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.
Derives Mirror Descent from gradient flow on a Riemannian manifold.
problem No specific problem stated; focuses on derivation.
method Derives Mirror Descent from gradient flow on a Riemannian manifold with a natural discretization.
result Generalizes Mirror Descent to non-Hessian metrics.
New methods using natural gradient for structured optimization.
problem Structured optimization problems.
method Structured second-order methods via natural gradient descent.
result Efficiency demonstrated on non-convex and deep learning problems.
We introduce a simple algorithm, True Asymptotic Natural Gradient Optimization (TANGO), that converges to a true natural gradient descent in the limit of small learning rates, without explicit Fisher matrix estimation. For quadratic models the algorithm is also an instance of averaged stochastic gradient, where the par…
This work proposes a new method for variational inference using Wasserstein gradient descent.
problem Optimizing variational parameters to match a true posterior distribution.
method Reinterpreting VI as an optimization problem over a variational parameter space, using Wasserstein gradient descent.
result The proposed Wasserstein gradient descent can be seen as a generalization of existing optimization techniques in VI.
NGD improves multivariate Gaussian inference by optimizing Fisher information.
problem Efficiently optimizing multivariate Gaussian models.
method Natural Gradient Descent applied to multivariate Gaussian parameters.
result NGD updates are more efficient for symmetric covariance matrices.
Optimizes graph neural networks using natural gradient descent.
problem Improving efficiency and performance of graph neural networks.
method Employing natural gradient descent to optimize graph neural networks.
result Natural gradient optimization leads to superior performance compared to existing methods.
Paper proposes ZO-NGD for more efficient black-box attacks.
problem Vulnerability of state-of-the-art DNNs to adversarial attacks.
method Zeroth-order natural gradient descent (ZO-NGD) for black-box attacks.
result ZO-NGD achieves significantly lower model query complexities.
Natural gradient learning improves synaptic plasticity in spiking neurons.
problem Parametrization dependence leads to inconsistencies in classical synaptic plasticity theories.
method Proposes natural gradient descent in Riemannian geometry for spiking neurons.
result Derives a synaptic learning rule that explains biological phenomena.
Proposes a new stochastic optimization method for MLR models.
problem Slow convergence of SGD in big data scenarios.
method Dual Stochastic Natural Gradient Descent (DNSGD) based on manifold optimization.
result DNSGD converges and has linear computational complexity.
Natural gradient descent has proven effective at mitigating the effects of pathological curvature in neural network optimization, but little is known theoretically about its convergence properties, especially for \emph{nonlinear} networks. In this work, we analyze for the first time the speed of convergence of natural …
In this paper, we provide an overview of first-order and second-order variants of the gradient descent method that are commonly used in machine learning. We propose a general framework in which 6 of these variants can be interpreted as different instances of the same approach. They are the vanilla gradient descent, the…
Natural gradient descent is an optimization method traditionally motivated from the perspective of information geometry, and works well for many applications as an alternative to stochastic gradient descent. In this paper we critically analyze this method and its properties, and show how it can be viewed as a type of 2…
We develop a more efficient NGD method for structured parameters.
problem Computational challenges in NGD for structured parameter spaces.
method Local-parameter coordinates to simplify Fisher-matrix computations.
result New structured second-order algorithms and learning methods.
Gradient descent solves robust mean estimation in high dimensions.
problem High-dimensional robust mean estimation in the presence of adversarial outliers.
method Gradient descent with a structural lemma showing near-optimal solutions.
result Gradient descent can solve the robust mean estimation problem directly.
Natural gradient descent, which preconditions a gradient descent update with the Fisher information matrix of the underlying statistical model, is a way to capture partial second-order information. Several highly visible works have advocated an approximation known as the empirical Fisher, drawing connections between ap…
Gradient descent struggles with high-dimensional data fitting.
problem Gradient descent struggles with high-dimensional data fitting.
method Gradient descent training of a two-layer neural network on empirical or population risk.
result Gradient descent training may not decrease population risk faster than t−4/(d−2) under mean field scaling. Proposes Coherent Gradients to explain and reduce overfitting in neural networks.
problem Why neural networks generalize well despite fitting random data.
method Hypothesis about gradient dynamics and a modification to gradient descent.
result Supports hypothesis with heuristic arguments and perturbative experiments.
This paper considers the problem of solving systems of quadratic equations, namely, recovering an object of interest x♮∈Rn from m quadratic equations/samples yi=(ai⊤x♮)2, 1≤i≤m. This problem, also dubbed as phase retrieval, span…
We study the Wasserstein natural gradient in parametric statistical models with continuous sample spaces. Our approach is to pull back the L2-Wasserstein metric tensor in the probability density space to a parameter space, equipping the latter with a positive definite metric tensor, under which it becomes a Riemanni…
Information geometry applies concepts in differential geometry to probability and statistics and is especially useful for parameter estimation in exponential families where parameters are known to lie on a Riemannian manifold. Connections between the geometric properties of the induced manifold and statistical properti…
We use differential equations based approaches to provide some {\it \textbf{physics}} insights into analyzing the dynamics of popular optimization algorithms in machine learning. In particular, we study gradient descent, proximal gradient descent, coordinate gradient descent, proximal coordinate gradient, and Newton's …
Improves understanding of stochastic NGVI convergence rates.
problem Lack of knowledge about non-asymptotic convergence rates in stochastic NGVI.
method Proved non-asymptotic convergence rates for conjugate likelihoods and showed implicit optimization for non-conjugate likelihoods.
result First O(T1) non-asymptotic convergence rate for stochastic NGVI in conjugate likelihoods. Gradient descent achieves exact linear convergence rate for symmetric matrix completion.
problem Low-rank symmetric matrix completion using gradient descent.
method Local analysis of gradient descent for symmetric matrices without additional assumptions.
result Closed-form expression of exact linear convergence rate matches practice.
We present a unifying framework for adapting the update direction in gradient-based iterative optimization methods. As natural special cases we re-derive classical momentum and Nesterov's accelerated gradient method, lending a new intuitive interpretation to the latter algorithm. We show that a new algorithm, which we …
Gradient descent dynamics in nonconvex models explained with universality.
problem Understanding long-time behavior of nonconvex gradient descent.
method Developed a state evolution system for tracking gradient descent iterates.
result Gradient descent iterates are approximately independent of data and strongly incoherent with feature vectors.
Gradient descent training of neural networks leads to solutions close to natural cubic splines.
problem Understanding the implicit bias of gradient descent in neural networks.
method Analysis of gradient descent training for wide neural networks, focusing on the curvature penalty and initialization schemes.
result The solutions of gradient descent training are polyharmonic splines for certain initialization schemes.
We study alignment in linear neural networks and its relation to gradient descent.
problem Understanding alignment in linear neural networks and its impact on training.
method Defined alignment for fully connected networks, analyzed alignment under gradient descent, and compared gradient descent to projected gradient descent for layer-constrained networks.
result Gradient descent can converge linearly to a global minimum when alignment is invariant, and alignment is impossible with large datasets in layer-constrained networks.
We present here a new model and algorithm which performs an efficient Natural gradient descent for Multilayer Perceptrons. Natural gradient descent was originally proposed from a point of view of information geometry, and it performs the steepest descent updates on manifolds in a Riemannian space. In particular, we ext…
A quantum generalization of Natural Gradient Descent is presented as part of a general-purpose optimization framework for variational quantum circuits. The optimization dynamics is interpreted as moving in the steepest descent direction with respect to the Quantum Information Geometry, corresponding to the real part of…
The natural gradient allows for more efficient gradient descent by removing dependencies and biases inherent in a function's parameterization. Several papers present the topic thoroughly and precisely. It remains a very difficult idea to get your head around however. The intent of this note is to provide simple intuiti…
Recent work has argued that stochastic gradient descent can approximate the Bayesian uncertainty in model parameters near local minima. In this work we develop a similar correspondence for minibatch natural gradient descent (NGD). We prove that for sufficiently small learning rates, if the model predictions on the trai…
Paper shows robustness of gradient descent in matrix sensing despite perturbations.
problem Understanding robustness of gradient descent in matrix sensing.
method Developed perturbed gradient flow to capture noise and improve robustness.
result Gradient descent is robust to perturbations in matrix sensing.
Square-root natural-gradient improves variational inference convergence.
problem Challenges in establishing theoretical convergence guarantees for natural-gradient descent.
method Square-root parameterization for Gaussian covariance.
result Establishes novel convergence guarantees for natural-gradient Gaussian inference.
New method learns good initialization for gradient descent from past solutions.
problem Challenges in solving non-convex optimization problems.
method Learning good initialization rules from previous solutions.
result Our approach performs better than random initialization in various non-convex problems.
Unified signSGD and gradient descent analysis for neural networks.
problem Performance of sign-based optimization methods in neural networks.
method Unified analysis of separable smoothness and ℓ∞-smoothness, isolating geometric properties affecting performance. result Sign-based methods are preferable over gradient descent under specific Hessian properties in deep networks.
CBO interprets as SGD, leading to global convergence for nonconvex functions.
problem Understanding and improving gradient-based learning algorithms.
method Interpreting CBO as a stochastic relaxation of SGD.
result CBO provably converges globally to minimizers for nonsmooth nonconvex functions.
Optimization algorithms that leverage gradient covariance information, such as variants of natural gradient descent (Amari, 1998), offer the prospect of yielding more effective descent directions. For models with many parameters, the covariance matrix they are based on becomes gigantic, making them inapplicable in thei…
New analysis shows GMD can converge linearly under PL-like conditions.
problem Establishing linear convergence for generalized mirror descent.
method PL-based analysis for time-dependent mirrors, Taylor-series approach for stochastic GMD.
result Linear convergence of stochastic GMD under PL-like conditions.
Algorithm learns which weights to share in deep multi-task learning.
problem Difficulty in deciding which weights to share between tasks in deep learning models.
method Combines natural evolution strategy and stochastic gradient descent to learn optimal weight sharing.
result Task-specific networks achieve lower test errors than existing methods on multi-task learning datasets.
Large deviations theory applied to policy gradient methods.
problem Understanding convergence of policy gradient methods in reinforcement learning.
method Large deviation rate function and contraction principle from large deviations theory.
result Convergence properties of policy gradient methods can be extended to various policy parametrizations.
Gradient descent recovers principal components of overparametrized asymmetric matrices without explicit regularization.
problem Asymmetric matrix factorization under overparametrization with minimal rank assumptions.
method Vanilla gradient descent with small random initialization and proper early stopping.
result Gradient descent produces the best low-rank approximation without explicit regularization.
The (stochastic) gradient descent and the multiplicative update method are probably the most popular algorithms in machine learning. We introduce and study a new regularization which provides a unification of the additive and multiplicative updates. This regularization is derived from an hyperbolic analogue of the entr…
MT-SGD samples from multiple target distributions using gradient descent.
problem Sampling from multiple unnormalized target distributions.
method Proposes MT-SGD, a flow of intermediate distributions to sample from multiple target distributions.
result Asymptotic analysis shows MT-SGD reduces to multiple-gradient descent for multi-objective optimization.
A new catnat function improves gradient descent for categorical variables.
problem Gradient descent challenges with discrete latent categorical variables.
method Replaced softmax with catnat, a hierarchical binary split function.
result Catnat function offers significant advantages in gradient descent.
Second-order optimization methods such as natural gradient descent have the potential to speed up training of neural networks by correcting for the curvature of the loss function. Unfortunately, the exact natural gradient is impractical to compute for large models, and most approximations either require an expensive it…
Proves FR-NGD optimally approximates evolutionary dynamics and continuous Bayesian inference.
problem Optimizing continuous time replicator equations and continuous Bayesian inference.
method Fisher-Rao natural gradient descent (FR-NGD) and its correspondence with evolutionary dynamics.
result FR-NGD optimally approximates continuous time replicator equations and continuous Bayesian inference.
BONG optimizes Bayesian inference online with natural gradient descent.
problem Sequential Bayesian inference in online settings.
method Bayesian online natural gradient (BONG) approach based on variational Bayes.
result BONG outperforms other online VB methods in non-conjugate settings.