Geometric Occam's Razor shapes deep learning solutions.
problem Understanding the regularization in over-parameterized neural networks.
method Analyzing the geometric model complexity and Dirichlet energy in neural networks.
result Over-parameterized neural networks are implicitly regularized by geometric model complexity.
Transformers prefer simpler explanations in hierarchical tasks.
problem Navigating tasks with varying complexity levels.
method Well-controlled testbeds based on Markov chains and linear regression.
result Transformers favor the least complex sufficient explanation when presented with simpler data.
The article applies Occam's Razor to non-parametric model building, minimizing the number of bits for data encoding.
problem Overlooking the role of model parameters in data encoding leads to inefficient probability density estimators.
method Extends bit counting to model parameters, providing a true measure of complexity for parametric models.
result Minimizing total bit requirement leads to smoother, more efficient probability density estimates and fewer relevant parameters.
Bayesian model selection can be misled by ELBO under certain conditions.
problem Misleading model selection when using ELBO for Bayesian inference.
method Analysis of ELBO-based hyperparameter learning in a simple regression model.
result Bayesian model selection can prefer overfit models when ELBO is used, contrary to evidence.
Deep neural networks perform well due to a balance of architecture, training, and structured data.
problem Understanding why overparameterized DNNs perform well.
method Bayesian approach to analyze the interplay between network architecture, training algorithms, and data structure.
result Structured data and an inductive bias towards simple functions explain DNN success.
Bayesian approach sparsifies neural networks efficiently.
problem Efficiently pruning neural networks to save resources.
method Sparsifiability via the Marginal likelihood (SpaM) framework.
result Prunes neural networks effectively without significant loss in performance.
New approach to nonuniform learnability using measure theory.
problem Nonuniform learnability of hypotheses with varying sample sizes.
method Measure theoretic approach to redefine nonuniform learnability, introducing a new algorithm (Generalize Measure Learnability).
result Achieved statistical consistency in learning countable hypothesis classes.
The paper improves deep learning generalization bounds using PAC-Bayes compression.
problem Improving generalization bounds for deep neural networks.
method Quantizing neural network parameters in a linear subspace to develop tight generalization bounds.
result Large models can be compressed significantly, explaining Occam's razor.
This work connects symmetries and conserved quantities in machine learning.
problem Improving machine learning models by learning conserved quantities.
method Using Noether's theorem, learn symmetries and conserved quantities directly from data.
result Correctly identifies conserved quantities and improves model performance.
The paper explores how smaller data sets can lead to better model selection decisions.
problem Model selection in small data regimes.
method Empirical study of generalization performance with varying training set sizes.
result Training on smaller subsets of data can lead to more reliable model selection decisions.
OccamNet finds interpretable symbolic fits to data efficiently.
problem Complex neural models extrapolate poorly and are hard to interpret.
method Samples functions, biases towards better fits, and uses cross-entropy matching.
result Outperforms state-of-the-art symbolic regression methods on real-world datasets.
Paper proposes a method to estimate scientific parameters in hybrid models without relying on model architecture.
problem Estimating unknown parameters in hybrid models combining machine learning and scientific models.
method Sharpness-aware minimization adapted for hybrid modeling, focusing on model simplicity.
result Demonstrates effectiveness of SAM-based hybrid model learning for scientific parameter estimation.
MDL principle aids in learning neural network-based causal structures.
problem Learning causal relationships from observations with neural networks.
method Prequential minimum description length (MDL) principle.
result Competitive results on synthetic and real-world data, often recovering correct structure.
This study explores how feature graphs enhance GNNs' performance in modeling interactions.
problem Improving GNNs' ability to model feature interactions effectively.
method Investigates feature graphs and their importance in GNNs, using experiments and theoretical support.
result Edges between interacting features are crucial for GNNs, while non-interaction edges can degrade performance.
ResNets minimize circuit size for fitting data in HTMC regime.
problem Finding the simplest algorithm that fits data.
method Defining HTMC and ResNet norms to relate circuit size and function fitting.
result Minimizing ResNet norm is equivalent to finding a circuit with minimal nodes.
New model estimates sparse transport maps for high-dimensional data.
problem Estimating optimal transport maps in high-dimensional spaces.
method Proposes a new model using a family of translation invariant costs and sparsity-inducing norms.
result Sparse transport maps that apply Occam's razor to reduce complexity.
Bayesian evidence helps compare models but can overfit.
problem Comparing hypotheses consistent with observations.
method Marginal likelihood, Occam's razor, PAC-Bayes bounds.
result Marginal likelihood can negatively correlate with generalization.
Bayesian inference simplified for machine learning models.
problem Difficulty in specifying general prior belief in machine learning architectures.
method Parsimonious inference using information theory and Kolmogorov complexity.
result Framework quantifies model complexity and prediction information, reducing memorization.
New method measures generalizability of deep neural networks based on decision boundary complexity.
problem Lack of generalization methods for deep neural networks.
method Created Decision Boundary Complexity (DBC) score to measure DNN complexity.
result Simpler decision boundaries lead to better generalizability, supporting Occam's Razor.
NLR models often perform worse than LR for outlying input data in environmental sciences.
problem NLR models often give poor predictions for input data outside the training domain.
method Screened input data for outliers, using linear extrapolation for outliers based on NLR within the non-outlier domain.
result NLROR approach reduces poor extrapolation and tends to outperform NLR and LR for outliers. We exhibit a strong link between frequentist PAC-Bayesian risk bounds and the Bayesian marginal likelihood. That is, for the negative log-likelihood loss function, we show that the minimization of PAC-Bayesian generalization risk bounds maximizes the Bayesian marginal likelihood. This provides an alternative explanatio…
Investigates principles of generalization in list learning, refutes sample compression conjecture.
problem Determining applicability of classical principles in list PAC learning.
method Examines uniform convergence and sample compression in list PAC learning.
result Sample compression fails in list PAC learning, refutes conjecture.
Compressed imitation learning uses simplicity priors for efficient expert behavior copying.
problem Efficiently learn expert behaviors with minimal data.
method Utilizes policy simplicity as a prior for sample-efficient imitation learning.
result Significantly higher scores achieved with limited expert demonstrations.
Bayesian nonparametric models, such as Gaussian processes, provide a compelling framework for automatic statistical modelling: these models have a high degree of flexibility, and automatically calibrated complexity. However, automating human expertise remains elusive; for example, Gaussian processes with standard kerne…
We present an asymptotic analysis of Viterbi Training (VT) and contrast it with a more conventional Maximum Likelihood (ML) approach to parameter estimation in Hidden Markov Models. While ML estimator works by (locally) maximizing the likelihood of the observed data, VT seeks to maximize the probability of the most lik…
Study reveals how model volume affects learning curves in machine learning.
problem Understanding the double descent risk phenomenon in machine learning.
method Investigates the role of model volume using MDL, Occam's Razor, and information geometry.
result Model volume can explain the double descent risk, suggesting better generalization with increased dimensionality.
The study examines causal razors and their logical relations, highlighting a dilemma in causal discovery.
problem Selecting a reasonable scoring criterion for causal discovery algorithms.
method Review and logical comparison of numerous causal razors, focusing on parameter minimality in multinomial models.
result Parameter minimality poses a dilemma in selecting a reasonable scoring criterion for causal discovery algorithms.
Why do deep neural networks (DNNs) benefit from very high dimensional parameter spaces? Their huge parameter complexities vs stunning performance in practice is all the more intriguing and not explainable using the standard theory of model selection for regular models. In this work, we propose a geometrically flavored …
The well known maximum-entropy principle due to Jaynes, which states that given mean parameters, the maximum entropy distribution matching them is in an exponential family, has been very popular in machine learning due to its "Occam's razor" interpretation. Unfortunately, calculating the potentials in the maximum-entro…
This paper presents a light-weight and accurate deep neural model for audiovisual emotion recognition. To design this model, the authors followed a philosophy of simplicity, drastically limiting the number of parameters to learn from the target datasets, always choosing the simplest earning methods: i) transfer learnin…
Transfer learning aims at transferring knowledge from a well-labeled domain to a similar but different domain with limited or no labels. Unfortunately, existing learning-based methods often involve intensive model selection and hyperparameter tuning to obtain good results. Moreover, cross-validation is not possible for…
Matrix completion and approximation are popular tools to capture a user's preferences for recommendation and to approximate missing data. Instead of using low-rank factorization we take a drastically different approach, based on the simple insight that an additive model of co-clusterings allows one to approximate matri…
Unified framework for deriving generalization bounds in supervised learning.
problem Generalization error bounds in supervised learning.
method Data Processing Inequality PAC-Bayesian framework.
result Unified bounds on binary Kullback-Leibler generalization gap for various divergences.
We present an integer programming framework to build accurate and interpretable discrete linear classification models. Unlike existing approaches, our framework is designed to provide practitioners with the control and flexibility they need to tailor accurate and interpretable models for a domain of choice. To this end…
Study improves neural network performance in sequential learning for image classification.
problem Improving neural network performance in sequential learning for image classification.
method Evaluation of approaches for computing prequential description lengths, proposing forward-calibration and replay-streams.
result Improved description lengths for image classification datasets, outperforming previous results.
Bayesian interpretation explains double descent in deep learning models.
problem Understanding the risk function behavior of over-parameterized models.
method Bayesian model selection, Dickey-Savage ratio, ridge regression, global-local shrinkage.
result Double descent phenomenon explained through Bayesian interpretation.
The paper explores how complex models can improve system identification beyond traditional limits.
problem Balancing model richness and spurious learning in system identification.
method Investigates the double-descent phenomenon in the context of dynamic systems.
result Complex models can improve system identification performance beyond the point of interpolation.
The recent empirical success of unsupervised cross-domain mapping algorithms, between two domains that share common characteristics, is not well-supported by theoretical justifications. This lacuna is especially troubling, given the clear ambiguity in such mappings. We work with adversarial training methods based on IP…
One of the objectives of designing feature selection learning algorithms is to obtain classifiers that depend on a small number of attributes and have verifiable future performance guarantees. There are few, if any, approaches that successfully address the two goals simultaneously. Performance guarantees become crucial…
Bayesian framework detects symmetries in chaotic dynamical systems.
problem Detecting symmetries in chaotic attractors for insights into dynamical system structure.
method Bayesian framework using Gibbs posterior constructed from Wasserstein distances.
result Bayesian framework accurately recovers symmetries under high noise and small sample sizes.
We consider the problem of identifying the causal direction between two discrete random variables using observational data. Unlike previous work, we keep the most general functional model but make an assumption on the unobserved exogenous variable: Inspired by Occam's razor, we assume that the exogenous variable is sim…
Bayesian model averaging has become a widely used approach to accounting for uncertainty about the structural form of the model generating the data. When data arrive sequentially and the generating model can change over time, Dynamic Model Averaging (DMA) extends model averaging to deal with this situation. Often in ma…
In this article, we investigate large sample properties of model selection procedures in a general Bayesian framework when a closed form expression of the marginal likelihood function is not available or a local asymptotic quadratic approximation of the log-likelihood function does not exist. Under appropriate identifi…
Leo Breiman's Rashomon Effect and Occam Dilemma are re-evaluated in the context of modern machine learning.
problem The tradeoff between model complexity and accuracy in machine learning.
method Modern perspective on Breiman's arguments using current computational capabilities.
result Algorithmic models can be accurate without being complex, nullifying the Occam Dilemma.
Machine learning selects the best prediction rules from noisy data.
problem Selection under uncertainty in machine learning.
method Statistical tools and inequalities to control noise in empirical estimates.
result Theoretical guarantees on selection outcomes under uncertainty.
This work characterizes optimal multiclass learning with regularization.
problem The empirical risk minimization (ERM) algorithm fails in multiclass learning settings.
method Using one-inclusion graphs (OIGs), the work introduces optimal learning algorithms that relax structural risk minimization and incorporate unsupervised learning.
result An optimal learner is introduced that uses a local regularization function and an unsupervised learning stage to learn the regularizer.
New estimate reduces overfitting risk in machine learning models.
problem Error rate on test data may not reflect true population error due to adaptive data analysis practices.
method Introduces Rip van Winkle's Razor, a simple estimate of overfit to test data based on information content.
result Shows non-vacuous estimate of deviation in many modern settings.
We introduce the problem of learning mixtures of k subcubes over {0,1}n, which contains many classic learning theory problems as a special case (and is itself a special case of others). We give a surprising nO(logk)-time learning algorithm based on higher-order multilinear moments. It is not possible to l…