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.
There has recently been much work on the "wide limit" of neural networks, where Bayesian neural networks (BNNs) are shown to converge to a Gaussian process (GP) as all hidden layers are sent to infinite width. However, these results do not apply to architectures that require one or more of the hidden layers to remain n…
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.
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.
The muti-layer information bottleneck (IB) problem, where information is propagated (or successively refined) from layer to layer, is considered. Based on information forwarded by the preceding layer, each stage of the network is required to preserve a certain level of relevance with regards to a specific hidden variab…
ResNet learns to compress information during training.
problem Applying information bottleneck principle to deep learning models.
method Trained PixelCNN++ models to measure mutual information between ResNet hidden layers and input data.
result Compression occurs in both classification and autoencoding training regimes.
New research shows non-bottlenecked autoencoders can outperform bottlenecked ones for anomaly detection.
problem The necessity of a bottleneck in autoencoders for anomaly detection.
method Investigated two ways to remove bottlenecks: overparameterising the latent layer and introducing skip connections. Carried out extensive experiments on various AE types and datasets.
result Non-bottlenecked autoencoders can outperform bottlenecked ones, improving anomaly detection performance.
CNNs use a bottleneck structure to focus on a few frequencies, affecting function representation.
problem Understanding how CNNs focus on specific frequencies in their feature learning.
method Defined Convolution Bottleneck (CBN) structure, measured CBN rank, and analyzed parameter norms.
result Parameter norm scales with depth and CBN rank, and networks with optimal parameters exhibit this structure.
Study on feature learning in Leaky ResNets, explaining bottleneck structure.
problem Understanding feature learning in deep neural networks.
method Lagrangian and Hamiltonian reformulation of representation geodesics.
result Emergence of a bottleneck structure in large effective depth networks.
We introduce the HSIC (Hilbert-Schmidt independence criterion) bottleneck for training deep neural networks. The HSIC bottleneck is an alternative to the conventional cross-entropy loss and backpropagation that has a number of distinct advantages. It mitigates exploding and vanishing gradients, resulting in the ability…
Recent years, many researches attempt to open the black box of deep neural networks and propose a various of theories to understand it. Among them, Information Bottleneck (IB) theory claims that there are two distinct phases consisting of fitting phase and compression phase in the course of training. This statement att…
PCBM improves neural network generalization by partially observing concepts.
problem Decreased generalization performance due to observing all concepts in CBM.
method Developed a theoretical analysis of PCBM's Bayesian generalization error.
result PCBM's generalization error is lower than CBM's due to partial concept observation.
We introduce a new framework for training deep generative models for high-dimensional conditional density estimation. The Bottleneck Conditional Density Estimator (BCDE) is a variant of the conditional variational autoencoder (CVAE) that employs layer(s) of stochastic variables as the bottleneck between the input x a…
LogDet estimator improves entropy estimation in neural networks.
problem Inconsistent observations and diversified interpretation in neural networks.
method Proposes LogDet estimator for reliable entropy approximation.
result LogDet estimator overcomes distributional diversity issues.
Information bottleneck (IB) is a method for extracting information from one random variable X that is relevant for predicting another random variable Y. To do so, IB identifies an intermediate "bottleneck" variable T that has low mutual information I(X;T) and high mutual information I(Y;T). The "IB curve" cha…
A new framework for interpretable models using sparse linear layers.
problem Performance degradation and lower interpretability in concept bottleneck models.
method Contrastive Language Image models and a single sparse linear layer with Bayesian inference.
result Our framework outperforms recent CBM approaches in accuracy and concept sparsity.
Theory explains deep nonlinear networks' plateaus and transitions.
problem Understanding long plateaus and feature acquisition transitions in deep nonlinear networks.
method Derived an exact identity for Frobenius norms, classified activation functions, and reduced matrix flow to a scalar ODE.
result Escape time law τ⋆=Θ(ε−(r−2)) for deep nonlinear networks, where r is the number of bottleneck layers. The Information Plane theory predicts autoencoders do not compress input information.
problem Understanding the training dynamics of hidden layers in autoencoders.
method Derive a theoretical convergence for the Information Plane of autoencoders using a Gram-matrix based mutual information estimator.
result Ideal autoencoders with a large bottleneck layer size do not compress input information, while a small size causes compression only in the encoder layers.
The Softmax function on top of a final linear layer is the de facto method to output probability distributions in neural networks. In many applications such as language models or text generation, this model has to produce distributions over large output vocabularies. Recently, this has been shown to have limited repres…
Recent works investigated the generalization properties in deep neural networks (DNNs) by studying the Information Bottleneck in DNNs. However, the mea- surement of the mutual information (MI) is often inaccurate due to the density estimation. To address this issue, we propose to measure the dependency instead of MI be…
Large neural networks learn low-dimensional representations that balance complexity and regularity.
problem Understanding the tradeoff between low-dimensional representations and complexity in deep neural networks.
method Computed finite depth corrections to reveal a measure of regularity that bounds the pseudo-determinant of the Jacobian.
result Proved the conjectured bottleneck structure in learned features as network depth increases, showing almost all hidden representations are approximately low-dimensional and weight matrices have singular values close to 1.
Deep Gaussian Processes are reinterpreted as deep trigonometric networks for tractable inference.
problem Challenging inference in DGPs due to intractable marginalization in latent function space.
method Viewing DGPs as deep trigonometric networks with Bochner's theorem, and using the wide limit with a bottleneck to translate DGPs into deep trigonometric networks.
result The weight space view yields the same effective covariance functions as obtained in function space, and varying prior distributions over network parameters is equivalent to employing different kernels.
We find the maximum mutual information for neural networks and its key determinants.
problem Understanding the maximum mutual information in neural architectures.
method Derived closed-form expression for maximum mutual information across neural network families.
result Maximum mutual information stems from a generalized formula and is influenced by network width and statistical invariances.
Despite their great success in practical applications, there is still a lack of theoretical and systematic methods to analyze deep neural networks. In this paper, we illustrate an advanced information theoretic methodology to understand the dynamics of learning and the design of autoencoders, a special type of deep lea…
New learning rules from information bottleneck improve deep learning without precise labels.
problem Training deep neural networks with backpropagation is biologically implausible.
method Kernelized information bottleneck principle with 3-factor Hebbian structure.
result The new learning rules perform nearly as well as backpropagation on image classification tasks.
Finet uses FBN for efficient, lightweight neural networks.
problem Building efficient neural networks with limited computational resources.
method Introduces Fine-grained Batch Normalization (FBN) and a novel light-weight network (Finet) that combines FBN with standard convolution.
result Finet achieves state-of-the-art performance on ImageNet classification with reduced computational complexity.
Soft-Radial Projection solves gradient saturation in constrained deep learning.
problem Gradient saturation in deep learning models when integrating hard constraints.
method Introduces Soft-Radial Projection, a differentiable layer that maps predictions onto constraint boundaries without rank-deficient Jacobians.
result Improves convergence and solution quality over state-of-the-art methods.
Adversarial examples in deep ReLU networks with constant depth.
problem Adversarial perturbations lead to large output changes in deep ReLU networks.
method Analyzes the phenomenon in networks with independent Gaussian parameters and constant depth.
result Adversarial examples arise due to functions being close to linear.
BLISS optimizes GNN training by adaptively sampling nodes.
problem High computational costs in training GNNs on large graphs.
method Uses Bandit Layer Importance Sampling to dynamically select nodes.
result Improves GNN performance with reduced computational cost.
New pruning method captures global correlations for efficient neural network inference.
problem Efficiently pruning neural networks for faster inference and reduced memory usage.
method Second-order structured pruning (SOSP-H) with innovative saliency-based approaches.
result SOSP-H scales to large-scale vision tasks and improves accuracy without compromising efficiency.
The study explores the compressive power of Boolean threshold autoencoders, finding that seven layers are necessary but three are not.
problem Understanding the compressive limits of Boolean threshold autoencoders.
method Investigation into the minimum number of layers and nodes required for autoencoders to accurately transform binary vectors.
result There exists a seven-layer autoencoder with a logarithmic middle layer size for any set of distinct vectors, but not a three-layer one.
This paper presents a method to explain how the information of each input variable is gradually discarded during the forward propagation in a deep neural network (DNN), which provides new perspectives to explain DNNs. We define two types of entropy-based metrics, i.e. (1) the discarding of pixel-wise information used i…
Optimizing data movement significantly improves transformer training efficiency.
problem Data movement is a major bottleneck in training transformers.
method Developed a recipe to globally optimize data movement in transformers.
result Achieved up to 1.30x performance improvement over state-of-the-art frameworks.
Tensor decomposition methods are widely used for model compression and fast inference in convolutional neural networks (CNNs). Although many decompositions are conceivable, only CP decomposition and a few others have been applied in practice, and no extensive comparisons have been made between available methods. Previo…
SCBMs model causal effects using low-dimensional bottlenecks.
problem Causal effect estimation in high-dimensional systems.
method Structural causal models with low-dimensional summary statistics.
result SCBMs provide a flexible framework for task-specific dimension reduction.
Softmax is an output activation function for modeling categorical probability distributions in many applications of deep learning. However, a recent study revealed that softmax can be a bottleneck of representational capacity of neural networks in language modeling (the softmax bottleneck). In this paper, we propose an…
Standard neural network architectures are non-linear only by virtue of a simple element-wise activation function, making them both brittle and excessively large. In this paper, we consider methods for making the feed-forward layer more flexible while preserving its basic structure. We develop simple drop-in replacement…
Transfer learning through fine-tuning a pre-trained neural network with an extremely large dataset, such as ImageNet, can significantly accelerate training while the accuracy is frequently bottlenecked by the limited dataset size of the new target task. To solve the problem, some regularization methods, constraining th…
A long standing open problem in the theory of neural networks is the development of quantitative methods to estimate and compare the capabilities of different architectures. Here we define the capacity of an architecture by the binary logarithm of the number of functions it can compute, as the synaptic weights are vari…
New approach reduces model size for Transformer architectures.
problem Large embedding dimensions limit model applicability.
method Identified low-rank bottleneck in multi-head attention.
result Reducing head size to sequence length improves model performance.
We present a method for feature interpretation that makes use of recent advances in autoregressive density estimation models to invert model representations. We train generative inversion models to express a distribution over input features conditioned on intermediate model representations. Insights into the invariance…
This paper quantifies hyperparameter transfer and finds embedding layer learning rate is key.
problem Quantifying optimal hyperparameters for large language models across scales.
method Developed three metrics to quantify hyperparameter transfer and investigated the importance of embedding layer learning rate.
result Maximal Update (μP) parameterization offers high-quality learning rate transfer compared to standard parameterization (SP).
Gradient oversmoothing and expansion hinder deep GNN training, solved with normalization.
problem Gradient oversmoothing and expansion prevent deep GNN training.
method Proposed normalization method to constrain the Lipschitz bound of each layer.
result Residual GNNs with hundreds of layers can be efficiently trained with the proposed normalization.
Study shows bottlenecks improve image segmentation quality.
problem Robust object discovery in real-world images remains challenging.
method Empirical investigation of reconstruction bottlenecks in GENESIS model.
result Reconstruction bottlenecks determine reconstruction and segmentation quality.
Deep ResNets favor low bottleneck rank with proper hyperparameters.
problem Understanding the inductive bias of deep neural networks.
method Computed minimum-norm weights of a deep linear ResNet.
result Deep nonlinear ResNets have an inductive bias towards minimizing bottleneck rank.
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.
CB-APM uses analyst consensus as a bottleneck to interpret stock returns.
problem Tackles the challenge of understanding and predicting stock returns using professional beliefs.
method Embeds analyst consensus as a structural bottleneck, treating it as a sufficient statistic for market information.
result CB-APM portfolios exhibit strong monotonic return gradients and robust across different economic conditions.
HaKAN uses Hahn-KAN blocks to forecast multivariate time series.
problem Long-term time series forecasting challenges with high complexity and spectral bias.
method HaKAN integrates channel independence, patching, and a stack of Hahn-KAN blocks with residual connections. It uses Hahn polynomial-based learnable activation functions.
result HaKAN consistently outperforms state-of-the-art methods on various forecasting benchmarks.