New algorithm trains ReLU gates provably in linear time.
problem Training ReLU gates in realizable settings with mild conditions.
method Iterative stochastic algorithm with moment assumptions.
result First recovery of true labels under data-poisoning attacks.
Deep gated networks help understand training and generalization in deep learning.
problem Understanding the role of SGD in training and generalization of deep neural networks with ReLU activation.
method Developed deep gated networks (DGNs) as a framework to analyze training and generalization in DNNs with ReLU activation.
result Gate adaptation is key for generalization in deep neural networks.
The paper analyzes the role of ReLU gates in deep learning networks.
problem Understanding the role of gates in deep learning networks.
method Developed neural path features (NPF) and neural path values (NPV) to characterize the active sub-networks during training.
result The neural path kernel associated with NPFs is a fundamental quantity that characterizes the information stored in the gates of a DNN.
For any positive integer k, there exist neural networks with Θ(k3) layers, Θ(1) nodes per layer, and Θ(1) distinct parameters which can not be approximated by networks with O(k) layers unless they are exponentially large --- they must possess Ω(2k) nodes. This result is proved here for a class o…
Study reveals heavy-tailed behavior in training ReLU gates.
problem Understanding heavy-tailed distribution in stochastic deep learning.
method Experimental study of heavy-tail index for S.G.D. and a variant.
result Two algorithms exhibit similar heavy-tail behavior on ReLU data.
ReLU neural-networks have been in the focus of many recent theoretical works, trying to explain their empirical success. Nonetheless, there is still a gap between current theoretical results and empirical observations, even in the case of shallow (one hidden-layer) networks. For example, in the task of memorizing a ran…
This work connects neural network training to convex optimization via the NTK.
problem Understanding and optimizing neural network training via convex programs.
method Interpreting gated ReLU network as MKL, showing NTK equivalence, and improving weights.
result The NTK cannot perform better than the optimal MKL kernel on the training set.
This work interprets GELU and related activations via a first-order loss function.
problem Understanding and optimizing activation functions in neural networks.
method Complementary interpretation using the Gaussian first-order loss function.
result Calibrated or learned uniform-threshold gates are competitive and often outperform GELU, ReLU, and SiLU/Swish.
Motivated by the resurgence of neural networks in being able to solve complex learning tasks we undertake a study of high depth networks using ReLU gates which implement the function x↦max{0,x}. We try to understand the role of depth in such neural networks by showing size lowerbounds against such network …
New method converts ANN gates to SNNs with AMOS neurons for improved image classification.
problem Efficiently converting ANN gates to SNNs for neuromorphic hardware.
method Introducing AMOS conversion for gates in ANNs, improving accuracy and throughput.
result Improved accuracy of SNNs for ImageNet from 74.60% to 80.97%.
We apply neural nets with ReLU gates in online reinforcement learning. Our goal is to train these networks in an incremental manner, without the computationally expensive experience replay. By studying how individual neural nodes behave in online training, we recognize that the global nature of ReLU gates can cause und…
GLU variants improve Transformer performance.
problem Improving the performance of Transformer models.
method Tested variations of Gated Linear Units (GLU) in Transformer feed-forward sublayers.
result Some GLU variants yield better performance than ReLU or GELU activations.
Sigmoid gating is more sample efficient than softmax in mixture of experts.
problem Softmax gating leads to unnecessary competition among experts, causing representation collapse.
method Theoretical analysis of a regression framework with mixture of experts, identifying identifiability conditions and convergence rates.
result Sigmoid gating requires fewer samples to achieve the same expert estimation error as softmax gating.
Study reveals biases in gradient descent for GLNs, improving neural network performance.
problem Understanding and improving the inductive biases of deep neural networks.
method Derive infinite-time training limit of gated linear networks and generalize to other networks.
result Theoretical framework captures key inductive biases of ReLU networks.
This paper presents a new family of backpropagation-free neural architectures, Gated Linear Networks (GLNs). What distinguishes GLNs from contemporary neural networks is the distributed and local nature of their credit assignment mechanism; each neuron directly predicts the target, forgoing the ability to learn feature…
The study calculates the injectivity capacity of ReLU networks using a novel mathematical approach.
problem Determining the injectivity capacity of ReLU networks layers.
method Employing fully lifted random duality theory (fl RDT) to handle the ℓ0 spherical perceptron and implicitly the ReLU layers injectivity. result The lifting mechanism converges remarkably fast with relative corrections not exceeding 0.1%.
Token-adaptive FFN design improves LLM expressivity.
problem Fixed activation functions limit FFN expressivity in LLMs.
method Mixture of Activations (MoA) and learnable activations (LA).
result MoA achieves lower terminal loss and better scaling than baselines.
Efficiently approximates neural network function space distance.
problem Estimating the average discrepancy between neural network outputs.
method Linearized Activation Function TRick (LAFTR) for ReLU networks.
result Parametric approximation outperforms nonparametric methods in memory and accuracy.
We propose a simple extension to the ReLU-family of activation functions that allows them to shift the mean activation across a layer towards zero. Combined with proper weight initialization, this alleviates the need for normalization layers. We explore the training of deep vanilla recurrent neural networks (RNNs) with…
In "Dictionary Learning" one tries to recover incoherent matrices A∗∈Rn×h (typically overcomplete and whose columns are assumed to be normalized) and sparse vectors x∗∈Rh with a small support of size hp for some 0<p<1 while having access to observations $y \in \mathbb{…
The paper studies 4-qubit Clifford states and their properties.
problem Understanding the set and properties of 4-qubit Clifford states.
method Analyzing the 293760 4-qubit Clifford states, splitting them into 18 groups, and studying the action of CNOT gates and local gates.
result There are 293760 4-qubit Clifford states with specific entanglement entropies, and any pair can be connected with local gates and at most 3 CNOT gates.
Improved logistic MoE with sigmoid gate shows better sample efficiency.
problem Improving sample efficiency in logistic MoE models.
method Comprehensive analysis of multinomial logistic MoE with modified sigmoid gate, incorporating temperature parameter and using Euclidean score.
result The sigmoid gate leads to lower sample complexity than softmax gate for both parameter and expert estimation.
New method prepares 3-qubit states using local gates and controlled-Z gates.
problem Preparation of 3-qubit states using quantum gates.
method Uses Ry(θ) gates and controlled-Z gates, with an optimal number of controlled-Z gates. result Optimal number of controlled-Z gates for preparing 3-qubit states is four. In quantum computation, series of quantum gates have to be arranged in a predefined sequence that led to a quantum circuit in order to solve a particular problem. What if the sequence of quantum gates is known but both the problem to be solved and the outcome of the so defined quantum circuit remain in the shadow? This…
Long Short-Term Memory (LSTM) infers the long term dependency through a cell state maintained by the input and the forget gate structures, which models a gate output as a value in [0,1] through a sigmoid function. However, due to the graduality of the sigmoid function, the sigmoid gate is not flexible in representing m…
Gated attention improves model curvature, enhancing performance on nonlinear tasks.
problem Understanding the geometric implications of gating in attention mechanisms.
method Modeling attention outputs as Gaussian distributions and analyzing Fisher--Rao geometry.
result Gated attention enables non-flat geometries, including positively curved manifolds.
Gated attention improves performance by using a hierarchical mixture of experts.
problem Improving performance of self-attention mechanisms in Transformers.
method Rigorously show that gated attention can be modeled as a hierarchical mixture of experts, providing a theoretical justification for its benefits.
result Gated attention is more sample-efficient than multi-head self-attention, requiring fewer data points to achieve the same estimation error.
Given the success of the gated recurrent unit, a natural question is whether all the gates of the long short-term memory (LSTM) network are necessary. Previous research has shown that the forget gate is one of the most important gates in the LSTM. Here we show that a forget-gate-only version of the LSTM with chrono-ini…
Improved HMoE models using Laplace gating function enhance expert specialization and performance.
problem Improving performance of hierarchical mixture of experts models.
method Used Laplace gating function instead of Softmax in hierarchical mixture of experts models.
result Laplace gating function accelerates expert convergence and enhances specialization.
Gating units in GRUs and LSTMs create slow modes and control phase-space complexity.
problem Training challenges in RNNs due to exploding or vanishing gradients.
method Random matrix theory and mean-field theory applied to GRUs and LSTMs.
result Gates in GRUs and LSTMs lead to accumulation of slow modes and control phase-space complexity.
The paper evaluates three variants of the Gated Recurrent Unit (GRU) in recurrent neural networks (RNN) by reducing parameters in the update and reset gates. We evaluate the three variant GRU models on MNIST and IMDB datasets and show that these GRU-RNN variant models perform as well as the original GRU RNN model while…
This study analyzes communication constraints in MoE architectures using information theory.
problem Communication constraints in Mixture-of-Experts (MoE) architectures.
method Developed a rate-distortion characterization of finite-rate gating in MoE architectures using information theory.
result Yielded capacity-aware limits for communication-constrained MoE systems.
Bayesian methods have been successfully applied to sparsify weights of neural networks and to remove structure units from the networks, e. g. neurons. We apply and further develop this approach for gated recurrent architectures. Specifically, in addition to sparsification of individual weights and neurons, we propose t…
GLCB uses Gated Linear Networks for online contextual bandits.
problem Online learning in contextual bandits with uncertainty estimation.
method Gated Linear Networks (GLNs) for prediction and uncertainty estimation.
result GLCB outperforms state-of-the-art methods in online contextual bandits.
Paper connects MoE and self-attention, proposing active-attention.
problem Improving efficiency and performance of self-attention mechanisms.
method Established connection between MoE and self-attention, analyzed quadratic gating functions, proposed active-attention mechanism.
result Active-attention outperforms standard self-attention in various tasks.
Gated recurrent neural networks have achieved remarkable results in the analysis of sequential data. Inside these networks, gates are used to control the flow of information, allowing to model even very long-term dependencies in the data. In this paper, we investigate whether the original gate equation (a linear projec…
Solves challenges in estimating parameters of softmax gating Gaussian mixture models.
problem Identifiability issues and complex interactions in Gaussian mixture of experts.
method Proposes novel Voronoi loss functions and establishes convergence rates of MLE.
result Connects convergence rate of MLE to a solvability problem of polynomial equations.
A major contributing factor to the recent advances in deep neural networks is structural units that let sensory information and gradients to propagate easily. Gating is one such structure that acts as a flow control. Gates are employed in many recent state-of-the-art recurrent models such as LSTM and GRU, and feedforwa…
This study connects ReLU neural networks to toric geometry to analyze function realization.
problem Determining which continuous piecewise linear functions can be realized by ReLU neural networks.
method Established a connection between toric geometry and ReLU neural networks, defining key structures like the ReLU fan, toric variety, and Cartier divisor.
result Proved a criterion for functions realizable by unbiased shallow ReLU networks using intersection numbers.
Recently, a lot of techniques were developed to sparsify the weights of neural networks and to remove networks' structure units, e.g. neurons. We adjust the existing sparsification approaches to the gated recurrent architectures. Specifically, in addition to the sparsification of weights and neurons, we propose sparsif…
Gradient Gating improves deep GNNs by modulating message passing updates.
problem Oversmoothing and performance degradation in deep GNNs.
method Gradient gating mechanism for multi-rate message passing.
result G2 framework alleviates oversmoothing and achieves state-of-the-art performance. LSTMs were introduced to combat vanishing gradients in simple RNNs by augmenting them with gated additive recurrent connections. We present an alternative view to explain the success of LSTMs: the gates themselves are versatile recurrent models that provide more representational power than previously appreciated. We do…
The paper establishes convergence rates for MoE models in classification problems.
problem Understanding the behavior of MoE models in classification settings.
method Established convergence rates for density and parameter estimation in softmax gating multinomial logistic MoE models.
result Parameter estimation rates are significantly improved with a novel modified softmax gating function.
Spatiotemporal sequence prediction is an important problem in deep learning. We study next-frame(s) video prediction using a deep-learning-based predictive coding framework that uses convolutional, long short-term memory (convLSTM) modules. We introduce a novel reduced-gate convolutional LSTM(rgcLSTM) architecture that…
New insights into the top-K sparse softmax gating function for deep learning.
problem Understanding the theoretical effects of the top-K sparse softmax gating function on density and parameter estimations.
method Using a Gaussian mixture of experts, novel loss functions, and theoretical analysis.
result The convergence rates of density and parameter estimations are parametric under certain conditions, but slow under over-specified models.
Model accurately gates ocean microbes from high-frequency flow cytometry data.
problem Gating of high-frequency flow cytometry data for ocean microbes is challenging.
method Trend filtered mixture of experts with smooth parameter variation.
result Model accurately matches human-annotated gating and corrects errors.
In order to quantize the gate parameters of the LSTM (Long Short-Term Memory) neural network model with almost no recognition performance degraded, a new quantization method named Quantization Loss Re-Learn Method is proposed in this paper. The method does lossy quantization on gate parameters during training iteration…
A new method for differentiable structured sparsity improves neural network performance and sparsity.
problem Non-differentiability of structured sparsity penalties in neural networks.
method Introducing D-Gating, a differentiable approach to structured overparameterization. result The D-Gating objective converges to the L2,2/D-regularized loss and induces sparse learning dynamics.