We formally study how ensemble of deep learning models can improve test accuracy, and how the superior performance of ensemble can be distilled into a single model using knowledge distillation. We consider the challenging case where the ensemble is simply an average of the outputs of a few independently trained neural networks with the SAME architecture, trained using the SAME algorithm on the SAME data set, and they only differ by the random seeds used in the initialization. 
We empirically show that ensemble/knowledge distillation in deep learning works very differently from traditional learning theory, especially differently from ensemble of random feature mappings or the neural-tangent-kernel feature mappings, and is potentially out of the scope of existing theorems. Thus, to properly understand ensemble and knowledge distillation in deep learning, we develop a theory showing that when data has a structure we refer to as "multi-view", then ensemble of independently trained neural networks can provably improve test accuracy, and such superior test accuracy can also be provably distilled into a single model by training a single model to match the output of the ensemble instead of the true label. Our result sheds light on how ensemble works in deep learning in a way that is completely different from traditional theorems, and how the "dark knowledge" is hidden in the outputs of the ensemble -- that can be used in knowledge distillation -- comparing to the true data labels. In the end, we prove that self-distillation can also be viewed as implicitly combining ensemble and knowledge distillation to improve test accuracy.

Towards Understanding Ensemble, Knowledge Distillation and Self-Distillation in Deep Learning.

How can neural networks such as ResNet efficiently learn CIFAR-10 with test accuracy more than 96%, while other methods, especially kernel methods, fall relatively behind? Can we more provide theoretical justifications for this gap? 
Recently, there is an influential line of work relating neural networks to kernels in the over-parameterized regime, proving they can learn certain concept class that is also learnable by kernels with similar test error. Yet, can neural networks provably learn some concept class BETTER than kernels? 
We answer this positively in the distribution-free setting. We prove neural networks can efficiently learn a notable class of functions, including those defined by three-layer residual networks with smooth activations, without any distributional assumption. At the same time, we prove there are simple functions in this class such that with the same number of training examples, the test error obtained by neural networks can be MUCH SMALLER than ANY kernel method, including neural tangent kernels (NTK). 
The main intuition is that multi-layer neural networks can implicitly perform hierarchical learning using different layers, which reduces the sample complexity comparing to "one-shot" learning algorithms such as kernel methods. In a follow-up work [2], this theory of hierarchical learning is further strengthened to incorporate the "backward feature correction" process when training deep networks. 
In the end, we also prove a computation complexity advantage of ResNet with respect to other learning methods including linear regression over arbitrary feature mappings.

What Can ResNet Learn Efficiently, Going Beyond Kernels?

How does a 110-layer ResNet learn a high-complexity classifier using relatively few training examples and short training time? We present a theory towards explaining this in terms of hierarchical learning. We refer hierarchical learning as the learner learns to represent a complicated target function by decomposing it into a sequence of simpler functions to reduce sample and time complexity. This paper formally analyzes how multi-layer neural networks can perform such hierarchical learning efficiently and automatically by applying SGD. 
On the conceptual side, we present, to the best of our knowledge, the FIRST theory result indicating how deep neural networks can be sample and time efficient on certain hierarchical learning tasks, when NO KNOWN non-hierarchical algorithms (such as kernel method, linear regression over feature mappings, tensor decomposition, sparse coding, and their simple combinations) are efficient. We establish a principle called "backward feature correction", where training higher layers in the network can improve the features of lower level ones. We believe this is the key to understand the deep learning process in multi-layer neural networks. 
On the technical side, we show for every input dimension $d > 0$, there is a concept class consisting of degree $\omega(1)$ multi-variate polynomials so that, using $\omega(1)$-layer neural networks as learners, SGD can learn any target function from this class in $\mathsf{poly}(d)$ time using $\mathsf{poly}(d)$ samples to any $\frac{1}{\mathsf{poly}(d)}$ error, through learning to represent it as a composition of $\omega(1)$ layers of quadratic functions. In contrast, we present lower bounds stating that several non-hierarchical learners, including any kernel methods, neural tangent kernels, must suffer from $d^{\omega(1)}$ sample or time complexity to learn this concept class even to $d^{-0.01}$ error.

Backward Feature Correction: How Deep Learning Performs Deep Learning

Despite the empirical success of using Adversarial Training to defend deep learning models against adversarial perturbations, so far, it still remains rather unclear what the principles are behind the existence of adversarial perturbations, and what adversarial training does to the neural network to remove them. In this paper, we present a principle that we call Feature Purification, where we show one of the causes of the existence of adversarial examples is the accumulation of certain small dense mixtures in the hidden weights during the training process of a neural network; and more importantly, one of the goals of adversarial training is to remove such mixtures to purify hidden weights. We present both experiments on the CIFAR-10 dataset to illustrate this principle, and a theoretical result proving that for certain natural classification tasks, training a two-layer neural network with ReLU activation using randomly initialized gradient descent indeed satisfies this principle. Technically, we give, to the best of our knowledge, the first result proving that the following two can hold simultaneously for training a neural network with ReLU activation. (1) Training over the original data is indeed non-robust to small adversarial perturbations of some radius. (2) Adversarial training, even with an empirical perturbation algorithm such as FGM, can in fact be provably robust against ANY perturbations of the same radius. Finally, we also prove a complexity lower bound, showing that low complexity models such as linear classifiers, low-degree polynomials, or even the neural tangent kernel for this network, CANNOT defend against perturbations of this same radius, no matter what algorithms are used to train them. 

/pdf/feature-purification-how-adversarial-training-performs-veskfga7.pdf

Feature Purification: How Adversarial Training Performs Robust Deep Learning

One of the trends in information technologies is implementing neural networks in modern software packages [1]. The fact that neural networks cannot be directly programmed (but trained) is their distinctive feature. In this regard, the urgent task is to ensure sufficient speed and quality of neural network training procedures. The process of neural network training can differ significantly depending on the problem. There are verification methods that correspond to the task’s constraints; they are used to assess the training results. Verification methods provide an estimate of the entire cardinal set of examples but do not allow to estimate which subset of those causes a significant error. This fact leads to neural networks’ failure to perform with the given set of hyperparameters, making training a new one time-consuming. On the other hand, existing empirical assessment methods of neural networks training use discrete sets of examples. With this approach, it is impossible to say that the network is suitable for classification on the whole cardinal set of examples. This paper proposes a criterion for assessing the quality of classification results. The criterion is formed by describing the training states of the neural network. Each state is specified by the correspondence of the set of errors to the function range representing a cardinal set of test examples. The criterion usage allows tracking the network’s classification defects and marking them as safe or unsafe. As a result, it is possible to formally assess how the training and validation data sets must be altered to improve the network’s performance, while existing verification methods do not provide any information on which part of the dataset causes the network to underperform.

/pdf/artificial-neural-network-training-criterion-formulation-10m8mazkff.pdf

Artificial Neural Network Training Criterion Formulation Using Error Continuous Domain

Word embedding is a distributed representation of words in a vector space. It involves a mathematical embedding from a space with one dimension per word to a continuous vector space with much lower dimension. It performs well on tasks including synonym and hyponym detection by grouping similar words. However, most existing word embeddings are insensitive to antonyms, since they are trained based on word distributions in a large amount of text data, where antonyms usually have similar contexts. To generate word embeddings that are capable of detecting antonyms, we firstly modify the objective function of Skip-Gram model, and then utilize the supervised synonym and antonym information in thesauri as well as the sentiment information of each word in SentiWordNet. We conduct evaluations on three relevant tasks, namely GRE antonym detection, word similarity, and semantic textual similarity. The experiment results show that our antonym-sensitive embedding outperforms common word embeddings in these tasks, demonstrating the efficacy of our methods.

Improving Word Embeddings for Antonym Detection Using Thesauri and SentiWordNet

In this paper, we analyze efficacy of the fast gradient sign method (FGSM) and the Carlini-Wagner's L2 (CW-L2) attack. We prove that, within a certain regime, the untargeted FGSM can fool any convolutional neural nets (CNNs) with ReLU activation; the targeted FGSM can mislead any CNNs with ReLU activation to classify any given image into any prescribed class. For a special two-layer neural network: a linear layer followed by the softmax output activation, we show that the CW-L2 attack increases the ratio of the classification probability between the target and ground truth classes. Moreover, we provide numerical results to verify all our theoretical results.

pdf/mathematical-analysis-of-adversarial-attacks-3elqztfrf1.pdf

Mathematical Analysis of Adversarial Attacks

Generative Adversarial Networks (GANs) are widely used models to learn complex real-world distributions. In GANs, the training of the generator usually stops when the discriminator can no longer distinguish the generator's output from the set of training examples. A central question of GANs is that when the training stops, whether the generated distribution is actually close to the target distribution. Previously, it was found that such closeness can only be achieved when there is a strict capacity trade-off between the generator and discriminator: Neither of the two models can be too powerful than the other. In this paper, we established one of the first theoretical results in explaining this trade-off. We show that when the generator is a class of two-layer neural networks, then it is necessary and sufficient for the discriminator to be a one-layer network with ReLU-type activation functions. With this trade-off, using polynomially many training examples, when the training stops, the generator will indeed output a distribution that is inverse-polynomially close to the target. Our result also sheds light on how GANs training can find such a generator efficiently.

When can Wasserstein GANs minimize Wasserstein Distance

Generative Adversarial Networks (GANs) are widely used models to learn complex real-world distributions. In GANs, the training of the generator usually stops when the discriminator can no longer distinguish the generator's output from the set of training examples. A central question of GANs is that when the training stops, whether the generated distribution is actually close to the target distribution, and how the training process reaches to such configurations efficiently? In this paper, we established a theoretical results towards understanding this generator-discriminator training process. We empirically observe that during the earlier stage of the GANs training, the discriminator is trying to force the generator to match the low degree moments between the generator's output and the target distribution. Moreover, only by matching these empirical moments over polynomially many training examples, we prove that the generator can already learn notable class of distributions, including those that can be generated by two-layer neural networks.

Making Method of Moments Great Again? -- How can GANs learn distributions

Nash equilibrium has long been a desired solution concept in multi-player games, especially for those on continuous strategy spaces, which have attracted a rapidly growing amount of interests due to advances in research applications such as the generative adversarial networks. Despite the fact that several deep learning based approaches are designed to obtain pure strategy Nash equilibrium, it is rather luxurious to assume the existence of such an equilibrium. In this paper, we present a new method to approximate mixed strategy Nash equilibria in multi-player continuous games, which always exist and include the pure ones as a special case. We remedy the pure strategy weakness by adopting the pushforward measure technique to represent a mixed strategy in continuous spaces. That allows us to generalize the Gradient-based Nikaido-Isoda (GNI) function to measure the distance between the players' joint strategy profile and a Nash equilibrium. Applying the gradient descent algorithm, our approach is shown to converge to a stationary Nash equilibrium under the convexity assumption on payoff functions, the same popular setting as in previous studies. In numerical experiments, our method consistently and significantly outperforms recent works on approximating Nash equilibrium for quadratic games, general blotto games, and GAMUT games.

/pdf/finding-mixed-strategy-nash-equilibrium-for-continuous-games-4sklxff9su.pdf

Finding Mixed Strategy Nash Equilibrium for Continuous Games through Deep Learning

Handle everyday research tasks with reliable, citation-backed results

Your personal Research Agent to handle research tasks with citation-backed results

Popular Tasks used by Researchers

How can I help with your research?

Meet SciSpace

Get more enhanced response by uploading the PDFs you want me to reference.

No relevant PDFs in your library

SciSpace is the AI research assistant for academics. Run systematic literature reviews on 280M+ papers, and write papers with cited sources. Trusted by 1M+ students, PhDs & researchers.

SciSpace | AI for Research

Analyze PDFs

Code & Manuscripts

Funding & Grants

Literature & Patents

Medical & Clinical Data

Systematic Review

Visualize & Present

Web & Data

Build a Google Scholar-like website for your research.

Build a website

Create charts and images for your research

Create a Chart

Write a paper for submission to a journal

Draft a manuscript

Patent Search

Design eye-catching scientific posters in minutes.

Scientific Poster Generation

Systematic Literature Review

One task is running at the moment. Your messages will be shown right after.

Drag and drop or click here to browse

Loved by <highlight>1 million+</highlight> researchers

Extract a list of specific topics and their sources from unstructured text

Topics

Compare and analyze relevant papers that matches with your search

Papers

Get insights from PDFs and bookmarked papers from your library

My library

Recent searches

Try searching for:

Catch AI-generated content in scholarly and non-scholarly content

{ai} Detector

Ai Writer

Get PDF Summaries, highlighted text explanations 

Chat with PDF

Effortlessly create in-text citations and bibliographies in APA and 2,500 other formats

Citation generator

Get explanations, summaries, and answers on academic papers

Ease up your research workflow with {scispace}'s cohort of exciting AI tools

Elevate your academic writing skills and convey your ideas the way you want

Paraphraser

Explore our range of reading and writing tools

Your file is being prepared and should be ready in a few minutes. If it's a large file, it might take a bit longer. You can close this window, and we'll email you the file when it's done.

You have reached a maximum limit of <strong>{limit}</strong> columns in the table. Remove at least <strong>1</strong> column to add or create another one.

Zehao Dou

Author Tools

Chat about Author