Improved LLM pre-training performance through better weight and variance control.
problem Improper weight and variance control in LLM pre-training affects downstream task performance.
method Introduced Layer Index Rescaling (LIR) and Target Variance Rescaling (TVR) techniques.
result Substantial improvements in downstream task performance (up to 4.6%) and reduced extreme activation values.
Optimizes deep neural network initialization variance for better performance.
problem Improving deep neural network performance through optimal initialization variance.
method Using SGD dynamics and Fokker-Planck equations, we study the relationship between initialization and expected loss function.
result An optimal condition for initialization variance that leads to lower training loss and higher test accuracy.
Before training a neural net, a classic rule of thumb is to randomly initialize the weights so the variance of activations is preserved across layers. This is traditionally interpreted using the total variance due to randomness in both weights \emph{and} samples. Alternatively, one can interpret the rule of thumb as pr…
This paper examines weight initialization for 1-Lipschitz networks to improve robustness against adversarial attacks.
problem Improving the robustness of deep neural networks against adversarial attacks.
method Examined weight parametrization of AOL and SLL networks, calculated weight variance bounds, and demonstrated weight decay.
result Weight initialization causes deep 1-Lipschitz networks to decay to zero, and weight variance does not affect output variance distribution.
Deep learning models can have low bias and variance, contrary to classical theory.
problem Understanding the performance of deep learning models at high complexity.
method Developed a fine-grained bias-variance decomposition for random feature kernel regression, analyzing the effects of sampling, initialization, and labels.
result The variance terms exhibit non-monotonic behavior and can diverge at the interpolation boundary, even in the absence of label noise.
EVA adapts LoRA for faster, more efficient fine-tuning.
problem Fast and efficient fine-tuning of large models for specific tasks.
method EVA uses directions capturing most activation variance for initialization, maximizing gradient signal and reducing parameters.
result EVA achieves faster convergence and higher average scores across tasks, reducing parameters.
Optimizes wide low-rank neural networks for reduced parameters and cost.
problem Reducing the number of learnable parameters in wide neural networks.
method Analyzed edge-of-chaos dynamics and derived formulae for optimal weight and bias variances.
result Optimal weight and bias variances for low-rank networks follow from multiplicative scaling.
DLNs dynamics change with variance, leading to saddle-to-saddle training phases.
problem Understanding the dynamics of DLNs with varying initialization variance.
method Analyzing the phase transition of DLNs' dynamics as variance changes.
result Gradient descent visits a sequence of saddles, reaching a sparse global minimum.
New model reveals significant impact of data and parameter variations on machine learning benchmarks.
problem Variance in machine learning benchmarks due to data sampling, initialization, and hyperparameters.
method Modeling the benchmarking process, analyzing comparison methods, and studying error rates.
result Adding more sources of variation reduces compute cost while improving performance.
This work uses ANOVA to understand how different factors contribute to test error in machine learning models.
problem Understanding why overparametrized models generalize well despite potentially fitting noise.
method Analysis of variance (ANOVA) to decompose test error into components of variance.
result The interaction between training samples and initialization can dominate variance, and there are phase transitions in variance behavior.
Study on fluctuations in neural network kernels and predictions, focusing on finite width effects.
problem Characterizing fluctuations in finite width neural networks.
method Dynamical mean field theory analysis of wide but finite feature learning neural networks.
result Fluctuations in kernels and predictions are dynamically coupled, leading to reduced variance in feature learning regimes.
New initialization schemes preserve fractional moments of weights in deep networks, improving training and test performance.
problem Heavy-tailed distribution of stochastic gradients in DNNs during training.
method Developed initialization schemes that preserve any given fractional moment of order s < 2 over layers for various activations.
result The network output admits a heavy-tailed distribution with finite moments, improving training and test performance.
A new method designs batches for Bayesian optimization more efficiently.
problem Efficiently designing batches for Bayesian optimization to reduce total time.
method Minimal Terminal Variance (MTV) acquisition function, optimizing I-optimality criterion.
result MTV designs batches more efficiently than other methods, as shown by numerical experiments.
The performance of the Self-Organizing Map (SOM) algorithm is dependent on the initial weights of the map. The different initialization methods can broadly be classified into random and data analysis based initialization approach. In this paper, the performance of random initialization (RI) approach is compared to that…
A new ratio, the Hansen ratio, simplifies mean-variance portfolio theory.
problem Simplifying mean-variance portfolio theory.
method Introducing the Hansen ratio and extending mean-variance theory.
result The Hansen ratio provides a parsimonious description of the mean-variance efficient frontier.
We identify and study two common failure modes for early training in deep ReLU nets. For each we give a rigorous proof of when it occurs and how to avoid it, for fully connected and residual architectures. The first failure mode, exploding/vanishing mean activation length, can be avoided by initializing weights from a …
Quantum circuit models learn better with specific initialization strategies.
problem Understanding and improving the optimization landscape of IQP-based generative models.
method Proved barren plateaus for random initialization, established lower bounds, and developed data-dependent initialization.
result Data-dependent initialization leads to faster convergence and better minimums.
New strategies for identifying the best arm in bandits with decreasing variances.
problem Best arm identification in bandits with time-varying variances.
method Two policies: initial wait followed by continuous sampling, and periodic sampling with weighted average.
result Analytical guarantees and simulations show improved performance over existing methods.
We consider the mean-variance hedging problem under partial Information. The underlying asset price process follows a continuous semimartingale and strategies have to be constructed when only part of the information in the market is available. We show that the initial mean variance hedging problem is equivalent to a ne…
New method solves continuous time mean-variance model for consistent investment strategy.
problem Time-consistent optimal strategy for continuous time mean-variance model.
method Developed a new Bellman principle method.
result Obtained a time-consistent dynamic optimal strategy.
New algorithms reduce regret in both stochastic and deterministic environments.
problem Designing algorithms that perform well in both types of MDPs.
method Proposed new environment norms and algorithms with variance-dependent regret bounds.
result First algorithm with simultaneously optimal bounds for both stochastic and deterministic MDPs.
New method shows random, diverse initializations are not essential for deep neural networks.
problem The necessity of random, diverse initializations in deep neural networks.
method Constructed a deep convolutional network with identical features by initializing weights to 0, enabling signal propagation and stable gradients.
result Random, diverse initializations are not necessary for training neural networks.
Dealing with high variance is a significant challenge in model-free reinforcement learning (RL). Existing methods are unreliable, exhibiting high variance in performance from run to run using different initializations/seeds. Focusing on problems arising in continuous control, we propose a functional regularization appr…
Based on the property that solving the system of linear matrix equations via the column space and the row space projections boils down to an approximation in the least squares error sense, a formulation for learning the weight matrices of the multilayer network can be derived. By exploiting into the vast number of feas…
Linear RNNs exhibit a bias towards shorter memory due to initialization variance.
problem Understanding the performance limitations of RNNs, especially linear ones.
method Kernel regime analysis to show equivalence to 1D-convolutional networks and analyze weightings.
result Linear RNNs with random initialization have a bias towards shorter memory periods.
It has been noted in existing literature that over-parameterization in ReLU networks generally improves performance. While there could be several factors involved behind this, we prove some desirable theoretical properties at initialization which may be enjoyed by ReLU networks. Specifically, it is known that He initia…
Residual networks (ResNet) and weight normalization play an important role in various deep learning applications. However, parameter initialization strategies have not been studied previously for weight normalized networks and, in practice, initialization methods designed for un-normalized networks are used as a proxy.…
Optimal insurance contract limits insurer's risk exposure variance.
problem Designing an optimal insurance contract limiting insurer's risk exposure variance.
method Derive optimal policy semi-analytically, focusing on actuarially fair case.
result Expected coverage is larger for wealthier insured, indicating normal good.
Deep networks generalize well even when they fit training data perfectly, thanks to overparametrization.
problem Understanding generalization in overparametrized deep networks.
method Random features regression, asymptotic analysis, ensemble averaging.
result Bias remains constant beyond the interpolation threshold, while variance components decay with overparametrization.
New CNN initialization scheme derived from modern architectures.
problem Stability of CNN model parameters initialization.
method Derived new initialization scheme from modern CNN architectures.
result New initialization method outperforms de facto standard schemes.
Our study analyzes how neural network initialization affects privacy and utility in overparameterized models.
problem Privacy and utility trade-off in overparameterized neural networks.
method Analytical proof of KL divergence privacy bound, focusing on initialization, width, and depth.
result Privacy bound improvement with increasing depth under certain initializations, degradation under others.
We propose a variance reduction framework for variational inference using the Multilevel Monte Carlo (MLMC) method. Our framework is built on reparameterized gradient estimators and "recycles" parameters obtained from past update history in optimization. In addition, our framework provides a new optimization algorithm …
Proposes a new method to initialize neural networks by estimating global curvature of weights.
problem Improving the initialization of neural networks for better training and convergence.
method Estimates the global curvature of weights across layers using the Hessian matrix norm.
result The proposed method helps in more rigorously initializing weights, leading to better performance.
Method estimates noise variance in Gaussian process regression.
problem Estimating noise variance in Gaussian process regression models.
method Reduces hyperparameter space, uses marginal likelihood function, derives bounds and asymptotes.
result Computational advantages and robustness compared to traditional methods.
We address the problem of estimating statistics of hidden units in a neural network using a method of analytic moment propagation. These statistics are useful for approximate whitening of the inputs in front of saturating non-linearities such as a sigmoid function. This is important for initialization of training and f…
The paper confirms a conjecture about optimal expected utility in markets with insider information.
problem Optimal expected utility in markets with insider information.
method An extension of the Black-Scholes-Merton model with a sequence of discrete-time economies.
result Optimal expected utility converges to the classic model when conditions are met.
Recent developments in system identification have brought attention to regularized kernel-based methods, where, adopting the recently introduced stable spline kernel, prior information on the unknown process is enforced. This reduces the variance of the estimates and thus makes kernel-based methods particularly attract…
Bayesian neural networks (BNNs) hold great promise as a flexible and principled solution to deal with uncertainty when learning from finite data. Among approaches to realize probabilistic inference in deep neural networks, variational Bayes (VB) is theoretically grounded, generally applicable, and computationally effic…
Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR) are two risk measures which are widely used in the practice of risk management. This paper deals with the problem of computing both VaR and CVaR using stochastic approximation (with decreasing steps): we propose a first Robbins-Monro procedure based on Rockaffela…
We study the dependence of volatility on the stock price in the stochastic volatility framework on the example of the Heston model. To be more specific, we consider the conditional expectation of variance (square of volatility) under fixed stock price return as a function of the return and time. The behavior of this fu…
A new estimator for evaluating policies in unknown environments.
problem Evaluating policies when both logging policy and value function are unknown.
method Doubly-Robust (DR) off-policy evaluation (OPE) estimator, DRUnknown, that estimates both the logging policy and value function.
result DRUnknown achieves the smallest asymptotic variance and is optimal when both models are correctly specified.
We give a rigorous analysis of the statistical behavior of gradients in a randomly initialized fully connected network N with ReLU activations. Our results show that the empirical variance of the squares of the entries in the input-output Jacobian of N is exponential in a simple architecture-dependent constant beta, gi…
We consider the optimization of a quadratic objective function whose gradients are only accessible through a stochastic oracle that returns the gradient at any given point plus a zero-mean finite variance random error. We present the first algorithm that achieves jointly the optimal prediction error rates for least-squ…
Many neural networks use the tanh activation function, however when given a probability distribution as input, the problem of computing the output distribution in neural networks with tanh activation has not yet been addressed. One important example is the initialization of the echo state network in reservoir computing…
Linking output sensitivity to deep learning generalization.
problem Understanding and comparing the generalization properties of deep neural networks.
method Linking the loss function to output sensitivity and analyzing its relation to bias-variance decomposition.
result Output sensitivity is a strong metric for comparing generalization performance of deep networks.
Exact solutions reveal how unbalanced initializations promote rapid feature learning in neural networks.
problem Understanding how neural networks efficiently extract features from data.
method Deriving exact solutions to a minimal model of neural networks transitioning between lazy and rich learning regimes.
result Unbalanced layer-specific initialization variances and learning rates determine the degree of feature learning.
This paper analyzes implicit bias in Deep Linear Discriminant Analysis.
problem The implicit bias of Deep Linear Discriminant Analysis.
method Analyzing gradient flow on a L-layer diagonal linear network.
result Under balanced initialization, the network transforms additive updates into multiplicative updates, conserving the (2/L) quasi-norm.
RL and DTSOC for final quadratic hedging performance studied.
problem Optimal hedging of European call options with and without transaction costs.
method Reinforcement Learning and Deep Trajectory-based Stochastic Optimal Control.
result RL and DTSOC perform similarly to variance-optimal hedging in various market models.