One-layer transformers can't solve induction heads task efficiently.
problem Solving the induction heads task efficiently with one-layer transformers.
method Communication complexity argument showing exponential size requirement.
result No one-layer transformer can solve the induction heads task efficiently.
The study characterizes sets for which one-layer neural networks are positive.
problem Characterizing sets of points for which one-layer neural networks are positive.
method Investigation of subsets of the real plane for one-layer ReLU neural networks.
result Full characterization of such sets for cones, and necessary condition for any subset of \(\mathbb{R}^d\).
Gradient descent fails to learn simple neural networks efficiently.
problem Learning one-layer neural networks efficiently using gradient descent.
method Gradient descent and statistical query algorithms.
result Superpolynomial lower bounds for learning one-layer neural networks.
Law derived for neural networks with sparse connections.
problem Understanding the behavior of neural networks with sparse connections.
method Law of large numbers for empirical distribution of parameters derived.
result Law for neural networks with sparse connections derived.
Generative adversarial networks (GANs) are a widely used framework for learning generative models. Wasserstein GANs (WGANs), one of the most successful variants of GANs, require solving a minmax optimization problem to global optimality, but are in practice successfully trained using stochastic gradient descent-ascent.…
Randomly initialized neural networks can linearly separate arbitrary sets.
problem Mapping two arbitrary sets to linearly separable sets.
method Randomly initialized one-layer neural networks with sufficient width.
result With high probability, these networks can transform two sets into linearly separable sets.
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.
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.
It is often hypothesized that a crucial role for recurrent connections in the brain is to constrain the set of possible response patterns, thereby shaping the neural code. This implies the existence of neural codes that cannot arise solely from feedforward processing. We set out to find such codes in the context of one…
We introduce a new method for training deep Boltzmann machines jointly. Prior methods require an initial learning pass that trains the deep Boltzmann machine greedily, one layer at a time, or do not perform well on classifi- cation tasks.
Transformers can learn optimal variable selection in group-sparse classification.
problem Understanding how transformers leverage attention to select relevant variables in group-sparse classification.
method Training a one-layer transformer using gradient descent to select variables from one group of input variables.
result A one-layer transformer can correctly leverage the attention mechanism to select variables, disregarding irrelevant ones.
Layer-wise networks have a closed-form solution and a stopping criterion.
problem Training networks one layer at a time without backpropagation.
method Proved the Kernel Mean Embedding as the closed-form solution and developed a stopping criterion.
result Layer-wise networks converge to a highly desirable kernel for classification.
Layer-wise networks have a closed-form solution and a stopping criterion.
problem Training networks one layer at a time without backpropagation.
method Proved the closed-form solution using the kernel Mean Embedding and Neural Indicator Kernel.
result Layer-wise networks have a closed-form solution and a stopping criterion.
This paper compares Transformers and RNNs in various tasks, showing size differences.
problem Comparing representational capabilities of Transformers and RNNs across tasks.
method Analysis of differences in tasks like index lookup, nearest neighbor, and string equality.
result Size differences in Transformers and RNNs for various tasks.
Study on calibration and consistency of adversarial surrogate losses.
problem Designing robust classifiers with theoretical guarantees.
method Extensive analysis of H-calibration and H-consistency of adversarial surrogate losses.
result Some convex loss functions and supremum-based convex losses are not H-calibrated for important hypothesis sets.
Neural networks are dense among Lipschitz functions with fixed Lipschitz constant.
problem Characterizing neural network approximations to Lipschitz functions.
method Analyzing L-Lipschitz neural networks and their density in L-Lipschitz functions. result One layer neural networks are dense in the set of all L-Lipschitz functions. An efficient way to learn deep density models that have many layers of latent variables is to learn one layer at a time using a model that has only one layer of latent variables. After learning each layer, samples from the posterior distributions for that layer are used as training data for learning the next layer. Thi…
Lower bounds set for infinite-precision transformers.
problem Understanding limitations of infinite-precision transformers.
method Used VC dimension technique to prove lower bounds.
result First lower bounds for two tasks: function composition and SUM2. Gradient flow on softmax attention minimizes nuclear norm of weight matrices.
problem Classification with separate key and query weight matrices.
method Gradient flow on exponential loss, separability assumption, reparameterization, approximate KKT conditions.
result Gradient flow implicitly minimizes nuclear norm of weight matrices, contrasting with Frobenius norm minimization.
This paper explores how Transformers predict next tokens in autoregressive tasks.
problem Understanding the success of Transformers in autoregressive learning.
method Trained a Transformer on a next-token prediction task, focusing on commuting orthogonal matrices.
result Trained Transformers can be seen as implementing gradient descent for a specific objective function.
One LN layer stabilizes neural network extrapolation.
problem Understanding neural networks' extrapolation behavior.
method Applied NTK theory to infinitely-wide networks.
result Adding one LN layer stabilizes network outputs.
In this work we propose a new deep learning tool called deep dictionary learning. Multi-level dictionaries are learnt in a greedy fashion, one layer at a time. This requires solving a simple (shallow) dictionary learning problem, the solution to this is well known. We apply the proposed technique on some benchmark deep…
Deep networks without non-linearities are equivalent to shallow ones.
problem Training deep orthogonal linear networks with no non-linearity.
method Riemannian gradient descent and gradient descent on factorization.
result Training deep overparametrized networks is equivalent to shallow ones.
This paper investigates how transformers can learn to generalize to unseen examples in context.
problem Understanding how transformers can generalize to unseen examples in a prompt.
method Gradient descent analysis of one-layer multi-head transformers for in-context learning.
result The training loss for a one-layer multi-head transformer converges linearly to a global minimum, effectively learning ridge regression over basis functions.
A generative Bayesian model is developed for deep (multi-layer) convolutional dictionary learning. A novel probabilistic pooling operation is integrated into the deep model, yielding efficient bottom-up and top-down probabilistic learning. After learning the deep convolutional dictionary, testing is implemented via dec…
Inspired by coarse-graining approaches used in physics, we show how similar algorithms can be adapted for data. The resulting algorithms are based on layered tree tensor networks and scale linearly with both the dimension of the input and the training set size. Computing most of the layers with an unsupervised algorith…
We describe \textit{deep exponential families} (DEFs), a class of latent variable models that are inspired by the hidden structures used in deep neural networks. DEFs capture a hierarchy of dependencies between latent variables, and are easily generalized to many settings through exponential families. We perform infere…
We consider the problem of estimating the parameters of a d-dimensional rectified Gaussian distribution from i.i.d. samples. A rectified Gaussian distribution is defined by passing a standard Gaussian distribution through a one-layer ReLU neural network. We give a simple algorithm to estimate the parameters (i.e., th…
Paper quantifies MEV on L2 networks, finding significant amounts on Polygon.
problem Lack of research on quantifying MEV on Ethereum Layer 2 networks.
method Analysis of Polygon's MEV, focusing on arbitrage opportunities and liquidations.
result Substantial MEV ($213 million) on L2s, mostly from arbitrage opportunities.
The R package abn is designed to fit additive Bayesian models to observational datasets. It contains routines to score Bayesian networks based on Bayesian or information theoretic formulations of generalized linear models. It is equipped with exact search and greedy search algorithms to select the best network. It supp…
While the optimization problem behind deep neural networks is highly non-convex, it is frequently observed in practice that training deep networks seems possible without getting stuck in suboptimal points. It has been argued that this is the case as all local minima are close to being globally optimal. We show that thi…
Study explores how neural networks and Transformers learn modular arithmetic with multiple inputs.
problem Understanding how neural networks and Transformers learn modular arithmetic with multiple inputs.
method Analytical characterization of features learned by neural networks and Transformers, focusing on margin maximization and Fourier spectra.
result Neural networks and Transformers require a minimum neuron count of \( m \geq 2^{2k-2} \cdot (p-1) \) to solve modular addition problems with \( k \) inputs and modulus \( p \).
Study on feature learning dynamics in infinite-depth neural networks, focusing on ResNets.
problem Understanding how features evolve during training in deep neural networks, especially in the large-depth limit.
method Conditional Gaussian representations and SDE system with decoupled backward weights.
result Depth-induced suppression of forward-backward coupling in infinite-depth networks, leading to a decoupled forward-backward SDE system.
We introduce a new method for training deep Boltzmann machines jointly. Prior methods of training DBMs require an initial learning pass that trains the model greedily, one layer at a time, or do not perform well on classification tasks. In our approach, we train all layers of the DBM simultaneously, using a novel train…
Transformers excel at sparse token selection, surpassing FCNs in both worst and average cases.
problem Sparse token selection task
method One-layer transformer trained with gradient descent
result Transformers learn sparse token selection and exhibit strong out-of-distribution length generalization
ELF simplifies normalizing flows, making them more efficient and universal.
problem Computational inefficiency of normalizing flows.
method ELF introduces a simple, one-layer network with closed-form Lipschitz constants, combining the ease of residual flows with the performance of autoregressive flows.
result ELF is a provably universal density approximator, more efficient computationally and parameter-wise.
Improved bounds on neural network regions using activation histograms.
problem Bounding the number of affine regions in ReLU networks.
method Analysis of algebraic topology problem, extension of framework to subnetwork composition.
result Slightly tighter bounds and insights into parameter initialization.
Deep neural networks (DNNs) may outperform human brains in complex tasks, but the lack of transparency in their decision-making processes makes us question whether we could fully trust DNNs with high stakes problems. As DNNs' operations rely on a massive number of both parallel and sequential linear/nonlinear computati…
RNNs solve modular addition tasks using low rank and sparse Fourier structures.
problem Solving modular addition tasks with recurrent neural networks.
method Identified low rank structures and sparse Fourier representations in RNN weights.
result RNNs robust to removing individual frequencies but degrade with more ablation.
An emerging problem in trustworthy machine learning is to train models that produce robust interpretations for their predictions. We take a step towards solving this problem through the lens of axiomatic attribution of neural networks. Our theory is grounded in the recent work, Integrated Gradients (IG), in axiomatical…
Unified framework LPCD optimizes quantization of complex submodules.
problem Quantization of complex submodules in neural networks.
method Layer-Projected Coordinate Descent (LPCD) for quantizing arbitrary submodules.
result LPCD enhances both layer-wise PTQ methods and existing submodule approaches.
Analytical solution found for a three-layer network with a specific activation function.
problem Understanding the power of depth in neural networks.
method Found analytical solutions for a three-layer network with a matrix exponential activation function.
result Analytical solutions for equations involving a three-layer network with a matrix exponential activation function.
Overparameterized Neural Networks (NN) display state-of-the-art performance. However, there is a growing need for smaller, energy-efficient, neural networks tobe able to use machine learning applications on devices with limited computational resources. A popular approach consists of using pruning techniques. While thes…
LLMs can be tricked into recalling facts based on context clues.
problem Manipulation of LLMs' factual recall through context changes.
method Mathematical exploration of transformers' associative memory properties.
result Transformers use self-attention and value matrix for associative memory.
We consider the problem of training generative models with a Generative Adversarial Network (GAN). Although GANs can accurately model complex distributions, they are known to be difficult to train due to instabilities caused by a difficult minimax optimization problem. In this paper, we view the problem of training GAN…
Neural networks with low-precision weights and activations offer compelling efficiency advantages over their full-precision equivalents. The two most frequently discussed benefits of quantization are reduced memory consumption, and a faster forward pass when implemented with efficient bitwise operations. We propose a t…
Improved texture synthesis using wavelet-based statistics with rectifier non-linearity.
problem Improving texture synthesis quality using wavelet representations.
method Proposes a family of statistics based on non-linear wavelet representations with a generalized rectifier non-linearity.
result Significantly improves visual quality of texture synthesis compared to classical wavelet-based models.
Softmax attention approximates complex functions and subsumes many known universal approximators.
problem Universal approximation of continuous sequence-to-sequence functions.
method Interpolation-based analysis of attention's internal mechanism, showing its ability to approximate ReLU functions.
result Softmax attention is a universal approximator for continuous sequence-to-sequence functions.