabilistic method in a simulation study.  ...  It has been theorized, for example, that different age groups exhibit distinct affective responses to stress (Scott, Sliwinski, & Fields, 2013) .  ... 
doi:10.1177/1073191119873714 pmid:31516030 fatcat:7fcprp3tbvfw5d2gtvcy3i3mky
« Previous Showing results 1 — 15 out of 213 results