Study proposes memory-efficient backpropagation for linear layers in neural networks.
problem Significant memory usage in backpropagation through linear layers in neural networks.
method Randomized matrix multiplications to reduce memory usage with a moderate decrease in test accuracy.
result Demonstrated benefits of the proposed method on fine-tuning pre-trained models.
PLN-Nets with two linear layers and parallel LN achieve universal approximation.
problem Limitations of standard neural network architectures in universal approximation.
method Introduced PLN-Nets combining two linear layers with parallel LN.
result PLN-Nets achieve universal approximation, while standard LN has limited power.
The study estimates the expressiveness of GCNs with bounds on the number of linear regions.
problem Characterizing the expressiveness of graph convolutional networks (GCNs).
method Estimates the number of linear regions for one-layer and multi-layer GCNs.
result GCNs with multiple layers have exponentially more expressivity per parameter than one-layer GCNs.
Capacity analysis has been recently introduced as a way to analyze how linear models distribute their modelling capacity across the input space. In this paper, we extend the notion of capacity allocation to the case of neural networks with non-linear layers. We show that under some hypotheses the problem is equivalent …
This article demonstrates that convolutional operation can be converted to matrix multiplication, which has the same calculation way with fully connected layer. The article is helpful for the beginners of the neural network to understand how fully connected layer and the convolutional layer work in the backend. To be c…
Improved recommendation systems using multi-layer embeddings reduce model size while maintaining accuracy.
problem Improving model accuracy in recommendation systems while minimizing model size.
method Introducing a multi-layer embedding training (MLET) architecture that trains embeddings via a sequence of linear layers.
result Substantial advantages in model accuracy and memory footprint are achieved with reduced embedding dimensions.
Despite the phenomenal success of deep learning in recent years, there remains a gap in understanding the fundamental mechanics of neural nets. More research is focussed on handcrafting complex and larger networks, and the design decisions are often ad-hoc and based on intuition. Some recent research has aimed to demys…
ReLU networks trained with MILPs match deep learning accuracy.
problem Training deep neural networks efficiently.
method Iterative training with Mixed Integer Linear Programs (MILPs).
result ReLU networks can be trained with MILPs achieving similar accuracy to deep learning methods.
New basis for permutation equivariant layers reduces computation costs.
problem Efficiently computing permutation equivariant layers in neural networks.
method Generalized partition algebra basis with low-rank tensors.
result Low-rank tensors enable faster computation compared to orbit basis.
SAEs struggle with curved activation manifolds, revealing layer-dependent scaling laws.
problem Sparse autoencoders' reconstruction error varies across layers, not fitting existing scaling laws.
method Cross-layer study of 844 SAE checkpoints, fitting and regressing on manifold geometry.
result Manifold geometry predicts layer-dependent width exponents in SAEs, with transferable coefficients.
Last-layer approximation improves UQ performance without sacrificing computational efficiency.
problem Epistemic uncertainty quantification for deep neural networks.
method Comparison of full-network and last-layer linearization using theoretical and empirical approaches.
result Last-layer approximation yields comparable UQ performance with improved computational efficiency.
The paper generalizes equivariant neural networks on homogeneous spaces to the non-linear setting.
problem Equivariant neural networks on homogeneous spaces.
method Deriving generalized steerability constraints for non-linear equivariant layers.
result The universality of the derived construction for non-linear equivariant layers.
This paper explains double descent in linear neural networks, identifying new factors.
problem Understanding double descent in linear neural networks.
method Gradient flow derivation and necessary conditions for double descent.
result Singular values of input-output covariance matrix are important for double descent in two-layer models.
Study clarifies Bayesian generalization error in CBM for 3-layered linear neural networks.
problem Understanding the generalization error in concept bottleneck models.
method Mathematical analysis of Bayesian generalization error and free energy in CBM for 3-layered linear neural networks.
result CBM significantly alters the parameter region and Bayesian generalization error compared to standard models.
In response to the development of recent efficient dense layers, this paper shows that something as simple as replacing linear components in pointwise convolutions with structured linear decompositions also produces substantial gains in the efficiency/accuracy tradeoff. Pointwise convolutions are fully connected layers…
Wide neural networks become linear, but adding bottlenecks makes them bilinear or multilinear.
problem Understanding the transition of neural networks from linearity to higher-order functions.
method Analyzing the behavior of randomly initialized wide neural networks with and without bottleneck layers.
result Bottleneck layers transform the network's function from linear to bilinear or multilinear.
We propose a new notion of `non-linearity' of a network layer with respect to an input batch that is based on its proximity to a linear system, which is reflected in the non-negative rank of the activation matrix. We measure this non-linearity by applying non-negative factorization to the activation matrix. Considering…
ButterflyFlow uses butterfly matrices for efficient invertible layers in normalizing flows.
problem Building efficient invertible layers for complex probability distributions.
method Proposes butterfly layers for invertible linear layers, leveraging their ability to capture complex structures.
result ButterflyFlow achieves strong density estimation and significantly better log-likelihoods on various datasets.
We consider deep linear networks with arbitrary convex differentiable loss. We provide a short and elementary proof of the fact that all local minima are global minima if the hidden layers are either 1) at least as wide as the input layer, or 2) at least as wide as the output layer. This result is the strongest possibl…
Transformer attention layers solve single-location regression tasks.
problem Understanding token-wise sparsity and internal linear representations in attention-based models.
method Introduce single-location regression task and a simplified predictor based on self-attention layers.
result Transformer attention layers are asymptotically Bayes optimal and can learn underlying structures effectively.
EigenGAN discovers interpretable dimensions in GAN layers for semantic control.
problem Lack of explicit dimensions to control semantic attributes in GAN layers.
method EigenGAN embeds linear subspaces with orthogonal bases into each generator layer, learning eigen-dimensions corresponding to semantic attributes via adversarial training.
result EigenGAN can produce samples with continuous changes corresponding to specific semantic attributes.
Wavelets help compress neural networks efficiently.
problem Efficiently compressing linear layers in neural networks.
method Learnable wavelet transforms to compress RNNs.
result Wavelet compressed RNNs have fewer parameters and perform competitively.
This work analyzes how bottleneck layers and skip connections affect linear denoising autoencoders' generalization.
problem Understanding the generalization of linear denoising autoencoders in overparameterized regimes.
method Analyzes two-layer linear denoising autoencoders with a bottleneck layer and skip connection, deriving test risk formulas.
result Bottleneck layers introduce an additional complexity measure, while skip connections can mitigate variance.
This paper explores loss landscapes of sparse neural networks, finding unique characteristics compared to dense networks.
problem Understanding the loss landscape of sparse neural networks, especially one-hidden-layer networks.
method Analyzes sparse networks with dense and sparse final layers, focusing on linear and non-linear models.
result Sparse networks can have no spurious valleys under certain conditions, but spurious valleys and minima can exist for wide sparse networks.
AdaLoss optimizes adaptive learning rates for efficient convergence in various models.
problem Efficiently optimizing adaptive learning rates for gradient descent methods.
method AdaLoss uses loss function information to dynamically adjust step sizes.
result AdaLoss achieves linear convergence in linear regression and robust global convergence in neural networks.
New framework finds more efficient linear layers over structured matrices.
problem Efficient alternatives for dense linear layers in neural networks.
method Unified framework searching over all linear operators, developing a taxonomy based on computational and algebraic properties.
result BTT-MoE provides substantial compute-efficiency gains over dense layers and standard MoE.
We prove that for an L-layer fully-connected linear neural network, if the width of every hidden layer is Ω~(L⋅r⋅dout⋅κ3), where r and κ are the rank and the condition number of the input data, and dout is the output dimension, then gradient descent with Gaussi…
New metric shows how different regularization methods affect deep linear networks.
problem Understanding the training dynamics of deep linear networks.
method Introduced a new metric called layer imbalance to analyze training dynamics. Demonstrated behavior of different regularization methods and stochastic gradient descent.
result Different regularization methods behave similarly, leading to a flat minima.
This work presents a novel approach to train invertible linear layers by adding rank-one perturbations.
problem Training invertible linear layers during optimization with gradient-based methods is challenging.
method Train rank-one perturbations and add them to weight matrices infrequently, keeping track of inverses and determinants.
result Invertible linear layers improve mixing and mode separation in normalizing flows.
Spectral analysis shows neural networks separate from linear methods in approximating functions.
problem Separating two-layer neural networks from linear methods in function approximation.
method Spectral-based approach using Kolmogorov width and kernel spectrum.
result Upper and lower bounds on separation, explicit hard functions identified.
Sparse butterfly network replaces dense layers in neural networks, improving expressibility and performance.
problem Improving expressibility and performance of neural networks with dense layers.
method Replacing dense layers with a butterfly network architecture.
result The proposed architecture significantly reduces the number of weights from quadratic to nearly linear, with comparable or better performance.
This work shows linear convergence for two-layer neural networks in mean-field regime.
problem Optimizing two-layer neural networks in the mean-field regime.
method Mean-field analysis and continuous-time noisy gradient descent.
result Establishes linear convergence rate for two-layer neural networks.
A simple 2-layer linear network outperforms neural networks in learning sparse targets.
problem Learning sparse targets from a sparse input with gradient descent.
method A 2-layer linear network with fully connected input layer and sparse targets.
result The 2-layer linear network achieves a lower expected square loss than neural networks.
A neural network with a single hidden layer can't represent certain multivariable functions.
problem Representing certain multivariable functions with a neural network having only one hidden layer.
method Developed a continuum version of a one-hidden-layer neural network with ReLU activation, and proved constraints on its parameters and second derivative.
result Existence of a smooth binary function that cannot be precisely represented by any such neural network.
Data augmentation methods improve worst-case model performance.
problem Ensuring fair predictions across subpopulations in large models.
method Linear last layer retraining with data augmentation techniques.
result Optimal worst-group accuracy achieved for Gaussian latent representation distribution.
Upper bounds on fixed points in PWL neural networks with hyperplane analysis.
problem Analyzing the number of fixed points in neural networks with PWL activation.
method Hyperplane arrangements to bound the number of fixed points.
result Upper bounds on the number of fixed points for PWL networks, showing exponential growth in layers.
This study analyzes how one-layer transformers learn regular language recognition tasks.
problem Understanding how one-layer transformers solve regular language recognition tasks like even pairs and parity check.
method Theoretical analysis of training dynamics and gradient descent for a one-layer transformer.
result A one-layer transformer can solve even pairs directly but needs CoT for parity check. Training phases show rapid growth in attention layer followed by logarithmic growth in linear layer.
Transformers can approximate Newton's method for logistic regression.
problem Implementing higher order optimization methods in Transformers.
method Linear attention Transformers with ReLU layers approximating second order optimization algorithms.
result Transformers can implement a single step of Newton's iteration for matrix inversion.
Neural network models have a reputation for being black boxes. We propose to monitor the features at every layer of a model and measure how suitable they are for classification. We use linear classifiers, which we refer to as "probes", trained entirely independently of the model itself. This helps us better understand …
Adding linear layers to ReLU networks favors functions with low mixed variation.
problem Understanding function space bias in overparameterized neural networks.
method Examined a family of networks with varying depths and same capacity but different representation costs, focusing on the effect of adding linear layers to the input side.
result Adding linear layers to shallow ReLU networks results in a bias towards functions with low mixed variation, which can be well approximated by single- or multi-index models.
Neural networks use their hidden layers to transform input data into linearly separable data clusters, with a linear or a perceptron type output layer making the final projection on the line perpendicular to the discriminating hyperplane. For complex data with multimodal distributions this transformation is difficult t…
Batch normalization biases linear models towards uniform margins, improving performance in binary classification.
problem Understanding the implicit bias of batch normalization in linear models and neural networks.
method Analyzing gradient descent convergence on linear models and two-layer CNNs with batch normalization.
result Gradient descent with batch normalization in linear models converges to a uniform margin classifier with an exponential convergence rate.
Paper analyzes why deeper layers of ViTs perform worse on out-of-distribution tasks.
problem Performance degradation of intermediate layers in ViTs under distribution shift.
method Extensive linear probing experiments across various benchmarks and fine-grained analysis of transformer modules.
result Probing feedforward network activations yields best performance under significant distribution shift.
Transformers can learn noisy linear systems with depth and IID data.
problem Learning noisy linear dynamical systems with transformers.
method Theoretical analysis of multi-layer and single-layer transformers with respect to L2-testing loss. result Single-layer transformers have a non-diminishing lower bound on approximation error, suggesting depth separation.
LIFE framework improves model accuracy and interpretability.
problem Achieving high prediction accuracy and interpretability in neural networks.
method Three-step process: subset definition, feature creation, and linear model combination.
result LIFE consistently outperforms other models in prediction accuracy and interpretability.
A method to automatically choose feature dimensions in linear attention for better approximation quality.
problem Choosing the feature dimension in linear attention to balance quality and efficiency.
method Statistical degrees of freedom for determining feature dimension, layer-wise training strategy.
result Our method achieves smaller approximation error compared to fixed dimensions and improves model performance.
A new optimizer for deep learning improves accuracy and reduces training time.
problem Training deep neural networks for classification tasks.
method Hybrid Newton/Gradient Descent (NGD) method exploiting convexity of cross-entropy loss.
result Improves validation error and provides qualitative differences in hidden layer basis functions.
Gradient descent trains both layers of a ReLU network to fit a linear model.
problem Training dynamics of a ReLU network to fit a linear target function.
method Jointly training both layers of a one-hidden-layer ReLU network in a realizable setting with Gaussian inputs and labels.
result Gradient descent from a small random initialization converges to a global minimizer at a linear rate with optimal sample complexity.