The paper addresses the invariance issue in Bayesian neural networks using linearized Laplace approximation.
problem Bayesian neural networks fail to maintain invariance under reparameterization, leading to different posterior densities for identical functions.
method Developed a geometric view of reparameterizations and a Riemannian diffusion process to extend reparameterization invariance to neural network predictive.
result Empirically improved posterior fit through approximate posterior sampling.
The performance of deep neural networks is often attributed to their automated, task-related feature construction. It remains an open question, though, why this leads to solutions with good generalization, even in cases where the number of parameters is larger than the number of samples. Back in the 90s, Hochreiter and…
This paper focuses on the study of open curves in a manifold M, and proposes a reparameterization invariant metric on the space of such paths. We use the square root velocity function (SRVF) introduced by Srivastava et al. in [11] to define a reparameterization invariant metric on the space of immersions M' = Imm([0,1]…
DiMS sampler explores neural network loss minima via dissipative dynamics.
problem Sampling reparameterization invariant solutions in neural networks.
method Dynamical system based on kinetic energy with dissipative friction.
result DiMS sampler samples exactly from minimum level sets.
Natural gradient descent is a robust optimization method for machine learning.
problem Training poorly parameterized networks efficiently.
method Optimization algorithms with natural transformation properties.
result Optimization algorithms with natural transformation properties are more efficient for poorly parameterized networks.
We investigate a local reparameterizaton technique for greatly reducing the variance of stochastic gradients for variational Bayesian inference (SGVB) of a posterior over model parameters, while retaining parallelizability. This local reparameterization translates uncertainty about global parameters into local noise th…
Geometric correspondence links flow metrics to reparameterizations.
problem Linking flow metrics to reparameterizations of geodesic flows.
method Analysis of Mineyev's flow space and Green metrics.
result First examples of continuous reparameterizations on negatively curved manifolds.
The reparameterization gradient has become a widely used method to obtain Monte Carlo gradients to optimize the variational objective. However, this technique does not easily apply to commonly used distributions such as beta or gamma without further approximations, and most practical applications of the reparameterizat…
Investigates invariant hulls of functionals on manifolds.
problem Understanding meaningful functionals on manifolds through reparameterizations.
method Uses inner-variations to transform arbitrary functionals into invariant realizations.
result Explicit computations for volume functional in N-dimensional manifolds, especially in N=2. Reparameterizes mirror descent as gradient descent for efficient sparse learning.
problem Efficiently training small sparse networks with mirror descent.
method Develops a framework to convert mirror descent updates into gradient descent updates on different parameters.
result Mirror descent can be reparameterized as gradient descent on modified parameters, facilitating standard backpropagation.
REP-GAN improves GANs by reparameterizing proposals for better sample quality and efficiency.
problem Poor sample efficiency in GANs due to independent proposal sampling.
method REParameterizing Markov chains into the latent space of the generator to create dependent proposals.
result Empirically shows significant improvement in sample efficiency and quality.
Variational inference using the reparameterization trick has enabled large-scale approximate Bayesian inference in complex probabilistic models, leveraging stochastic optimization to sidestep intractable expectations. The reparameterization trick is applicable when we can simulate a random variable by applying a differ…
The reparameterization trick has become one of the most useful tools in the field of variational inference. However, the reparameterization trick is based on the standardization transformation which restricts the scope of application of this method to distributions that have tractable inverse cumulative distribution fu…
Geometric approach improves motion alignment accuracy and efficiency.
problem Temporal alignment of human motion data for various applications.
method Geometric point of view, principal fiber bundle, reparameterization invariant projection, dynamic programming, keyframe correspondences.
result Temporal alignment procedures are more accurate and computationally efficient.
By providing a simple and efficient way of computing low-variance gradients of continuous random variables, the reparameterization trick has become the technique of choice for training a variety of latent variable models. However, it is not applicable to a number of important continuous distributions. We introduce an a…
LEARN-SAM improves RL from sub-optimal demonstrations by localizing expert policies and selectively using demonstrations.
problem Improving RL from sub-optimal or sparse demonstrations.
method Local Ensemble and Reparameterization with Split and Merge of expert policies (LEARN-SAM).
result LEARN-SAM boosts learning speed and accuracy by selectively using demonstrations.
A new method for optimizing models with categorical variables using diffusion.
problem Optimizing models with categorical variables, especially in discrete distributions.
method Introducing ReDGE, a diffusion-based soft reparameterization method for categorical distributions.
result ReDGE consistently matches or outperforms existing gradient-based methods in experiments.
Two new estimators improve VAE training for hierarchical and prior parameters.
problem Efficient gradient estimation for VAEs with hierarchical and prior parameters.
method Developed two generalizations of Doubly-Reparameterized Gradient Estimators (DReGs) for VAEs.
result Improved training of conditional and hierarchical VAEs on image modeling tasks.
We observe that gradients computed via the reparameterization trick are in direct correspondence with solutions of the transport equation in the formalism of optimal transport. We use this perspective to compute (approximate) pathwise gradients for probability distributions not directly amenable to the reparameterizati…
The reparameterization trick is widely used in variational inference as it yields more accurate estimates of the gradient of the variational objective than alternative approaches such as the score function method. Although there is overwhelming empirical evidence in the literature showing its success, there is relative…
Low-variance gradient estimation is crucial for learning directed graphical models parameterized by neural networks, where the reparameterization trick is widely used for those with continuous variables. While this technique gives low-variance gradient estimates, it has not been directly applicable to discrete variable…
Most neural networks are trained using first-order optimization methods, which are sensitive to the parameterization of the model. Natural gradient descent is invariant to smooth reparameterizations because it is defined in a coordinate-free way, but tractable approximations are typically defined in terms of coordinate…
A new method for stochastic optimal control improves accuracy over existing techniques.
problem Improving the accuracy of stochastic optimal control for noisy systems.
method Stochastic Optimal Control Matching (SOCM) using Iterative Diffusion Optimization (IDO) with path-wise reparameterization trick.
result SOCM achieves lower error than existing techniques for three out of four control problems, sometimes by an order of magnitude.
New measure of maximal entropy found for a class of geometrically finite groups.
problem Finding a measure of maximal entropy for relatively Anosov groups.
method Constructing reparameterizations and using exponential expansion along unstable foliations.
result The Bowen-Margulis-Sullivan measure is finite and unique for relatively Anosov groups.
Paper presents a reparameterized DP-DLGMM for clustering.
problem Non-parametric DP priors in DLGMM are hard to couple with variational inference.
method Closed-form updates for DP-DLGMM's variational posterior.
result Model generates realistic samples and performs competitively in semi-supervised settings.
We consider spaces of smooth immersed plane curves (modulo translations and/or rotations), equipped with reparameterization invariant weak Riemannian metrics involving second derivatives. This includes the full H2-metric without zero order terms. We find isometries (called R-transforms) from some of these spaces i…
PeL separates sensory interface optimization from decision learning.
problem Optimizing sensory interfaces without task-specific information.
method Formal separation of perception and decision learning, using metrics for stability, informativeness, and geometry.
result Updates preserving invariants are orthogonal to decision gradients.
The notion of Nonlocal Mean Curvature (NMC) appears recently in the mathematics literature. It is an extrinsic geometric quantity that is invariant under global reparameterization of a surface and provide a natural extension of the classical mean curvature. We describe some properties of the NMC and the quasilinear dif…
The importance of the fusion relation of loops was recognized in the context of spin structures on the loop space by Stolz and Teichner and further developed by Waldorf. On a spin manifold M the equivalence classes of `fusive' spin structures on the loop space LM, incorporating the fusion property, strong regularity an…
Natural gradient descent avoids the magic of model parametrization, leading to different optimization outcomes.
problem Understanding the impact of model parametrization on optimization and generalization in deep learning.
method Characterization of natural gradient flow in deep linear networks and nonlinear neural networks.
result Natural gradient descent fails to generalize in some cases, while gradient descent with the right architecture performs well.
Bayesian optimization is a sample-efficient approach to solving global optimization problems. Along with a surrogate model, this approach relies on theoretically motivated value heuristics (acquisition functions) to guide the search process. Maximizing acquisition functions yields the best performance; unfortunately, t…
Conventional prior for Variational Auto-Encoder (VAE) is a Gaussian distribution. Recent works demonstrated that choice of prior distribution affects learning capacity of VAE models. We propose a general technique (embedding-reparameterization procedure, or ER) for introducing arbitrary manifold-valued variables in VAE…
The reparameterization trick enables optimizing large scale stochastic computation graphs via gradient descent. The essence of the trick is to refactor each stochastic node into a differentiable function of its parameters and a random variable with fixed distribution. After refactoring, the gradients of the loss propag…
This research simplifies verification of machine learning systems using reparameterization.
problem Reduce or eliminate serious bugs in machine learning systems.
method Use proof assistants to construct machine-checked proofs of correctness, leveraging reparameterization to handle probabilistic claims.
result Demonstrates broad applicability of reparameterization to verify different types of machine learning systems.
We present a new algorithm for stochastic variational inference that targets at models with non-differentiable densities. One of the key challenges in stochastic variational inference is to come up with a low-variance estimator of the gradient of a variational objective. We tackle the challenge by generalizing the repa…
Recent breakthroughs in computer vision make use of large deep neural networks, utilizing the substantial speedup offered by GPUs. For applications running on limited hardware, however, high precision real-time processing can still be a challenge. One approach to solving this problem is training networks with binary or…
We stabilize the Kumaraswamy distribution for efficient sampling and differentiation.
problem Numerical instabilities in the Kumaraswamy distribution's inverse CDF and log-pdf.
method Identified and resolved numerical issues, introduced a stabilized KS distribution.
result Stabilized Kumaraswamy distribution supports efficient sampling and differentiation.
Optimization with noisy gradients has become ubiquitous in statistics and machine learning. Reparameterization gradients, or gradient estimates computed via the "reparameterization trick," represent a class of noisy gradients often used in Monte Carlo variational inference (MCVI). However, when these gradient estimator…
New method enhances model fine-tuning with minimal data.
problem Improving model performance on new tasks with limited data.
method Introducing α-LoRA, a reparameterization method for fine-tuning. result Enhanced generalization ability of fine-tuned models.
The Gumbel-Softmax is a continuous distribution over the simplex that is often used as a relaxation of discrete distributions. Because it can be readily interpreted and easily reparameterized, it enjoys widespread use. We propose a modular and more flexible family of reparameterizable distributions where Gaussian noise…
Unified view of LR and RP gradients explained via divergence theorem.
problem Explaining the nature and relationship of LR and RP gradients.
method First principles approach using divergence theorem.
result Characterization of all possible estimators combining LR and RP.
Paper tackles performative prediction without convexity assumptions.
problem Performative prediction where data distribution changes with model deployment.
method Reparameterization framework to transform non-convex objective into convex one.
result Provably sublinear regret guarantees for learnable model.
Monge SAM improves deep learning by making sharpness-aware minimization invariant to reparametrizations.
problem Non-invariance of sharpness-aware minimization (SAM) to reparametrizations.
method Introduces Monge SAM, a reparametrization-invariant version of SAM using a Riemannian metric.
result Monge SAM enhances robustness and generalization compared to previous methods.
EXPO framework eliminates need for reward model, achieving better optimization.
problem Optimizing large language model responses without a separate reward model.
method Introduces EXPO framework that avoids reparameterization, directly optimizing preferences.
result Demonstrates better regularization and intuitive interpolation behaviors.
New method for analyzing learning dynamics in singular models.
problem Challenges in analyzing learning of singular models with no one-to-one parameter space.
method Relative reparameterization technique to extract regular sub-models.
result Demonstrated differences in convergence behavior due to algorithmic and intrinsic aspects.
New method improves imitation learning from expert observations.
problem Challenges in imitation learning from observation setting.
method Reparameterized Variational Divergence Minimization.
result Our method outperforms baseline approaches in low-dimensional tasks.
Reparameterizable densities are an important way to learn probability distributions in a deep learning setting. For many distributions it is possible to create low-variance gradient estimators by utilizing a `reparameterization trick'. Due to the absence of a general reparameterization trick, much research has recently…
Gradient descent implicitly favors group sparsity in neural networks.
problem Understanding implicit regularization in neural networks for structured sparsity.
method Novel neural reparameterization for diagonally grouped linear networks.
result Gradient descent without explicit regularization biases towards group sparsity.