Zeyuan hu (zeyuan)

To protect your privacy, all features that rely on external API calls from your browser are turned off by default. You need to opt-in for them to become active. All settings here will be stored as cookies with your web browser. For more information see our F. Add open access links from to the list of external document links if available.

Privacy notice: By enabling the option above, your browser will contact the API of unpaywall. Although we do not have any reason to believe that your call will be tracked, we do not have any control over how the remote server uses your data.

So please proceed with care and consider checking the Unpaywall privacy policy. For web page which are no longer available, try to retrieve content from the of the Internet Archive if available. Privacy notice: By enabling the option above, your browser will contact the API of web. So please proceed with care and consider checking the Internet Archive privacy policy. Add a list of references from and to record detail pages. Privacy notice: By enabling the option above, your browser will contact the APIs of crossref.

So please proceed with care and consider checking the Crossref privacy policy and the OpenCitations privacy policy. Add a list of citing articles from to record detail pages. Privacy notice: By enabling the option above, your browser will contact the API of opencitations. So please proceed with care and consider checking the OpenCitations privacy policy. Show tweets from on the dblp homepage. Privacy notice: By enabling the option above, your browser will contact twitter. At the same time, Twitter will persitently store several cookies with your web browser.

While we did signal Twitter to not track our users by setting the "dnt" flagwe do not have any control over how Twitter uses your data. So please proceed with care and consider checking the Twitter privacy policy. Authors: no matches. Venues: no matches. Publications: no matches. Trier 1 Trier 2. Unpaywalled article links Add open access links from to the list of external document links if available. Archived links via Wayback Machine For web page which are no longer available, try to retrieve content from the of the Internet Archive if available.

Reference lists Add a list of references from and to record detail pages. Citation data Add a list of citing articles from to record detail pages. Tweets on dblp homepage Show tweets from on the dblp homepage.Skip to search form Skip to main content You are currently offline. Some features of the site may not work correctly. Zeyuan Hu. Follow Author Publications Citations Highly Influential Citations 1. Publications Influence. Claim Your Author Page. Ensure your research is discoverable on Semantic Scholar.

Claiming your author page allows you to personalize the information displayed and manage publications all current information on this profile has been aggregated automatically from publisher and metadata sources. Answering visual questions need acquire daily common knowledge and model the semantic connection among different parts in images, which is too difficult for VQA systems to learn from images with the … Continue Reading.

Visual question answering VQA and image captioning require a shared body of general knowledge connecting language and vision. AbstractRegional sea surface temperature SST mode variabilities, especially the La Nina—like Pacific Ocean temperature pattern known as the negative phase of the interdecadal Pacific oscillation … Continue Reading. The goal of our research is to contribute information about how useful the crowd is at anticipating stereotypes that may be biasing a data set without a researcher's knowledge.

The results of the … Continue Reading. In recent years, deep Convolutional neural networks CNNs have made fantastic progress in static image recognition, but the ability to model motion information on behavioral video is weak. Therefore, … Continue Reading.

Log-structured Merge Tree LSM [17] is a data structure that is widely used in write-intensive storage system. However, it suffers from write amplifications, which can hinder the write throughput.

Source code can be treated similar as corpus constructed by natural language Hindle et al. In this paper, we use the neural network model to study identifer naming convention problem. We … Continue Reading. By clicking accept or continuing to use the site, you agree to the terms outlined in our Privacy PolicyTerms of Serviceand Dataset License.To protect your privacy, all features that rely on external API calls from your browser are turned off by default.

You need to opt-in for them to become active. All settings here will be stored as cookies with your web browser. For more information see our F. Add open access links from to the list of external document links if available.

Privacy notice: By enabling the option above, your browser will contact the API of unpaywall. Although we do not have any reason to believe that your call will be tracked, we do not have any control over how the remote server uses your data.

So please proceed with care and consider checking the Unpaywall privacy policy. For web page which are no longer available, try to retrieve content from the of the Internet Archive if available. Privacy notice: By enabling the option above, your browser will contact the API of web.

zeyuan hu (zeyuan)

So please proceed with care and consider checking the Internet Archive privacy policy. Add a list of references from and to record detail pages. Privacy notice: By enabling the option above, your browser will contact the APIs of crossref. So please proceed with care and consider checking the Crossref privacy policy and the OpenCitations privacy policy.

Add a list of citing articles from to record detail pages. Privacy notice: By enabling the option above, your browser will contact the API of opencitations. So please proceed with care and consider checking the OpenCitations privacy policy. Show tweets from on the dblp homepage. Privacy notice: By enabling the option above, your browser will contact twitter. At the same time, Twitter will persitently store several cookies with your web browser. While we did signal Twitter to not track our users by setting the "dnt" flagwe do not have any control over how Twitter uses your data.

So please proceed with care and consider checking the Twitter privacy policy. Authors: no matches. Venues: no matches. Publications: no matches. Trier 2 Dagstuhl. Aditya Bhaskara. Wei Chen Zheng Chen Alessandro Chiesa.

Rati Gelashvili. Michael I. Jonathan A.I work on the mathematical foundations of machine learning and optimization, and apply them to deep learning, theoretical computer science, operations research, and statistics.

I am also interested in the mathematical modeling for physical, social, economic, and biological systems. JMLR Math. I would love to thank my wonderful collaborators without whom these results below would never have been accomplished. In inverse chronological order:. My middle name "Allen" was legally merged into my family name in Feburarybecoming "Allen-Zhu". How does a 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 simply by applying stochastic gradient descent SGD. On the conceptual side, we present, to the best of our knowledge, the FIRST theory result indicating how very deep neural networks can still 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 are efficient.

We establish a new principle called "backward feature correction", which we believe is the key to understand the hierarchical learning in multi-layer neural networks.

Click here to show only equal first-author papers. Click here to show all papers. The experimental design problem concerns the selection of k points from a potentially large design pool of p-dimensional vectors, so as to maximize the statistical efficiency regressed on the selected k design points. Except for the T-optimality, exact optimization is NP-hard. Can we more provide theoretical justifications for this gap?

There is an influential line of work relating neural networks to kernels in the over-parameterized regime, proving that they can learn certain concept class that is also learnable by kernels, with similar test error.

Yet, can we show neural networks provably learn some concept class better than kernels? We answer this positively in the PAC learning language. 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 that the test error obtained by neural networks can be much smaller than any "generic" kernel method, including neural tangent kernels, conjugate kernels, etc.

The main intuition is that multi-layer neural networks can implicitly perform hierarchal learning using different layers, which reduces the sample complexity comparing to "one-shot" learning algorithms such as kernel methods.

The fundamental learning theory behind neural networks remains largely open. What classes of functions can neural networks actually learn? Why doesn't the trained neural networks overfit when the it is overparameterized namely, having more parameters than statistically needed to overfit training data? In this work, we prove that overparameterized neural networks can learn some notable concept classes, including two and three-layer networks with fewer parameters and smooth activations.Example 5.

By convention, this argument is always named self. You need to specify self explicitly when defining the method, you do not specify it when calling the method; Python will add it for you automatically. The return value will be the newly created object. In this case, each instance of UserDict will have a data attribute data. The update method is a dictionary duplicator:.

Think of update as a merge function, not a copy function. Unlike Java, Python does not support function overloading by argument list i. Method are defined solely by their name, and there can be only one method per class with a given name. And the same rule applies to any other method. Then you immediately return the new UserDict. In versions of Python prior to 2. To compensate for this, Python comes with wrapper classes that mimic the behavior of these built-in datatypes: UserStringUserListUserDict.

In Python 2. You are inheriting from dict directly, instead of from UserDict. In addition to normal class methods, there are a number of special methods that Python classes can define.

Instead of being called directly by your code like normal methodsspecial methods are called for you by Python in particular cirmustances or when specific syntax is used. This allows you to define classes that act like dictionaries in some ways but define their own behavior above and beyong the built-in dictionary.

When accessing data attribute within a class, you need to qualify the attribute name: self. When calling other methods within a class, you need to qualify the method name: self. This example shows some of the other special methods in UserDict.

Class attributes can be used as class-level constants, but they are not really constants. You can also change them. It is a reference to the class that self is an instance of in this case, the counter class. This affects the class itself, not just the newly created instance. This only because self.

Diving In 5. Importing Modules Using from module import 5. Defining Classes 5. Instantiating Classes 5. Special Class Methods 5. Advanced Special Class Methods 5.

Introducing Class Attributes 5. Private Functions About this site About search this site. Instantiate appropriate class with filename.

zeyuan hu (zeyuan)

Returned object acts like a dictionary, with key-value pairs for each piece of metadata. Framework can be extended by adding classes for particular file types, e. Each class is completely responsible for parsing its files appropriately; see MP3FileInfo for example. In [14]: import types In [16]: types. When you call a method of an ancestor class from within your class, you must include self argument.How does a layer ResNet learn a high-complexity classifier using rel We propose a rank-k variant of the classical Frank-Wolfe algorithm to so Given a nonconvex function f x that is an average of n smooth functions The online problem of computing the top eigenvector is fundamental to ma We design a non-convex second-order optimization algorithm that is guara We solve principal component regression PCRup to a multiplicative ac We study streaming principal component analysis PCAthat is to find, We study k-GenEV, the problem of finding the top k generalized eigenvect We study k-SVD that is to obtain the first k singular vectors of a matri Nesterov's momentum trick is famously known for accelerating gradient de We consider the fundamental problem in non-convex optimization of effici The diverse world of machine learning applications has given rise to a p The amount of data available in the world is growing faster than our abi Accelerated coordinate descent is widely used in optimization due to its In this paper, we provide a novel construction of the linear-sized spect Many classical algorithms are found until several years later to outlive First-order methods play a central role in large-scale machine learning Given a subset S of vertices of an undirected graph G, the cut-improveme Motivated by applications of large-scale graph clustering, we study rand We propose a reduction for non-convex optimization that can 1 turn a s In convex stochastic optimization, convergence rates in terms of minimiz The problem of minimizing sum-of-nonconvex functions i.

This paper studies the problem of distributed stochastic optimization in The experimental design problem concerns the selection of k points from Model-free reinforcement learning RL algorithms, such as Q-learning, d We propose a new second-order method for geodesically convex optimizatio Despite the huge success of deep learning, our understanding to how the Deep neural networks DNNs have demonstrated dominating performance in Neural networks have great success in many machine learning applications Classically, the time complexity of a first-order method is estimated by Are you a researcher?

Expose your work to one of the largest A. Already have an account? Login here. Don't have an account? Signup here. Zeyuan Allen-Zhu is this you? Featured Co-authors. Michael I.Modify char in another function. Generalized binary search. Cache, Lease, Consistency, Invalidation. State Machine Replication Approach. Lamport Clocks, Vector Clocks.

Zeyuan Allen-Zhu

Distributed System Reference Guide. How to write binary search correctly. Introduction to Conditional Random Fields.

zeyuan hu (zeyuan)

MAW Chapter 8: Disjoint set. Understanding how function call works. Andrew Ng's ML Week 06, The tortoise and the hare. Python case study: leetcode scraper. Solving recurrence relations part 2. Draw a Neural Network through Graphviz. Andrew Ng's ML Week 04 - Andrew Ng's ML Week 01 - Simple sorting algorithms. Introducing the "Andrew Ng's ML course study notes". MAW Chapter 7: Sorting writing questions. MAW: Chapter 6 Reflection.

MAW Chapter 5: Hashing writing questions. MAW: Chapter 4 Reflection. Solving recurrence relations in a nutshell. MAW Chapter 4: Tree writing questions. MAW: Chapter 3 Reflection.

Num of function calls in recursive Fibonacci routine. Modify array inside function in C. A peek in code optimization. Reflection on integer arithmetic package problem.

A small C trick I learned today. Print singly linked list in reverse order. Environment variable substitution using Sed. What's the difference between sourcing a script and executing a script? Polynomial Multiplication.

【绝代双骄】花絮:陈哲远特辑 - Handsome Siblings

Lesson Learned: Portability. Reverse Singly Linked List. Generate a Linked List from a given array.


Zeyuan hu (zeyuan)