Unnamed: 0.1
int64
0
41k
Unnamed: 0
int64
0
41k
author
stringlengths
9
1.39k
id
stringlengths
11
18
summary
stringlengths
25
3.66k
title
stringlengths
4
258
year
int64
1.99k
2.02k
arxiv_url
stringlengths
32
39
info
stringlengths
523
3.18k
embeddings
stringlengths
16.9k
17.1k
1,000
1,000
['Pierre Baldi', 'Peter Sadowski']
1506.06472v2
In a physical neural system, where storage and processing are intimately intertwined, the rules for adjusting the synaptic weights can only depend on variables that are available locally, such as the activity of the pre- and post-synaptic neurons, resulting in local learning rules. A systematic framework for studying t...
A Theory of Local Learning, the Learning Channel, and the Optimality of Backpropagation
2,015
http://arxiv.org/pdf/1506.06472v2
Title Theory Local Learning Learning Channel Optimality Backpropagation Summary physical neural system storage processing intimately intertwined rule adjusting synaptic weight depend variable available locally activity pre postsynaptic neuron resulting local learning rule systematic framework studying space local learn...
[0.00760247278958559, -0.0013965636026114225, -0.011607717722654343, 0.050949905067682266, 0.007959111593663692, -0.033396292477846146, 0.041610006242990494, -0.011325334198772907, -0.03875073418021202, 0.015684904530644417, -0.019185174256563187, -0.005184045527130365, 0.005953799933195114, 0.04171661287546158, 0.0443...
1,001
1,001
['Majid Janzamin', 'Hanie Sedghi', 'Anima Anandkumar']
1506.08473v3
Training neural networks is a challenging non-convex optimization problem, and backpropagation or gradient descent can get stuck in spurious local optima. We propose a novel algorithm based on tensor decomposition for guaranteed training of two-layer neural networks. We provide risk bounds for our proposed method, with...
Beating the Perils of Non-Convexity: Guaranteed Training of Neural Networks using Tensor Methods
2,015
http://arxiv.org/pdf/1506.08473v3
Title Beating Perils NonConvexity Guaranteed Training Neural Networks using Tensor Methods Summary Training neural network challenging nonconvex optimization problem backpropagation gradient descent get stuck spurious local optimum propose novel algorithm based tensor decomposition guaranteed training twolayer neural n...
[-0.011419358663260937, 0.06404642760753632, -0.025790397077798843, 0.03155630826950073, 0.006474959198385477, -0.004853494465351105, 0.043829306960105896, 0.0015866432804614305, -0.0032157341483980417, 0.00563422916457057, -0.0043036011047661304, -0.018643533810973167, 0.002184829441830516, -0.0006265626288950443, 0.0...
1,002
1,002
['Guillaume Desjardins', 'Karen Simonyan', 'Razvan Pascanu', 'Koray Kavukcuoglu']
1507.00210v1
We introduce Natural Neural Networks, a novel family of algorithms that speed up convergence by adapting their internal representation during training to improve conditioning of the Fisher matrix. In particular, we show a specific example that employs a simple and efficient reparametrization of the neural network weigh...
Natural Neural Networks
2,015
http://arxiv.org/pdf/1507.00210v1
Title Natural Neural Networks Summary introduce Natural Neural Networks novel family algorithm speed convergence adapting internal representation training improve conditioning Fisher matrix particular show specific example employ simple efficient reparametrization neural network weight implicitly whitening representati...
[-0.023316465318202972, 0.04403283819556236, -0.008403029292821884, 0.04698909819126129, 0.006800304166972637, -0.028438495472073555, 0.009840313345193863, 0.003920476883649826, 0.007934129796922207, 0.025590742006897926, -0.04626111686229706, 0.030625753104686737, 0.024648960679769516, -0.0024142409674823284, 0.023473...
1,003
1,003
['Antti Rasmus', 'Harri Valpola', 'Mikko Honkala', 'Mathias Berglund', 'Tapani Raiko']
1507.02672v2
We combine supervised learning with unsupervised learning in deep neural networks. The proposed model is trained to simultaneously minimize the sum of supervised and unsupervised cost functions by backpropagation, avoiding the need for layer-wise pre-training. Our work builds on the Ladder network proposed by Valpola (...
Semi-Supervised Learning with Ladder Networks
2,015
http://arxiv.org/pdf/1507.02672v2
Title SemiSupervised Learning Ladder Networks Summary combine supervised learning unsupervised learning deep neural network proposed model trained simultaneously minimize sum supervised unsupervised cost function backpropagation avoiding need layerwise pretraining work build Ladder network proposed Valpola 2015 extend ...
[-0.039266400039196014, 0.0740010216832161, 0.00045614270493388176, 0.012034674175083637, 0.013148428872227669, 0.006837876047939062, 0.026733368635177612, -0.05170517787337303, 0.022176256403326988, 0.04301025718450546, -0.04440532997250557, 0.011744906194508076, -0.00117692188359797, 0.06924199312925339, 0.0464486442...
1,004
1,004
['Rossella Cancelliere', 'Mario Gai', 'Patrick Gallinari', 'Luca Rubini']
1508.06095v1
In this paper we consider the training of single hidden layer neural networks by pseudoinversion, which, in spite of its popularity, is sometimes affected by numerical instability issues. Regularization is known to be effective in such cases, so that we introduce, in the framework of Tikhonov regularization, a matricia...
OCReP: An Optimally Conditioned Regularization for Pseudoinversion Based Neural Training
2,015
http://arxiv.org/pdf/1508.06095v1
Title OCReP Optimally Conditioned Regularization Pseudoinversion Based Neural Training Summary paper consider training single hidden layer neural network pseudoinversion spite popularity sometimes affected numerical instability issue Regularization known effective case introduce framework Tikhonov regularization matric...
[-0.01080725621432066, 0.028573397547006607, -0.0020157950930297375, -0.0019419288728386164, 0.023659273982048035, -0.03376099467277527, 0.0176223274320364, 0.042202483862638474, -0.021722324192523956, 0.01823808066546917, -0.02855880744755268, 0.022159788757562637, 0.0282483771443367, -0.014263048768043518, 0.04104025...
1,005
1,005
['Yanping Huang', 'Sai Zhang']
1508.07096v1
Deep learning methods have shown great promise in many practical applications, ranging from speech recognition, visual object recognition, to text processing. However, most of the current deep learning methods suffer from scalability problems for large-scale applications, forcing researchers or users to focus on small-...
Partitioning Large Scale Deep Belief Networks Using Dropout
2,015
http://arxiv.org/pdf/1508.07096v1
Title Partitioning Large Scale Deep Belief Networks Using Dropout Summary Deep learning method shown great promise many practical application ranging speech recognition visual object recognition text processing However current deep learning method suffer scalability problem largescale application forcing researcher use...
[-0.01561122015118599, 0.05842526629567146, -0.01607709378004074, 0.062400564551353455, -0.03784314915537834, 0.018610667437314987, 0.09203869849443436, -0.02367348037660122, 0.0642785057425499, -0.0076059335842728615, -0.02280244044959545, -0.03673293814063072, -0.02257743664085865, 0.08498809486627579, -0.01718771643...
1,006
1,006
['Minhyung Cho', 'Chandra Shekhar Dhir', 'Jaehyung Lee']
1509.03475v2
Multidimensional recurrent neural networks (MDRNNs) have shown a remarkable performance in the area of speech and handwriting recognition. The performance of an MDRNN is improved by further increasing its depth, and the difficulty of learning the deeper network is overcome by using Hessian-free (HF) optimization. Given...
Hessian-free Optimization for Learning Deep Multidimensional Recurrent Neural Networks
2,015
http://arxiv.org/pdf/1509.03475v2
Title Hessianfree Optimization Learning Deep Multidimensional Recurrent Neural Networks Summary Multidimensional recurrent neural network MDRNNs shown remarkable performance area speech handwriting recognition performance MDRNN improved increasing depth difficulty learning deeper network overcome using Hessianfree HF o...
[-0.02317788451910019, 0.050238821655511856, -0.0011568962363526225, 0.07076552510261536, 0.006518575828522444, 0.014560024254024029, 0.03768059238791466, -0.01388279814273119, 0.007555229589343071, -0.022181354463100433, -0.041420578956604004, -0.06325777620077133, 0.03561465069651604, -0.010540228337049484, 0.0050849...
1,007
1,007
['Nadav Cohen', 'Or Sharir', 'Amnon Shashua']
1509.05009v3
It has long been conjectured that hypotheses spaces suitable for data that is compositional in nature, such as text or images, may be more efficiently represented with deep hierarchical networks than with shallow ones. Despite the vast empirical evidence supporting this belief, theoretical justifications to date are li...
On the Expressive Power of Deep Learning: A Tensor Analysis
2,015
http://arxiv.org/pdf/1509.05009v3
Title Expressive Power Deep Learning Tensor Analysis Summary long conjectured hypothesis space suitable data compositional nature text image may efficiently represented deep hierarchical network shallow one Despite vast empirical evidence supporting belief theoretical justification date limited particular account local...
[-0.025506973266601562, 0.05522305518388748, -0.024910204112529755, 0.05504650995135307, -0.018744299188256264, 0.011751165613532066, 0.05135875195264816, 0.02122006006538868, -0.030108461156487465, -0.03114894963800907, -0.02738891914486885, -0.003963015973567963, 0.02069639042019844, 0.057300031185150146, 0.032450508...
1,008
1,008
['David Nova', 'Pablo A. Estevez']
1509.07093v1
In this work we present a review of the state of the art of Learning Vector Quantization (LVQ) classifiers. A taxonomy is proposed which integrates the most relevant LVQ approaches to date. The main concepts associated with modern LVQ approaches are defined. A comparison is made among eleven LVQ classifiers using one r...
A review of learning vector quantization classifiers
2,015
http://arxiv.org/pdf/1509.07093v1
Title review learning vector quantization classifier Summary work present review state art Learning Vector Quantization LVQ classifier taxonomy proposed integrates relevant LVQ approach date main concept associated modern LVQ approach defined comparison made among eleven LVQ classifier using one realworld two artificia...
[-0.05977300554513931, 0.004736436530947685, -0.029403463006019592, 0.014884956181049347, 0.022385310381650925, 0.007275397889316082, 0.005885349586606026, 0.059824470430612564, -0.000953124719671905, -0.02724575437605381, 0.04968015104532242, -0.009075534529983997, -0.017313819378614426, 0.027814656496047974, -0.01105...
1,009
1,009
['Uri Shaham', 'Alexander Cloninger', 'Ronald R. Coifman']
1509.07385v3
We discuss approximation of functions using deep neural nets. Given a function $f$ on a $d$-dimensional manifold $\Gamma \subset \mathbb{R}^m$, we construct a sparsely-connected depth-4 neural network and bound its error in approximating $f$. The size of the network depends on dimension and curvature of the manifold $\...
Provable approximation properties for deep neural networks
2,015
http://arxiv.org/pdf/1509.07385v3
Title Provable approximation property deep neural network Summary discus approximation function using deep neural net Given function f ddimensional manifold Gamma subset mathbbRm construct sparselyconnected depth4 neural network bound error approximating f size network depends dimension curvature manifold Gamma complex...
[-0.05590597912669182, 0.01858479715883732, 0.028689654543995857, 0.0371275320649147, 0.0035450495779514313, -0.01502046175301075, 0.011005731299519539, -0.027167141437530518, 0.0021194193977862597, 0.0738958790898323, 0.006760123651474714, -0.008111672475934029, -0.01495177298784256, 0.04407349228858948, 0.03492267429...
1,010
1,010
['David Balduzzi']
1509.08627v1
Deep learning is currently the subject of intensive study. However, fundamental concepts such as representations are not formally defined -- researchers "know them when they see them" -- and there is no common language for describing and analyzing algorithms. This essay proposes an abstract framework that identifies th...
Semantics, Representations and Grammars for Deep Learning
2,015
http://arxiv.org/pdf/1509.08627v1
Title Semantics Representations Grammars Deep Learning Summary Deep learning currently subject intensive study However fundamental concept representation formally defined researcher know see common language describing analyzing algorithm essay proposes abstract framework identifies essential feature current practice ma...
[-0.0013996713096275926, 0.020266596227884293, -0.022982850670814514, 0.01866045780479908, -0.047065068036317825, -0.0113881416618824, 0.037124473601579666, -0.010451005771756172, -0.05943715199828148, -0.05291685461997986, 0.00984866637736559, -0.010585596784949303, 0.0031278925016522408, 0.09333193302154541, 0.024312...
1,011
1,011
['Chen-Yu Lee', 'Patrick W. Gallagher', 'Zhuowen Tu']
1509.08985v2
We seek to improve deep neural networks by generalizing the pooling operations that play a central role in current architectures. We pursue a careful exploration of approaches to allow pooling to learn and to adapt to complex and variable patterns. The two primary directions lie in (1) learning a pooling function via (...
Generalizing Pooling Functions in Convolutional Neural Networks: Mixed, Gated, and Tree
2,015
http://arxiv.org/pdf/1509.08985v2
Title Generalizing Pooling Functions Convolutional Neural Networks Mixed Gated Tree Summary seek improve deep neural network generalizing pooling operation play central role current architecture pursue careful exploration approach allow pooling learn adapt complex variable pattern two primary direction lie 1 learning p...
[-0.005272310227155685, 0.025088487192988396, 0.0010660000843927264, 0.03123207576572895, 0.017109090462327003, -0.010477676056325436, 0.013040158897638321, -0.016844142228364944, -0.04249121993780136, -0.002917418023571372, -0.02521940879523754, -0.04021741822361946, 0.02709580585360527, 0.03698455169796944, 0.0172950...
1,012
1,012
['César Laurent', 'Gabriel Pereyra', 'Philémon Brakel', 'Ying Zhang', 'Yoshua Bengio']
1510.01378v1
Recurrent Neural Networks (RNNs) are powerful models for sequential data that have the potential to learn long-term dependencies. However, they are computationally expensive to train and difficult to parallelize. Recent work has shown that normalizing intermediate representations of neural networks can significantly im...
Batch Normalized Recurrent Neural Networks
2,015
http://arxiv.org/pdf/1510.01378v1
Title Batch Normalized Recurrent Neural Networks Summary Recurrent Neural Networks RNNs powerful model sequential data potential learn longterm dependency However computationally expensive train difficult parallelize Recent work shown normalizing intermediate representation neural network significantly improve converge...
[-0.027521861717104912, 0.03971173241734505, 0.023022178560495377, 0.019709840416908264, -0.012142783030867577, -0.00977847259491682, 0.04466661438345909, 0.0017847318667918444, -0.029096905142068863, 0.001522875507362187, -0.02627219259738922, -0.011297101154923439, 0.03435905650258064, 0.027394583448767662, -0.019226...
1,013
1,013
['Ben Krause']
1510.04953v1
Recurrent Neural Networks (RNNs) have long been recognized for their potential to model complex time series. However, it remains to be determined what optimization techniques and recurrent architectures can be used to best realize this potential. The experiments presented take a deep look into Hessian free optimization...
Optimizing and Contrasting Recurrent Neural Network Architectures
2,015
http://arxiv.org/pdf/1510.04953v1
Title Optimizing Contrasting Recurrent Neural Network Architectures Summary Recurrent Neural Networks RNNs long recognized potential model complex time series However remains determined optimization technique recurrent architecture used best realize potential experiment presented take deep look Hessian free optimizatio...
[0.018120326101779938, 0.06562254577875137, 0.00030987634090706706, 0.02336784452199936, -0.02435116656124592, -0.012681765481829643, 0.018707696348428726, 0.014799770899116993, -0.033419013023376465, -0.008177623152732849, -0.017388323321938515, -0.08129463344812393, 0.043445803225040436, 0.00026696265558712184, 0.020...
1,014
1,014
['Alexandre de Brébisson', 'Pascal Vincent']
1511.05042v3
In a multi-class classification problem, it is standard to model the output of a neural network as a categorical distribution conditioned on the inputs. The output must therefore be positive and sum to one, which is traditionally enforced by a softmax. This probabilistic mapping allows to use the maximum likelihood pri...
An Exploration of Softmax Alternatives Belonging to the Spherical Loss Family
2,015
http://arxiv.org/pdf/1511.05042v3
Title Exploration Softmax Alternatives Belonging Spherical Loss Family Summary multiclass classification problem standard model output neural network categorical distribution conditioned input output must therefore positive sum one traditionally enforced softmax probabilistic mapping allows use maximum likelihood princ...
[-0.011322312988340855, 0.0020932150073349476, 0.011889001354575157, 0.007676499430090189, 0.022251417860388756, -0.015849504619836807, 0.030473507940769196, -0.00724671920761466, -0.03328094258904457, -0.018747184425592422, -0.04571295529603958, -0.0598076656460762, 0.015659049153327942, 0.04293930158019066, 0.0299765...
1,015
1,015
['Uri Shaham', 'Yutaro Yamada', 'Sahand Negahban']
1511.05432v3
We propose a general framework for increasing local stability of Artificial Neural Nets (ANNs) using Robust Optimization (RO). We achieve this through an alternating minimization-maximization procedure, in which the loss of the network is minimized over perturbed examples that are generated at each parameter update. We...
Understanding Adversarial Training: Increasing Local Stability of Neural Nets through Robust Optimization
2,015
http://arxiv.org/pdf/1511.05432v3
Title Understanding Adversarial Training Increasing Local Stability Neural Nets Robust Optimization Summary propose general framework increasing local stability Artificial Neural Nets ANNs using Robust Optimization RO achieve alternating minimizationmaximization procedure loss network minimized perturbed example genera...
[-0.004029142204672098, 0.015982283279299736, -0.017641833052039146, 0.02705393359065056, 0.027294622734189034, -0.04971231520175934, 0.0039017926901578903, -0.017038116231560707, 0.0030511992517858744, 0.0006444669561460614, 0.007238368503749371, 0.003184942062944174, 0.008727705106139183, -0.004711437039077282, 0.062...
1,016
1,016
['Martin Arjovsky', 'Amar Shah', 'Yoshua Bengio']
1511.06464v4
Recurrent neural networks (RNNs) are notoriously difficult to train. When the eigenvalues of the hidden to hidden weight matrix deviate from absolute value 1, optimization becomes difficult due to the well studied issue of vanishing and exploding gradients, especially when trying to learn long-term dependencies. To cir...
Unitary Evolution Recurrent Neural Networks
2,015
http://arxiv.org/pdf/1511.06464v4
Title Unitary Evolution Recurrent Neural Networks Summary Recurrent neural network RNNs notoriously difficult train eigenvalue hidden hidden weight matrix deviate absolute value 1 optimization becomes difficult due well studied issue vanishing exploding gradient especially trying learn longterm dependency circumvent pr...
[0.0019006810616701841, 0.0763901025056839, 0.00197855313308537, 0.02733837068080902, 0.0023853941820561886, 0.005786016117781401, -0.0056060669012367725, -0.0029748764354735613, -0.024224046617746353, -0.008765719830989838, -0.011217683553695679, -0.022924989461898804, -0.008992144837975502, 0.00693783862516284, 0.016...
1,017
1,017
['Nicolas Papernot', 'Patrick McDaniel', 'Somesh Jha', 'Matt Fredrikson', 'Z. Berkay Celik', 'Ananthram Swami']
1511.07528v1
Deep learning takes advantage of large datasets and computationally efficient training algorithms to outperform other approaches at various machine learning tasks. However, imperfections in the training phase of deep neural networks make them vulnerable to adversarial samples: inputs crafted by adversaries with the int...
The Limitations of Deep Learning in Adversarial Settings
2,015
http://arxiv.org/pdf/1511.07528v1
Title Limitations Deep Learning Adversarial Settings Summary Deep learning take advantage large datasets computationally efficient training algorithm outperform approach various machine learning task However imperfection training phase deep neural network make vulnerable adversarial sample input crafted adversary inten...
[0.02380998805165291, 0.06371823698282242, -0.040228307247161865, 0.011923722922801971, -0.02521265298128128, -0.01379898376762867, 0.06262260675430298, -0.021116524934768677, -0.0199461430311203, -0.03555198386311531, 0.012148430570960045, 0.01339671015739441, 0.026439659297466278, 0.039241306483745575, 0.066861122846...
1,018
1,018
['Ronen Eldan', 'Ohad Shamir']
1512.03965v4
We show that there is a simple (approximately radial) function on $\reals^d$, expressible by a small 3-layer feedforward neural networks, which cannot be approximated by any 2-layer network, to more than a certain constant accuracy, unless its width is exponential in the dimension. The result holds for virtually all kn...
The Power of Depth for Feedforward Neural Networks
2,015
http://arxiv.org/pdf/1512.03965v4
Title Power Depth Feedforward Neural Networks Summary show simple approximately radial function realsd expressible small 3layer feedforward neural network cannot approximated 2layer network certain constant accuracy unless width exponential dimension result hold virtually known activation function including rectified l...
[-0.025231948122382164, -0.015978693962097168, -0.019615404307842255, 0.030008630827069283, -0.008808150887489319, -0.03327452391386032, 0.0336979515850544, -0.039953604340553284, -0.01616176962852478, 0.0377151295542717, 0.04056018963456154, -0.00040590125718154013, 0.02106557972729206, 0.050211530178785324, 0.0497252...
1,019
1,019
['Sajid Anwar', 'Kyuyeon Hwang', 'Wonyong Sung']
1512.08571v1
Real time application of deep learning algorithms is often hindered by high computational complexity and frequent memory accesses. Network pruning is a promising technique to solve this problem. However, pruning usually results in irregular network connections that not only demand extra representation efforts but also ...
Structured Pruning of Deep Convolutional Neural Networks
2,015
http://arxiv.org/pdf/1512.08571v1
Title Structured Pruning Deep Convolutional Neural Networks Summary Real time application deep learning algorithm often hindered high computational complexity frequent memory access Network pruning promising technique solve problem However pruning usually result irregular network connection demand extra representation ...
[-0.010910236276686192, 0.015242780558764935, -0.020245423540472984, 0.07711878418922424, 0.018374234437942505, -0.05077040195465088, 0.04685645550489426, 0.03043891116976738, -0.045074425637722015, 0.017209837213158607, 0.012775520794093609, -0.01960749737918377, -0.00017783624934963882, 0.038504406809806824, -0.01493...
1,020
1,020
['Uri Shaham', 'Roy Lederman']
1512.08806v3
We consider the statistical problem of learning common source of variability in data which are synchronously captured by multiple sensors, and demonstrate that Siamese neural networks can be naturally applied to this problem. This approach is useful in particular in exploratory, data-driven applications, where neither ...
Common Variable Learning and Invariant Representation Learning using Siamese Neural Networks
2,015
http://arxiv.org/pdf/1512.08806v3
Title Common Variable Learning Invariant Representation Learning using Siamese Neural Networks Summary consider statistical problem learning common source variability data synchronously captured multiple sensor demonstrate Siamese neural network naturally applied problem approach useful particular exploratory datadrive...
[-0.0544157475233078, 0.024771135300397873, -0.008793048560619354, 0.014006986282765865, -0.026938587427139282, 0.025043006986379623, 0.04505455493927002, -0.01592770591378212, -0.013735727407038212, -0.007581098936498165, -0.0060582952573895454, 0.04268019273877144, 0.030583487823605537, 0.032423462718725204, 0.027445...
1,021
1,021
['Siamak Ravanbakhsh', 'Barnabas Poczos', 'Jeff Schneider', 'Dale Schuurmans', 'Russell Greiner']
1601.00034v4
We propose a Laplace approximation that creates a stochastic unit from any smooth monotonic activation function, using only Gaussian noise. This paper investigates the application of this stochastic approximation in training a family of Restricted Boltzmann Machines (RBM) that are closely linked to Bregman divergences....
Stochastic Neural Networks with Monotonic Activation Functions
2,016
http://arxiv.org/pdf/1601.00034v4
Title Stochastic Neural Networks Monotonic Activation Functions Summary propose Laplace approximation creates stochastic unit smooth monotonic activation function using Gaussian noise paper investigates application stochastic approximation training family Restricted Boltzmann Machines RBM closely linked Bregman diverge...
[-0.022105617448687553, -0.012747292406857014, -0.013226406648755074, 0.007011051755398512, -0.010834104381501675, -0.018352391198277473, 0.0627669170498848, -0.026270363479852676, -0.03777092695236206, 0.04276793822646141, -0.0025533935986459255, -0.04123224690556526, 0.0041479626670479774, 0.07758092880249023, 0.0235...
1,022
1,022
['Conrado S. Miranda', 'Fernando J. Von Zuben']
1602.01164v1
This paper introduces the hypervolume maximization with a single solution as an alternative to the mean loss minimization. The relationship between the two problems is proved through bounds on the cost function when an optimal solution to one of the problems is evaluated on the other, with a hyperparameter to control t...
Single-Solution Hypervolume Maximization and its use for Improving Generalization of Neural Networks
2,016
http://arxiv.org/pdf/1602.01164v1
Title SingleSolution Hypervolume Maximization use Improving Generalization Neural Networks Summary paper introduces hypervolume maximization single solution alternative mean loss minimization relationship two problem proved bound cost function optimal solution one problem evaluated hyperparameter control similarity two...
[-0.0027464719023555517, 0.009001024067401886, -0.008193577639758587, 0.02726939134299755, 0.02189466543495655, -0.023885119706392288, 0.04536872357130051, 0.019456245005130768, -0.015219427645206451, 0.012663464993238449, -0.02107631228864193, 0.042969170957803726, 0.005190565250813961, -0.012795457616448402, 0.017654...
1,023
1,023
['Mark Tygert']
1602.02823v1
Poor (even random) starting points for learning/training/optimization are common in machine learning. In many settings, the method of Robbins and Monro (online stochastic gradient descent) is known to be optimal for good starting points, but may not be optimal for poor starting points -- indeed, for poor starting point...
Poor starting points in machine learning
2,016
http://arxiv.org/pdf/1602.02823v1
Title Poor starting point machine learning Summary Poor even random starting point learningtrainingoptimization common machine learning many setting method Robbins Monro online stochastic gradient descent known optimal good starting point may optimal poor starting point indeed poor starting point Nesterov acceleration ...
[-0.03629722073674202, 0.018068384379148483, -0.008501116186380386, -0.019850803539156914, -0.005484607070684433, 0.01216861791908741, 0.022362254559993744, 0.033891402184963226, -0.00452084606513381, -0.012383509427309036, 0.08977865427732468, 0.0027387761510908604, 0.017566345632076263, 0.03579370304942131, -0.004166...
1,024
1,024
['Matus Telgarsky']
1602.04485v2
For any positive integer $k$, there exist neural networks with $\Theta(k^3)$ layers, $\Theta(1)$ nodes per layer, and $\Theta(1)$ distinct parameters which can not be approximated by networks with $\mathcal{O}(k)$ layers unless they are exponentially large --- they must possess $\Omega(2^k)$ nodes. This result is prove...
Benefits of depth in neural networks
2,016
http://arxiv.org/pdf/1602.04485v2
Title Benefits depth neural network Summary positive integer k exist neural network Thetak3 layer Theta1 node per layer Theta1 distinct parameter approximated network mathcalOk layer unless exponentially large must posse Omega2k node result proved class node termed semialgebraic gate includes common choice ReLU maximum...
[-0.05724705755710602, 0.05195591598749161, -0.010790308937430382, 0.07703837752342224, -0.032453905791044235, -0.05051332339644432, 0.027538588270545006, -0.005830776412039995, -0.012861654162406921, -0.017301520332694054, 0.07778756320476532, 0.04497380182147026, -0.008873707614839077, 0.069909006357193, 0.0183462016...
1,025
1,025
['Prateek Jain', 'Chi Jin', 'Sham M. Kakade', 'Praneeth Netrapalli', 'Aaron Sidford']
1602.06929v2
This work provides improved guarantees for streaming principle component analysis (PCA). Given $A_1, \ldots, A_n\in \mathbb{R}^{d\times d}$ sampled independently from distributions satisfying $\mathbb{E}[A_i] = \Sigma$ for $\Sigma \succeq \mathbf{0}$, this work provides an $O(d)$-space linear-time single-pass streaming...
Streaming PCA: Matching Matrix Bernstein and Near-Optimal Finite Sample Guarantees for Oja's Algorithm
2,016
http://arxiv.org/pdf/1602.06929v2
Title Streaming PCA Matching Matrix Bernstein NearOptimal Finite Sample Guarantees Ojas Algorithm Summary work provides improved guarantee streaming principle component analysis PCA Given A1 ldots Anin mathbbRdtimes sampled independently distribution satisfying mathbbEAi Sigma Sigma succeq mathbf0 work provides Odspace...
[-0.05177688226103783, 0.0025803453754633665, -0.032424625009298325, 0.030767343938350677, -0.0019016548758372664, 0.0030492173973470926, -0.0014358807820826769, 0.0364130362868309, -0.017607303336262703, 0.002995114540681243, 0.040183454751968384, -0.0178676750510931, 0.05465877801179886, 0.09350184351205826, -0.00825...
1,026
1,026
['Gaétan Marceau-Caron', 'Yann Ollivier']
1602.08007v1
We provide the first experimental results on non-synthetic datasets for the quasi-diagonal Riemannian gradient descents for neural networks introduced in [Ollivier, 2015]. These include the MNIST, SVHN, and FACE datasets as well as a previously unpublished electroencephalogram dataset. The quasi-diagonal Riemannian alg...
Practical Riemannian Neural Networks
2,016
http://arxiv.org/pdf/1602.08007v1
Title Practical Riemannian Neural Networks Summary provide first experimental result nonsynthetic datasets quasidiagonal Riemannian gradient descent neural network introduced Ollivier 2015 include MNIST SVHN FACE datasets well previously unpublished electroencephalogram dataset quasidiagonal Riemannian algorithm consis...
[-0.01422437746077776, 0.015026831068098545, 0.001314919674769044, 0.03444475308060646, 0.007506160996854305, 0.007985351607203484, 0.03505011275410652, -0.021717244759202003, 0.019317619502544403, 0.06532058119773865, -0.050407420843839645, 0.0018943356117233634, 0.042873673141002655, 0.0022290360648185015, 0.06274704...
1,027
1,027
['Ryan Spring', 'Anshumali Shrivastava']
1602.08194v2
Current deep learning architectures are growing larger in order to learn from complex datasets. These architectures require giant matrix multiplication operations to train millions of parameters. Conversely, there is another growing trend to bring deep learning to low-power, embedded devices. The matrix operations, ass...
Scalable and Sustainable Deep Learning via Randomized Hashing
2,016
http://arxiv.org/pdf/1602.08194v2
Title Scalable Sustainable Deep Learning via Randomized Hashing Summary Current deep learning architecture growing larger order learn complex datasets architecture require giant matrix multiplication operation train million parameter Conversely another growing trend bring deep learning lowpower embedded device matrix o...
[0.0010114377364516258, 0.03945699706673622, -0.009542764164507389, 0.03983019292354584, 0.002935656812041998, -0.002689250512048602, 0.03509049117565155, -0.01152210496366024, -0.01446304377168417, -0.009551228024065495, 0.0005053200293332338, 0.015780463814735413, -0.016175471246242523, 0.05343971401453018, -0.002425...
1,028
1,028
['Caglar Gulcehre', 'Marcin Moczulski', 'Misha Denil', 'Yoshua Bengio']
1603.00391v3
Common nonlinear activation functions used in neural networks can cause training difficulties due to the saturation behavior of the activation function, which may hide dependencies that are not visible to vanilla-SGD (using first order gradients only). Gating mechanisms that use softly saturating activation functions t...
Noisy Activation Functions
2,016
http://arxiv.org/pdf/1603.00391v3
Title Noisy Activation Functions Summary Common nonlinear activation function used neural network cause training difficulty due saturation behavior activation function may hide dependency visible vanillaSGD using first order gradient Gating mechanism use softly saturating activation function emulate discrete switching ...
[-0.02933085896074772, -0.005668508354574442, -0.014781647361814976, 0.023311136290431023, 0.015231486409902573, -0.04363773763179779, 0.04838943108916283, -0.017878500744700432, -0.0022393171675503254, -0.009032201021909714, 0.0012381107080727816, 0.05114392191171646, 0.018531769514083862, 0.06450876593589783, 0.03022...
1,029
1,029
['Hanie Sedghi', 'Anima Anandkumar']
1603.00954v5
We consider the problem of training input-output recurrent neural networks (RNN) for sequence labeling tasks. We propose a novel spectral approach for learning the network parameters. It is based on decomposition of the cross-moment tensor between the output and a non-linear transformation of the input, based on score ...
Training Input-Output Recurrent Neural Networks through Spectral Methods
2,016
http://arxiv.org/pdf/1603.00954v5
Title Training InputOutput Recurrent Neural Networks Spectral Methods Summary consider problem training inputoutput recurrent neural network RNN sequence labeling task propose novel spectral approach learning network parameter based decomposition crossmoment tensor output nonlinear transformation input based score func...
[0.014532573521137238, 0.011313416063785553, 0.018280792981386185, 0.027680059894919395, -0.018323451280593872, -0.019043298438191414, -0.013410831801593304, -0.007060344330966473, -0.010939265601336956, -0.0278327576816082, -0.029019979760050774, -0.009128183126449585, 0.018436765298247337, 0.037789370864629745, -0.00...
1,030
1,030
['Wei Li', 'Melvin Gauci', 'Roderich Gross']
1603.04904v2
We propose Turing Learning, a novel system identification method for inferring the behavior of natural or artificial systems. Turing Learning simultaneously optimizes two populations of computer programs, one representing models of the behavior of the system under investigation, and the other representing classifiers. ...
Turing learning: a metric-free approach to inferring behavior and its application to swarms
2,016
http://arxiv.org/pdf/1603.04904v2
Title Turing learning metricfree approach inferring behavior application swarm Summary propose Turing Learning novel system identification method inferring behavior natural artificial system Turing Learning simultaneously optimizes two population computer program one representing model behavior system investigation rep...
[-0.02025068923830986, 0.012597393244504929, -0.024991095066070557, -0.002386539475992322, -0.00498027540743351, -0.002023395849391818, 0.001367067568935454, 0.037467971444129944, -0.005943096708506346, -0.02246209979057312, -0.012003985233604908, 0.05816323682665825, -0.035783879458904266, 0.005197503603994846, 0.0232...
1,031
1,031
['Sashank J. Reddi', 'Ahmed Hefny', 'Suvrit Sra', 'Barnabas Poczos', 'Alex Smola']
1603.06160v2
We study nonconvex finite-sum problems and analyze stochastic variance reduced gradient (SVRG) methods for them. SVRG and related methods have recently surged into prominence for convex optimization given their edge over stochastic gradient descent (SGD); but their theoretical analysis almost exclusively assumes convex...
Stochastic Variance Reduction for Nonconvex Optimization
2,016
http://arxiv.org/pdf/1603.06160v2
Title Stochastic Variance Reduction Nonconvex Optimization Summary study nonconvex finitesum problem analyze stochastic variance reduced gradient SVRG method SVRG related method recently surged prominence convex optimization given edge stochastic gradient descent SGD theoretical analysis almost exclusively assumes conv...
[-0.017825353890657425, -0.004558364395052195, 0.0011095072841271758, 0.02004215493798256, 0.007348756771534681, -0.03906097635626793, -0.02140718512237072, -0.0020931249018758535, -0.05875280126929283, 0.040182650089263916, -0.002938437508419156, -0.009760996326804161, 0.019895298406481743, 0.04601144790649414, 0.0272...
1,032
1,032
['R. Devon Hjelm', 'Sergey M. Plis', 'Vince C. Calhoun']
1603.06624v1
Independent component analysis (ICA), as an approach to the blind source-separation (BSS) problem, has become the de-facto standard in many medical imaging settings. Despite successes and a large ongoing research effort, the limitation of ICA to square linear transformations have not been overcome, so that general INFO...
Variational Autoencoders for Feature Detection of Magnetic Resonance Imaging Data
2,016
http://arxiv.org/pdf/1603.06624v1
Title Variational Autoencoders Feature Detection Magnetic Resonance Imaging Data Summary Independent component analysis ICA approach blind sourceseparation BSS problem become defacto standard many medical imaging setting Despite success large ongoing research effort limitation ICA square linear transformation overcome ...
[-0.029280539602041245, 0.06551916152238846, -0.03624175861477852, 0.03891884163022041, 0.01904877834022045, 0.03921347111463547, 0.034090276807546616, 0.02394743822515011, -0.0011691383551806211, 0.05344498157501221, 0.019847573712468147, -0.019907116889953613, 0.007700404617935419, 0.07628156989812851, 0.051622044295...
1,033
1,033
['Vincent Dumoulin', 'Francesco Visin']
1603.07285v2
We introduce a guide to help deep learning practitioners understand and manipulate convolutional neural network architectures. The guide clarifies the relationship between various properties (input shape, kernel shape, zero padding, strides and output shape) of convolutional, pooling and transposed convolutional layers...
A guide to convolution arithmetic for deep learning
2,016
http://arxiv.org/pdf/1603.07285v2
Title guide convolution arithmetic deep learning Summary introduce guide help deep learning practitioner understand manipulate convolutional neural network architecture guide clarifies relationship various property input shape kernel shape zero padding stride output shape convolutional pooling transposed convolutional ...
[-0.027366986498236656, -0.025669442489743233, -0.02870669960975647, 0.07416903227567673, -0.020394761115312576, -0.02669597417116165, 0.06578706949949265, -0.028624963015317917, -0.06529634445905685, 0.006211624015122652, -0.0026299890596419573, -0.032536134123802185, 0.030836543068289757, 0.069731704890728, 0.0371372...
1,034
1,034
['Tayfun Gokmen', 'Yurii Vlasov']
1603.07341v1
In recent years, deep neural networks (DNN) have demonstrated significant business impact in large scale analysis and classification tasks such as speech recognition, visual object detection, pattern extraction, etc. Training of large DNNs, however, is universally considered as time consuming and computationally intens...
Acceleration of Deep Neural Network Training with Resistive Cross-Point Devices
2,016
http://arxiv.org/pdf/1603.07341v1
Title Acceleration Deep Neural Network Training Resistive CrossPoint Devices Summary recent year deep neural network DNN demonstrated significant business impact large scale analysis classification task speech recognition visual object detection pattern extraction etc Training large DNNs however universally considered ...
[-0.012012002989649773, 0.010204619728028774, -0.048822712153196335, 0.07510970532894135, -0.018761029466986656, -0.014101080596446991, 0.08337471634149551, -0.021849609911441803, 0.02988363802433014, -0.004502861760556698, -0.009541725739836693, 0.008387312293052673, 0.005764232482761145, 0.02505887672305107, 0.009183...
1,035
1,035
['Benjamin Migliori', 'Riley Zeller-Townson', 'Daniel Grady', 'Daniel Gebhardt']
1605.05239v1
Automatic modulation classification (AMC) is an important task for modern communication systems; however, it is a challenging problem when signal features and precise models for generating each modulation may be unknown. We present a new biologically-inspired AMC method without the need for models or manually specified...
Biologically Inspired Radio Signal Feature Extraction with Sparse Denoising Autoencoders
2,016
http://arxiv.org/pdf/1605.05239v1
Title Biologically Inspired Radio Signal Feature Extraction Sparse Denoising Autoencoders Summary Automatic modulation classification AMC important task modern communication system however challenging problem signal feature precise model generating modulation may unknown present new biologicallyinspired AMC method with...
[0.004506845027208328, 0.04426071047782898, -0.012995606288313866, -0.035241685807704926, 0.0505380742251873, 0.03157045319676399, 0.04344221577048302, 0.00989223551005125, -0.09838833659887314, -0.03230344504117966, -0.00015275782789103687, 0.049426715821027756, 0.003318492090329528, 0.04024110361933708, 0.01992030814...
1,036
1,036
['Simone Scardapane', 'Michele Scarpiniti', 'Danilo Comminiello', 'Aurelio Uncini']
1605.05509v2
Neural networks require a careful design in order to perform properly on a given task. In particular, selecting a good activation function (possibly in a data-dependent fashion) is a crucial step, which remains an open problem in the research community. Despite a large amount of investigations, most current implementat...
Learning activation functions from data using cubic spline interpolation
2,016
http://arxiv.org/pdf/1605.05509v2
Title Learning activation function data using cubic spline interpolation Summary Neural network require careful design order perform properly given task particular selecting good activation function possibly datadependent fashion crucial step remains open problem research community Despite large amount investigation cu...
[-0.010667058639228344, 0.031108738854527473, 0.0010392069816589355, 0.03380172327160835, 0.024867622181773186, -0.029170237481594086, 0.017008671537041664, -0.013125554658472538, -0.007886748760938644, 0.038197822868824005, -0.03163711726665497, 0.014261537231504917, 0.0019798658322542906, 0.07050881534814835, -0.0134...
1,037
1,037
['Devansh Arpit', 'Yingbo Zhou', 'Hung Q. Ngo', 'Nils Napp', 'Venu Govindaraju']
1605.07145v2
Auto-Encoders are unsupervised models that aim to learn patterns from observed data by minimizing a reconstruction cost. The useful representations learned are often found to be sparse and distributed. On the other hand, compressed sensing and sparse coding assume a data generating process, where the observed data is g...
On Optimality Conditions for Auto-Encoder Signal Recovery
2,016
http://arxiv.org/pdf/1605.07145v2
Title Optimality Conditions AutoEncoder Signal Recovery Summary AutoEncoders unsupervised model aim learn pattern observed data minimizing reconstruction cost useful representation learned often found sparse distributed hand compressed sensing sparse coding assume data generating process observed data generated true la...
[-0.03757653012871742, 0.07418263703584671, -0.008229772560298443, 0.056047700345516205, -0.007916820235550404, 0.009706190787255764, 0.031618766486644745, 0.07330065220594406, -0.06116146221756935, -0.021794728934764862, 0.014838628470897675, 0.020254770293831825, 0.01566500775516033, 0.07702795416116714, 0.0302096884...
1,038
1,038
['Zhengping Che', 'Sanjay Purushotham', 'Kyunghyun Cho', 'David Sontag', 'Yan Liu']
1606.01865v2
Multivariate time series data in practical applications, such as health care, geoscience, and biology, are characterized by a variety of missing values. In time series prediction and other related tasks, it has been noted that missing values and their missing patterns are often correlated with the target labels, a.k.a....
Recurrent Neural Networks for Multivariate Time Series with Missing Values
2,016
http://arxiv.org/pdf/1606.01865v2
Title Recurrent Neural Networks Multivariate Time Series Missing Values Summary Multivariate time series data practical application health care geoscience biology characterized variety missing value time series prediction related task noted missing value missing pattern often correlated target label aka informative mis...
[-0.012370516546070576, 0.05453154817223549, -0.006999610923230648, -0.03447144106030464, 0.003401670604944229, 0.006981666199862957, 0.03035300225019455, -0.00612240843474865, -0.006608666852116585, 0.05855646729469299, 0.07647636532783508, -0.04249528422951698, 0.03462782874703407, 0.07694125175476074, -0.00040219174...
1,039
1,039
['Zachary C. Lipton', 'David C. Kale', 'Randall Wetzel']
1606.04130v5
We demonstrate a simple strategy to cope with missing data in sequential inputs, addressing the task of multilabel classification of diagnoses given clinical time series. Collected from the pediatric intensive care unit (PICU) at Children's Hospital Los Angeles, our data consists of multivariate time series of observat...
Modeling Missing Data in Clinical Time Series with RNNs
2,016
http://arxiv.org/pdf/1606.04130v5
Title Modeling Missing Data Clinical Time Series RNNs Summary demonstrate simple strategy cope missing data sequential input addressing task multilabel classification diagnosis given clinical time series Collected pediatric intensive care unit PICU Childrens Hospital Los Angeles data consists multivariate time series o...
[0.011131436564028263, 0.05854105204343796, -0.011960997246205807, -0.03848806023597717, -0.01573006622493267, 0.026324205100536346, 0.02152855694293976, -0.0020901032257825136, -0.014020009897649288, 0.018067780882120132, 0.08838817477226257, -0.06427156925201416, 0.024277424439787865, 0.04747456684708595, -0.00392925...
1,040
1,040
['Stefan Hosein', 'Patrick Hosein']
1606.05018v1
Recently there has been significant research on power generation, distribution and transmission efficiency especially in the case of renewable resources. The main objective is reduction of energy losses and this requires improvements on data acquisition and analysis. In this paper we address these concerns by using con...
Improving Power Generation Efficiency using Deep Neural Networks
2,016
http://arxiv.org/pdf/1606.05018v1
Title Improving Power Generation Efficiency using Deep Neural Networks Summary Recently significant research power generation distribution transmission efficiency especially case renewable resource main objective reduction energy loss requires improvement data acquisition analysis paper address concern using consumer e...
[-0.03407803550362587, 0.05291131138801575, -0.015052832663059235, 0.024366991594433784, 0.05610652267932892, -0.001329451915808022, 0.03773383051156998, -0.03784646838903427, -0.019566189497709274, 0.03193863853812218, -0.003670220961794257, 0.035747263580560684, 0.013334088027477264, 0.05060300603508949, 0.0537500642...
1,041
1,041
['Trang Pham', 'Truyen Tran', 'Dinh Phung', 'Svetha Venkatesh']
1608.03639v1
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...
Faster Training of Very Deep Networks Via p-Norm Gates
2,016
http://arxiv.org/pdf/1608.03639v1
Title Faster Training Deep Networks Via pNorm Gates Summary major contributing factor recent advance deep neural network structural unit let sensory information gradient propagate easily Gating one structure act flow control Gates employed many recent stateoftheart recurrent model LSTM GRU feedforward model Residual Ne...
[-0.038673896342515945, 0.07525280117988586, -0.027758315205574036, 0.06350535899400711, -0.038745224475860596, -0.005100186448544264, 0.02834760770201683, 0.016896206885576248, -0.017322776839137077, 0.024394335225224495, -0.02355293184518814, -0.055701106786727905, 0.01647026464343071, 0.10443779826164246, 0.01980346...
1,042
1,042
['Zachary C. Lipton', 'Xiujun Li', 'Jianfeng Gao', 'Lihong Li', 'Faisal Ahmed', 'Li Deng']
1608.05081v4
We present a new algorithm that significantly improves the efficiency of exploration for deep Q-learning agents in dialogue systems. Our agents explore via Thompson sampling, drawing Monte Carlo samples from a Bayes-by-Backprop neural network. Our algorithm learns much faster than common exploration strategies such as ...
BBQ-Networks: Efficient Exploration in Deep Reinforcement Learning for Task-Oriented Dialogue Systems
2,016
http://arxiv.org/pdf/1608.05081v4
Title BBQNetworks Efficient Exploration Deep Reinforcement Learning TaskOriented Dialogue Systems Summary present new algorithm significantly improves efficiency exploration deep Qlearning agent dialogue system agent explore via Thompson sampling drawing Monte Carlo sample BayesbyBackprop neural network algorithm learn...
[0.01756395399570465, -0.002453365596011281, -0.006325583439320326, 0.007751567754894495, 0.0036683231592178345, 0.004726892337203026, -0.002970527857542038, -0.003614576067775488, -0.025236308574676514, -0.012071250006556511, -0.06712344288825989, 0.004469552543014288, -0.030904719606041908, 0.08419699221849442, -0.01...
1,043
1,043
['Henry W. Lin', 'Max Tegmark', 'David Rolnick']
1608.08225v4
We show how the success of deep learning could depend not only on mathematics but also on physics: although well-known mathematical theorems guarantee that neural networks can approximate arbitrary functions well, the class of functions of practical interest can frequently be approximated through "cheap learning" with ...
Why does deep and cheap learning work so well?
2,016
http://arxiv.org/pdf/1608.08225v4
Title deep cheap learning work well Summary show success deep learning could depend mathematics also physic although wellknown mathematical theorem guarantee neural network approximate arbitrary function well class function practical interest frequently approximated cheap learning exponentially fewer parameter generic ...
[-0.05316087603569031, 0.05271539092063904, -0.0023642354644834995, 0.043239373713731766, -0.0020512137562036514, -0.029747430235147476, 0.07406356185674667, -0.0006282221875153482, -0.03803754597902298, 0.008790702559053898, -0.02639138698577881, 0.015553015284240246, 0.013995025306940079, 0.04369424656033516, 0.03127...
1,044
1,044
['Ohad Shamir']
1609.01037v2
Although neural networks are routinely and successfully trained in practice using simple gradient-based methods, most existing theoretical results are negative, showing that learning such networks is difficult, in a worst-case sense over all data distributions. In this paper, we take a more nuanced view, and consider w...
Distribution-Specific Hardness of Learning Neural Networks
2,016
http://arxiv.org/pdf/1609.01037v2
Title DistributionSpecific Hardness Learning Neural Networks Summary Although neural network routinely successfully trained practice using simple gradientbased method existing theoretical result negative showing learning network difficult worstcase sense data distribution paper take nuanced view consider whether specif...
[-0.01701528951525688, 0.05463641881942749, -0.03403698652982712, 0.045400600880384445, -0.03130735456943512, -0.025421524420380592, 0.021013692021369934, -0.018950212746858597, -0.030385931953787804, 0.0040805102325975895, 0.03568645194172859, 0.013218891806900501, 0.03322708234190941, 0.013763370923697948, 0.02486516...
1,045
1,045
['Tom White']
1609.04468v3
We introduce several techniques for sampling and visualizing the latent spaces of generative models. Replacing linear interpolation with spherical linear interpolation prevents diverging from a model's prior distribution and produces sharper samples. J-Diagrams and MINE grids are introduced as visualizations of manifol...
Sampling Generative Networks
2,016
http://arxiv.org/pdf/1609.04468v3
Title Sampling Generative Networks Summary introduce several technique sampling visualizing latent space generative model Replacing linear interpolation spherical linear interpolation prevents diverging model prior distribution produce sharper sample JDiagrams MINE grid introduced visualization manifold created analogy...
[-0.013721062801778316, 0.04511566832661629, -0.04850391298532486, 0.012499610893428326, -0.023033253848552704, -0.008584017865359783, 0.054385073482990265, -0.02627229504287243, -0.04574090987443924, 0.021454133093357086, -0.0011326386593282223, 0.037688031792640686, 0.00016109456191770732, 0.043710559606552124, 0.051...
1,046
1,046
['Alexandre de Brébisson', 'Pascal Vincent']
1609.05866v1
The softmax content-based attention mechanism has proven to be very beneficial in many applications of recurrent neural networks. Nevertheless it suffers from two major computational limitations. First, its computations for an attention lookup scale linearly in the size of the attended sequence. Second, it does not enc...
A Cheap Linear Attention Mechanism with Fast Lookups and Fixed-Size Representations
2,016
http://arxiv.org/pdf/1609.05866v1
Title Cheap Linear Attention Mechanism Fast Lookups FixedSize Representations Summary softmax contentbased attention mechanism proven beneficial many application recurrent neural network Nevertheless suffers two major computational limitation First computation attention lookup scale linearly size attended sequence Seco...
[0.019965214654803276, 0.017598077654838562, -0.005461378023028374, 0.030027788132429123, -0.022849038243293762, -0.011782084591686726, 0.025866692885756493, -0.018963102251291275, 0.0034295115619897842, -0.06577866524457932, 0.028971312567591667, -0.013260751962661743, -0.02671240083873272, 0.07389020919799805, 0.0232...
1,047
1,047
['Sakyasingha Dasgupta', 'Takayuki Yoshizumi', 'Takayuki Osogami']
1610.01989v1
We introduce Delay Pruning, a simple yet powerful technique to regularize dynamic Boltzmann machines (DyBM). The recently introduced DyBM provides a particularly structured Boltzmann machine, as a generative model of a multi-dimensional time-series. This Boltzmann machine can have infinitely many layers of units but al...
Regularized Dynamic Boltzmann Machine with Delay Pruning for Unsupervised Learning of Temporal Sequences
2,016
http://arxiv.org/pdf/1610.01989v1
Title Regularized Dynamic Boltzmann Machine Delay Pruning Unsupervised Learning Temporal Sequences Summary introduce Delay Pruning simple yet powerful technique regularize dynamic Boltzmann machine DyBM recently introduced DyBM provides particularly structured Boltzmann machine generative model multidimensional timeser...
[-0.034820955246686935, 0.015977049246430397, -0.04693754389882088, 0.009623349644243717, -0.010128208436071873, -0.011577432975172997, 0.04971451312303543, -0.022431986406445503, -0.04560653492808342, 0.04773739352822304, 0.023729410022497177, 0.004891319666057825, -0.007737250532954931, 0.06745558977127075, 0.0119249...
1,048
1,048
['Or Sharir', 'Ronen Tamari', 'Nadav Cohen', 'Amnon Shashua']
1610.04167v4
Casting neural networks in generative frameworks is a highly sought-after endeavor these days. Existing methods, such as Generative Adversarial Networks, capture some of the generative capabilities, but not all. To truly leverage the power of generative models, tractable marginalization is needed, a feature outside the...
Tractable Generative Convolutional Arithmetic Circuits
2,016
http://arxiv.org/pdf/1610.04167v4
Title Tractable Generative Convolutional Arithmetic Circuits Summary Casting neural network generative framework highly soughtafter endeavor day Existing method Generative Adversarial Networks capture generative capability truly leverage power generative model tractable marginalization needed feature outside realm curr...
[-0.026159675791859627, 0.1338084191083908, 0.0044853538274765015, 0.04394255578517914, 0.0053377398289740086, -0.01644374616444111, 0.0651826485991478, 0.004866212606430054, -0.07171468436717987, -0.007775270380079746, -0.02130560204386711, -0.014915110543370247, 0.03056446462869644, 0.020416980609297752, 0.0446075983...
1,049
1,049
['Jimmy Ba', 'Geoffrey Hinton', 'Volodymyr Mnih', 'Joel Z. Leibo', 'Catalin Ionescu']
1610.06258v3
Until recently, research on artificial neural networks was largely restricted to systems with only two types of variable: Neural activities that represent the current or recent input and weights that learn to capture regularities among inputs, outputs and payoffs. There is no good reason for this restriction. Synapses ...
Using Fast Weights to Attend to the Recent Past
2,016
http://arxiv.org/pdf/1610.06258v3
Title Using Fast Weights Attend Recent Past Summary recently research artificial neural network largely restricted system two type variable Neural activity represent current recent input weight learn capture regularity among input output payoff good reason restriction Synapses dynamic many different timescales suggests...
[0.0042250556871294975, 0.06523290276527405, -0.018991587683558464, -0.012209667824208736, -0.025263655930757523, -0.02415217086672783, 0.0346950925886631, 0.022041847929358482, -0.008921664208173752, -0.008219674229621887, 0.014070840552449226, -0.05241796746850014, -0.004168669693171978, 0.07041047513484955, 0.032904...
1,050
1,050
['Yevgeniy Bodyanskiy', 'Olena Vynokurova', 'Volodymyr Savvo', 'Tatiana Tverdokhlib', 'Pavlo Mulesa']
1610.07857v1
The hybrid clustering-classification neural network is proposed. This network allows increasing a quality of information processing under the condition of overlapping classes due to the rational choice of a learning rate parameter and introducing a special procedure of fuzzy reasoning in the clustering process, which o...
Hybrid clustering-classification neural network in the medical diagnostics of reactive arthritis
2,016
http://arxiv.org/pdf/1610.07857v1
Title Hybrid clusteringclassification neural network medical diagnostics reactive arthritis Summary hybrid clusteringclassification neural network proposed network allows increasing quality information processing condition overlapping class due rational choice learning rate parameter introducing special procedure fuzzy...
[-0.04642784222960472, -0.02085297740995884, -0.032319098711013794, 0.006815905682742596, 0.020100697875022888, 0.018149347975850105, 0.05504541099071503, 0.0762290358543396, 0.08380621671676636, 0.014367308467626572, 0.03662105277180672, 0.01445157639682293, 0.0012496589915826917, 0.01340700313448906, 0.01372126117348...
1,051
1,051
['Scott Wisdom', 'Thomas Powers', 'John R. Hershey', 'Jonathan Le Roux', 'Les Atlas']
1611.00035v1
Recurrent neural networks are powerful models for processing sequential data, but they are generally plagued by vanishing and exploding gradient problems. Unitary recurrent neural networks (uRNNs), which use unitary recurrence matrices, have recently been proposed as a means to avoid these issues. However, in previous ...
Full-Capacity Unitary Recurrent Neural Networks
2,016
http://arxiv.org/pdf/1611.00035v1
Title FullCapacity Unitary Recurrent Neural Networks Summary Recurrent neural network powerful model processing sequential data generally plagued vanishing exploding gradient problem Unitary recurrent neural network uRNNs use unitary recurrence matrix recently proposed mean avoid issue However previous experiment recur...
[-0.030734920874238014, 0.052602607756853104, 0.0006041592569090426, 0.03892534598708153, 0.026264594867825508, 0.00020796956960111856, 0.04626817628741264, -0.051186393946409225, -0.006274225190281868, 0.007698203902691603, -0.000739139795769006, -0.05875573307275772, -0.016060475260019302, 0.02370956540107727, 0.0273...
1,052
1,052
['Sihan Li', 'Jiantao Jiao', 'Yanjun Han', 'Tsachy Weissman']
1611.01186v2
The Residual Network (ResNet), proposed in He et al. (2015), utilized shortcut connections to significantly reduce the difficulty of training, which resulted in great performance boosts in terms of both training and generalization error. It was empirically observed in He et al. (2015) that stacking more layers of res...
Demystifying ResNet
2,016
http://arxiv.org/pdf/1611.01186v2
Title Demystifying ResNet Summary Residual Network ResNet proposed et al 2015 utilized shortcut connection significantly reduce difficulty training resulted great performance boost term training generalization error empirically observed et al 2015 stacking layer residual block shortcut 2 result smaller training error t...
[-0.028482018038630486, 0.014436997473239899, -0.010756446979939938, 0.045224107801914215, 0.03620850667357445, -0.004042657557874918, -0.00974203273653984, -0.022486062720417976, -0.07816368341445923, 0.0018403935246169567, -8.067340240813792e-05, 0.03390999510884285, 0.01682714931666851, -0.0018463052110746503, 0.009...
1,053
1,053
['Zachary C. Lipton', 'Kamyar Azizzadenesheli', 'Abhishek Kumar', 'Lihong Li', 'Jianfeng Gao', 'Li Deng']
1611.01211v8
Many practical environments contain catastrophic states that an optimal agent would visit infrequently or never. Even on toy problems, Deep Reinforcement Learning (DRL) agents tend to periodically revisit these states upon forgetting their existence under a new policy. We introduce intrinsic fear (IF), a learned reward...
Combating Reinforcement Learning's Sisyphean Curse with Intrinsic Fear
2,016
http://arxiv.org/pdf/1611.01211v8
Title Combating Reinforcement Learnings Sisyphean Curse Intrinsic Fear Summary Many practical environment contain catastrophic state optimal agent would visit infrequently never Even toy problem Deep Reinforcement Learning DRL agent tend periodically revisit state upon forgetting existence new policy introduce intrinsi...
[-0.019792810082435608, 0.022421997040510178, -0.023461276665329933, -0.05199126899242401, 0.02416764572262764, 0.03155511990189552, -0.052378132939338684, -0.0004106828710064292, -0.05425058305263519, 0.04195147752761841, 0.00558619387447834, 0.04211270436644554, -0.05505409091711044, 0.03129200637340546, 0.0038214386...
1,054
1,054
['Juan Maroñas Molano', 'Alberto Albiol Colomer', 'Roberto Paredes Palacios']
1611.02320v1
The use of unsupervised data in addition to supervised data in training discriminative neural networks has improved the performance of this clas- sification scheme. However, the best results were achieved with a training process that is divided in two parts: first an unsupervised pre-training step is done for initializ...
Adversarial Ladder Networks
2,016
http://arxiv.org/pdf/1611.02320v1
Title Adversarial Ladder Networks Summary use unsupervised data addition supervised data training discriminative neural network improved performance clas sification scheme However best result achieved training process divided two part first unsupervised pretraining step done initializing weight network weight refined u...
[-0.006075280252844095, 0.06084270402789116, -0.015357154421508312, 0.025425808504223824, 0.028498176485300064, -0.026468999683856964, 0.03860553726553917, -0.031894292682409286, 0.05148080363869667, 0.03106657974421978, -0.027396125718951225, 0.021526359021663666, 0.009733046405017376, 0.017647814005613327, 0.04040824...
1,055
1,055
['David Balduzzi', 'Brian McWilliams', 'Tony Butler-Yeoman']
1611.02345v2
Modern convolutional networks, incorporating rectifiers and max-pooling, are neither smooth nor convex. Standard guarantees therefore do not apply. Nevertheless, methods from convex optimization such as gradient descent and Adam are widely used as building blocks for deep learning algorithms. This paper provides the fi...
Neural Taylor Approximations: Convergence and Exploration in Rectifier Networks
2,016
http://arxiv.org/pdf/1611.02345v2
Title Neural Taylor Approximations Convergence Exploration Rectifier Networks Summary Modern convolutional network incorporating rectifier maxpooling neither smooth convex Standard guarantee therefore apply Nevertheless method convex optimization gradient descent Adam widely used building block deep learning algorithm ...
[-0.03941693156957626, 0.0035694059915840626, -0.010908045805990696, 0.044448450207710266, 0.006061987020075321, -0.04058769345283508, -0.00604519248008728, 0.01572156697511673, -0.035835932940244675, 0.012853969819843769, -0.0072275856509804726, -0.05246378853917122, 0.04613615572452545, 0.05757799744606018, 0.0047241...
1,056
1,056
['Nat Dilokthanakul', 'Pedro A. M. Mediano', 'Marta Garnelo', 'Matthew C. H. Lee', 'Hugh Salimbeni', 'Kai Arulkumaran', 'Murray Shanahan']
1611.02648v2
We study a variant of the variational autoencoder model (VAE) with a Gaussian mixture as a prior distribution, with the goal of performing unsupervised clustering through deep generative models. We observe that the known problem of over-regularisation that has been shown to arise in regular VAEs also manifests itself i...
Deep Unsupervised Clustering with Gaussian Mixture Variational Autoencoders
2,016
http://arxiv.org/pdf/1611.02648v2
Title Deep Unsupervised Clustering Gaussian Mixture Variational Autoencoders Summary study variant variational autoencoder model VAE Gaussian mixture prior distribution goal performing unsupervised clustering deep generative model observe known problem overregularisation shown arise regular VAEs also manifest model lea...
[-0.034761372953653336, 0.04398951306939125, -0.04592457786202431, 0.006551637779921293, 0.0191801805049181, 0.0023211322259157896, 0.02561366930603981, -0.01467252429574728, -0.04048890992999077, 0.046030595898628235, 0.01781448721885681, 0.015582984313368797, 0.007907512597739697, 0.0870896503329277, 0.03281857818365...
1,057
1,057
['Moritz Hardt', 'Tengyu Ma']
1611.04231v2
An emerging design principle in deep learning is that each layer of a deep artificial neural network should be able to easily express the identity transformation. This idea not only motivated various normalization techniques, such as \emph{batch normalization}, but was also key to the immense success of \emph{residual ...
Identity Matters in Deep Learning
2,016
http://arxiv.org/pdf/1611.04231v2
Title Identity Matters Deep Learning Summary emerging design principle deep learning layer deep artificial neural network able easily express identity transformation idea motivated various normalization technique emphbatch normalization also key immense success emphresidual network work put principle emphidentity param...
[0.005022880155593157, 0.06987550854682922, -0.00843107420951128, 0.03470754995942116, 0.019836002960801125, -0.018270617350935936, 0.08320346474647522, -0.014904382638633251, -0.03487559035420418, 0.0009235618636012077, -0.009102459996938705, 0.005694924388080835, 0.014065858907997608, 0.01970188319683075, 0.029320327...
1,058
1,058
['Siamak Ravanbakhsh', 'Jeff Schneider', 'Barnabas Poczos']
1611.04500v3
We introduce a simple permutation equivariant layer for deep learning with set structure.This type of layer, obtained by parameter-sharing, has a simple implementation and linear-time complexity in the size of each set. We use deep permutation-invariant networks to perform point-could classification and MNIST-digit sum...
Deep Learning with Sets and Point Clouds
2,016
http://arxiv.org/pdf/1611.04500v3
Title Deep Learning Sets Point Clouds Summary introduce simple permutation equivariant layer deep learning set structureThis type layer obtained parametersharing simple implementation lineartime complexity size set use deep permutationinvariant network perform pointcould classification MNISTdigit summation case output ...
[-0.08200529962778091, -0.003272661939263344, -0.033327195793390274, 0.02612738125026226, -0.018871000036597252, 0.013011643663048744, 0.055305685847997665, -0.04713607206940651, -0.008249286562204361, 0.034251317381858826, 0.012211918830871582, 0.01897413656115532, 0.0017339433543384075, 0.06297551840543747, 0.0358491...
1,059
1,059
['Yotaro Kubo', 'George Tucker', 'Simon Wiesler']
1611.06148v2
We introduce dropout compaction, a novel method for training feed-forward neural networks which realizes the performance gains of training a large model with dropout regularization, yet extracts a compact neural network for run-time efficiency. In the proposed method, we introduce a sparsity-inducing prior on the per u...
Compacting Neural Network Classifiers via Dropout Training
2,016
http://arxiv.org/pdf/1611.06148v2
Title Compacting Neural Network Classifiers via Dropout Training Summary introduce dropout compaction novel method training feedforward neural network realizes performance gain training large model dropout regularization yet extract compact neural network runtime efficiency proposed method introduce sparsityinducing pr...
[-0.031969599425792694, 0.05555571988224983, -0.005156401544809341, 0.007052110508084297, -0.00039117818232625723, 0.0005590925575233996, 0.07456076890230179, 0.0061189280822873116, -0.03168214112520218, 0.03842666745185852, -0.03931460529565811, 0.042380694299936295, 0.06471101194620132, 0.060442499816417694, -0.00516...
1,060
1,060
['Anders Søgaard']
1611.06245v1
We present a confidence-based single-layer feed-forward learning algorithm SPIRAL (Spike Regularized Adaptive Learning) relying on an encoding of activation spikes. We adaptively update a weight vector relying on confidence estimates and activation offsets relative to previous activity. We regularize updates proportion...
Spikes as regularizers
2,016
http://arxiv.org/pdf/1611.06245v1
Title Spikes regularizers Summary present confidencebased singlelayer feedforward learning algorithm SPIRAL Spike Regularized Adaptive Learning relying encoding activation spike adaptively update weight vector relying confidence estimate activation offset relative previous activity regularize update proportionally item...
[-0.05139300599694252, -0.023373285308480263, -0.01818792149424553, 0.007478630170226097, 0.028674595057964325, -0.01217301283031702, -0.0011967546306550503, -0.006157814525067806, -0.0016135062323883176, 0.06982237845659256, 0.021045703440904617, 0.03801541402935982, -0.008924521505832672, 0.07269978523254395, 0.04228...
1,061
1,061
['Grzegorz Swirszcz', 'Wojciech Marian Czarnecki', 'Razvan Pascanu']
1611.06310v2
There has been a lot of recent interest in trying to characterize the error surface of deep models. This stems from a long standing question. Given that deep networks are highly nonlinear systems optimized by local gradient methods, why do they not seem to be affected by bad local minima? It is widely believed that tra...
Local minima in training of neural networks
2,016
http://arxiv.org/pdf/1611.06310v2
Title Local minimum training neural network Summary lot recent interest trying characterize error surface deep model stem long standing question Given deep network highly nonlinear system optimized local gradient method seem affected bad local minimum widely believed training deep model using gradient method work well ...
[-0.025040727108716965, 0.018828732892870903, 0.0020180989522486925, 0.03851878643035889, 0.016874512657523155, -0.029997412115335464, 0.048505693674087524, 0.010365145280957222, -0.013248194009065628, 0.04150661826133728, 0.007211761083453894, 0.031743504106998444, 0.0005282129859551787, 0.00604730611667037, 0.0477937...
1,062
1,062
['Zhiguang Wang', 'Weizhong Yan', 'Tim Oates']
1611.06455v4
We propose a simple but strong baseline for time series classification from scratch with deep neural networks. Our proposed baseline models are pure end-to-end without any heavy preprocessing on the raw data or feature crafting. The proposed Fully Convolutional Network (FCN) achieves premium performance to other state-...
Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline
2,016
http://arxiv.org/pdf/1611.06455v4
Title Time Series Classification Scratch Deep Neural Networks Strong Baseline Summary propose simple strong baseline time series classification scratch deep neural network proposed baseline model pure endtoend without heavy preprocessing raw data feature crafting proposed Fully Convolutional Network FCN achieves premiu...
[-0.06151938810944557, 0.044059768319129944, -0.03874833136796951, 0.004693314433097839, -0.017601080238819122, 0.02240394987165928, 0.042283061891794205, -0.01776360347867012, -0.007370558101683855, -0.004207574296742678, 0.04923185333609581, 0.008842306211590767, 0.040825922042131424, 0.08796941488981247, -0.00677602...
1,063
1,063
['Gil Keren', 'Sivan Sabato', 'Björn Schuller']
1611.07743v1
When humans learn a new concept, they might ignore examples that they cannot make sense of at first, and only later focus on such examples, when they are more useful for learning. We propose incorporating this idea of tunable sensitivity for hard examples in neural network learning, using a new generalization of the cr...
Tunable Sensitivity to Large Errors in Neural Network Training
2,016
http://arxiv.org/pdf/1611.07743v1
Title Tunable Sensitivity Large Errors Neural Network Training Summary human learn new concept might ignore example cannot make sense first later focus example useful learning propose incorporating idea tunable sensitivity hard example neural network learning using new generalization crossentropy gradient step used pla...
[0.0001895406749099493, -0.009627725929021835, -0.04425737261772156, 0.03413127735257149, -0.004997651558369398, -0.03961694613099098, 0.02413165010511875, -0.009105011820793152, -0.024802515283226967, -0.007966424338519573, -0.04739246517419815, 0.04947291687130928, 0.04345014691352844, 0.05209972336888313, -0.0052366...
1,064
1,064
['Meshia Cédric Oveneke', 'Mitchel Aliosha-Perez', 'Yong Zhao', 'Dongmei Jiang', 'Hichem Sahli']
1611.09232v1
The omnipresence of deep learning architectures such as deep convolutional neural networks (CNN)s is fueled by the synergistic combination of ever-increasing labeled datasets and specialized hardware. Despite the indisputable success, the reliance on huge amounts of labeled data and specialized hardware can be a limiti...
Efficient Convolutional Auto-Encoding via Random Convexification and Frequency-Domain Minimization
2,016
http://arxiv.org/pdf/1611.09232v1
Title Efficient Convolutional AutoEncoding via Random Convexification FrequencyDomain Minimization Summary omnipresence deep learning architecture deep convolutional neural network CNNs fueled synergistic combination everincreasing labeled datasets specialized hardware Despite indisputable success reliance huge amount ...
[-0.002733794506639242, 0.07162054628133774, -0.006765868980437517, 0.07325270026922226, 0.036439448595047, -0.015761621296405792, 0.021122336387634277, -0.0011417802888900042, -0.07397232949733734, 0.04892192780971527, -0.010136629454791546, -0.0067984964698553085, -0.0021008949261158705, 0.046684496104717255, 0.00093...
1,065
1,065
['Caglar Gulcehre', 'Sarath Chandar', 'Yoshua Bengio']
1701.08718v1
Recent empirical results on long-term dependency tasks have shown that neural networks augmented with an external memory can learn the long-term dependency tasks more easily and achieve better generalization than vanilla recurrent neural networks (RNN). We suggest that memory augmented neural networks can reduce the ef...
Memory Augmented Neural Networks with Wormhole Connections
2,017
http://arxiv.org/pdf/1701.08718v1
Title Memory Augmented Neural Networks Wormhole Connections Summary Recent empirical result longterm dependency task shown neural network augmented external memory learn longterm dependency task easily achieve better generalization vanilla recurrent neural network RNN suggest memory augmented neural network reduce effe...
[0.020384853705763817, 0.043414779007434845, -0.019371645525097847, 0.037320446223020554, -0.029785962775349617, -0.017390504479408264, -0.00362427718937397, -0.021455192938447, 0.0033462168648838997, -0.011203029192984104, -0.021561110392212868, -0.01146629173308611, 0.007640272844582796, 0.05182391032576561, 0.029674...
1,066
1,066
['Andrew Sohn', 'Randal S. Olson', 'Jason H. Moore']
1702.01780v1
Machine learning has been gaining traction in recent years to meet the demand for tools that can efficiently analyze and make sense of the ever-growing databases of biomedical data in health care systems around the world. However, effectively using machine learning methods requires considerable domain expertise, which ...
Toward the automated analysis of complex diseases in genome-wide association studies using genetic programming
2,017
http://arxiv.org/pdf/1702.01780v1
Title Toward automated analysis complex disease genomewide association study using genetic programming Summary Machine learning gaining traction recent year meet demand tool efficiently analyze make sense evergrowing database biomedical data health care system around world However effectively using machine learning met...
[-0.01747054234147072, 0.040198031812906265, -0.04613858461380005, -0.049972712993621826, 0.018341965973377228, 0.049574196338653564, 0.03016352467238903, 0.02071346528828144, 0.048469722270965576, 0.023042185232043266, 0.06908728182315826, 0.0017413890454918146, 0.03552968055009842, 0.0758456364274025, -0.011088464409...
1,067
1,067
['Moshe Looks', 'Marcello Herreshoff', 'DeLesley Hutchins', 'Peter Norvig']
1702.02181v2
Neural networks that compute over graph structures are a natural fit for problems in a variety of domains, including natural language (parse trees) and cheminformatics (molecular graphs). However, since the computation graph has a different shape and size for every input, such networks do not directly support batched t...
Deep Learning with Dynamic Computation Graphs
2,017
http://arxiv.org/pdf/1702.02181v2
Title Deep Learning Dynamic Computation Graphs Summary Neural network compute graph structure natural fit problem variety domain including natural language parse tree cheminformatics molecular graph However since computation graph different shape size every input network directly support batched training inference also...
[-0.008267956785857677, -0.014635797590017319, -0.04825250059366226, 0.005312778986990452, -0.026782169938087463, -0.029162375256419182, -0.00019441884069237858, -0.012170019559562206, 0.03338523954153061, 0.028249187394976616, 0.012643994763493538, 0.00818469375371933, 0.012616272084414959, 0.13440857827663422, 0.0190...
1,068
1,068
['Michael Kampffmeyer', 'Sigurd Løkse', 'Filippo Maria Bianchi', 'Robert Jenssen', 'Lorenzo Livi']
1702.02526v1
In this paper we introduce the deep kernelized autoencoder, a neural network model that allows an explicit approximation of (i) the mapping from an input space to an arbitrary, user-specified kernel space and (ii) the back-projection from such a kernel space to input space. The proposed method is based on traditional a...
Deep Kernelized Autoencoders
2,017
http://arxiv.org/pdf/1702.02526v1
Title Deep Kernelized Autoencoders Summary paper introduce deep kernelized autoencoder neural network model allows explicit approximation mapping input space arbitrary userspecified kernel space ii backprojection kernel space input space proposed method based traditional autoencoders trained new unsupervised loss funct...
[-0.0424896776676178, 0.040260255336761475, -0.009203486144542694, 0.07529452443122864, 0.01783420331776142, -0.02568083629012108, 0.03321047127246857, 0.010553828440606594, -0.045716144144535065, 0.019435163587331772, -0.014130617491900921, 0.0237068273127079, 0.004100973252207041, 0.07678387314081192, 0.0412604734301...
1,069
1,069
['Eric W. Tramel', 'Marylou Gabrié', 'Andre Manoel', 'Francesco Caltagirone', 'Florent Krzakala']
1702.03260v2
Restricted Boltzmann machines (RBMs) are energy-based neural-networks which are commonly used as the building blocks for deep architectures neural architectures. In this work, we derive a deterministic framework for the training, evaluation, and use of RBMs based upon the Thouless-Anderson-Palmer (TAP) mean-field appro...
A Deterministic and Generalized Framework for Unsupervised Learning with Restricted Boltzmann Machines
2,017
http://arxiv.org/pdf/1702.03260v2
Title Deterministic Generalized Framework Unsupervised Learning Restricted Boltzmann Machines Summary Restricted Boltzmann machine RBMs energybased neuralnetworks commonly used building block deep architecture neural architecture work derive deterministic framework training evaluation use RBMs based upon ThoulessAnders...
[-0.04032556712627411, -0.0022950677666813135, -0.042117245495319366, 0.031403448432683945, 0.010676312260329723, 0.009829383343458176, 0.029624799266457558, -0.019647827371954918, -0.01837093010544777, 0.04781458154320717, -0.015097693540155888, 0.017857804894447327, -0.013851631432771683, 0.07567187398672104, 0.02860...
1,070
1,070
['Dianhui Wang', 'Ming Li']
1702.04459v2
Neural networks have been widely used as predictive models to fit data distribution, and they could be implemented through learning a collection of samples. In many applications, however, the given dataset may contain noisy samples or outliers which may result in a poor learner model in terms of generalization. This pa...
Robust Stochastic Configuration Networks with Kernel Density Estimation
2,017
http://arxiv.org/pdf/1702.04459v2
Title Robust Stochastic Configuration Networks Kernel Density Estimation Summary Neural network widely used predictive model fit data distribution could implemented learning collection sample many application however given dataset may contain noisy sample outlier may result poor learner model term generalization paper ...
[-0.02971598505973816, -0.005998740904033184, -0.03633934631943703, 0.008485712110996246, 0.014882173389196396, -0.05329795554280281, -0.0026674116961658, -0.04438336566090584, -0.08371355384588242, 0.035479944199323654, 0.028966840356588364, 0.025127308443188667, 0.037924639880657196, 0.08133848756551743, 0.0146772274...
1,071
1,071
['Mevlana Gemici', 'Chia-Chun Hung', 'Adam Santoro', 'Greg Wayne', 'Shakir Mohamed', 'Danilo J. Rezende', 'David Amos', 'Timothy Lillicrap']
1702.04649v2
We consider the general problem of modeling temporal data with long-range dependencies, wherein new observations are fully or partially predictable based on temporally-distant, past observations. A sufficiently powerful temporal model should separate predictable elements of the sequence from unpredictable elements, exp...
Generative Temporal Models with Memory
2,017
http://arxiv.org/pdf/1702.04649v2
Title Generative Temporal Models Memory Summary consider general problem modeling temporal data longrange dependency wherein new observation fully partially predictable based temporallydistant past observation sufficiently powerful temporal model separate predictable element sequence unpredictable element express uncer...
[0.01757582649588585, 0.0632890909910202, -0.023872578516602516, -0.017164133489131927, -0.03139231353998184, -0.03345639631152153, -0.009081427939236164, 0.004312305711209774, -0.006534235551953316, -0.002540197456255555, 0.07283607870340347, -0.029338544234633446, -0.00865310337394476, 0.0681542456150055, -0.00308079...
1,072
1,072
['Haohan Wang', 'Bhiksha Raj']
1702.07800v4
This paper is a review of the evolutionary history of deep learning models. It covers from the genesis of neural networks when associationism modeling of the brain is studied, to the models that dominate the last decade of research in deep learning like convolutional neural networks, deep belief networks, and recurrent...
On the Origin of Deep Learning
2,017
http://arxiv.org/pdf/1702.07800v4
Title Origin Deep Learning Summary paper review evolutionary history deep learning model cover genesis neural network associationism modeling brain studied model dominate last decade research deep learning like convolutional neural network deep belief network recurrent neural network addition review model paper primari...
[0.022658944129943848, 0.04412216320633888, -0.045865725725889206, -0.013182910159230232, -0.007671241648495197, 0.004469872917979956, 0.0490889810025692, 0.0064223092049360275, -0.051274076104164124, -0.00012445323227439076, 0.009608177468180656, -0.021364765241742134, 0.01883193850517273, 0.0534035786986351, 0.026474...
1,073
1,073
['David Balduzzi', 'Marcus Frean', 'Lennox Leary', 'JP Lewis', 'Kurt Wan-Duo Ma', 'Brian McWilliams']
1702.08591v1
A long-standing obstacle to progress in deep learning is the problem of vanishing and exploding gradients. The problem has largely been overcome through the introduction of carefully constructed initializations and batch normalization. Nevertheless, architectures incorporating skip-connections such as resnets perform m...
The Shattered Gradients Problem: If resnets are the answer, then what is the question?
2,017
http://arxiv.org/pdf/1702.08591v1
Title Shattered Gradients Problem resnets answer question Summary longstanding obstacle progress deep learning problem vanishing exploding gradient problem largely overcome introduction carefully constructed initialization batch normalization Nevertheless architecture incorporating skipconnections resnets perform much ...
[-0.05306291580200195, 0.08462988585233688, 0.003910399042069912, 0.05466694384813309, 0.015272182412445545, -0.02508697472512722, 0.036428868770599365, -0.03013664484024048, -0.06501620262861252, 0.04085691273212433, 0.0043370164930820465, -0.0031224077101796865, 0.021023709326982498, 0.0664471983909607, 0.01241520419...
1,074
1,074
['Sanjeev Arora', 'Rong Ge', 'Yingyu Liang', 'Tengyu Ma', 'Yi Zhang']
1703.00573v5
We show that training of generative adversarial network (GAN) may not have good generalization properties; e.g., training may appear successful but the trained distribution may be far from target distribution in standard metrics. However, generalization does occur for a weaker metric called neural net distance. It is a...
Generalization and Equilibrium in Generative Adversarial Nets (GANs)
2,017
http://arxiv.org/pdf/1703.00573v5
Title Generalization Equilibrium Generative Adversarial Nets GANs Summary show training generative adversarial network GAN may good generalization property eg training may appear successful trained distribution may far target distribution standard metric However generalization occur weaker metric called neural net dist...
[0.0004752659588120878, 0.09058456122875214, -0.02210065722465515, -0.002524065785109997, -0.0011906065046787262, -0.01998322270810604, 0.018323440104722977, 0.012895728461444378, -0.05703878402709961, 0.03898852691054344, -0.003697628853842616, 0.004836162086576223, -0.021246090531349182, 0.0032145746517926455, 0.0695...
1,075
1,075
['Antti Tarvainen', 'Harri Valpola']
1703.01780v5
The recently proposed Temporal Ensembling has achieved state-of-the-art results in several semi-supervised learning benchmarks. It maintains an exponential moving average of label predictions on each training example, and penalizes predictions that are inconsistent with this target. However, because the targets change ...
Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results
2,017
http://arxiv.org/pdf/1703.01780v5
Title Mean teacher better role model Weightaveraged consistency target improve semisupervised deep learning result Summary recently proposed Temporal Ensembling achieved stateoftheart result several semisupervised learning benchmark maintains exponential moving average label prediction training example penalizes predic...
[0.011877510696649551, 0.06664077937602997, -0.023936349898576736, 0.02185123786330223, 0.03708719089627266, -0.013958770781755447, 0.04595377296209335, -0.04740389809012413, -0.0678393617272377, 0.002637114142999053, -0.017652131617069244, -0.02702568843960762, 0.004606575705111027, 0.06541435420513153, 0.007885558530...
1,076
1,076
['Or Sharir', 'Amnon Shashua']
1703.02065v4
Expressive efficiency refers to the relation between two architectures A and B, whereby any function realized by B could be replicated by A, but there exists functions realized by A, which cannot be replicated by B unless its size grows significantly larger. For example, it is known that deep networks are exponentially...
On the Expressive Power of Overlapping Architectures of Deep Learning
2,017
http://arxiv.org/pdf/1703.02065v4
Title Expressive Power Overlapping Architectures Deep Learning Summary Expressive efficiency refers relation two architecture B whereby function realized B could replicated exists function realized cannot replicated B unless size grows significantly larger example known deep network exponentially efficient respect shal...
[-0.028094440698623657, 0.05080948397517204, -0.019131144508719444, 0.07281231880187988, -0.017669323831796646, -0.03467448428273201, 0.038127392530441284, -0.02683134563267231, -0.04156096652150154, 0.01800176128745079, -0.05253916233778, 0.008035933598876, 0.011621705256402493, 0.0717770978808403, 0.02769280970096588...
1,077
1,077
['Olga Wichrowska', 'Niru Maheswaranathan', 'Matthew W. Hoffman', 'Sergio Gomez Colmenarejo', 'Misha Denil', 'Nando de Freitas', 'Jascha Sohl-Dickstein']
1703.04813v4
Learning to learn has emerged as an important direction for achieving artificial intelligence. Two of the primary barriers to its adoption are an inability to scale to larger problems and a limited ability to generalize to new tasks. We introduce a learned gradient descent optimizer that generalizes well to new tasks, ...
Learned Optimizers that Scale and Generalize
2,017
http://arxiv.org/pdf/1703.04813v4
Title Learned Optimizers Scale Generalize Summary Learning learn emerged important direction achieving artificial intelligence Two primary barrier adoption inability scale larger problem limited ability generalize new task introduce learned gradient descent optimizer generalizes well new task significantly reduced memo...
[0.00554680498316884, 0.02597377635538578, -0.018615232780575752, 0.03551197424530983, 0.004616288002580404, -0.007247465662658215, 0.0568012073636055, 0.007626159116625786, -0.08842956274747849, -0.014408163726329803, -0.03654739633202553, 0.056028690189123154, 0.025739187374711037, 0.012816272675991058, -0.0009048736...
1,078
1,078
['Ramin M. Hasani', 'Guodong Wang', 'Radu Grosu']
1703.06272v1
This paper studies an intelligent ultimate technique for health-monitoring and prognostic of common rotary machine components, particularly bearings. During a run-to-failure experiment, rich unsupervised features from vibration sensory data are extracted by a trained sparse auto-encoder. Then, the correlation of the ex...
An Automated Auto-encoder Correlation-based Health-Monitoring and Prognostic Method for Machine Bearings
2,017
http://arxiv.org/pdf/1703.06272v1
Title Automated Autoencoder Correlationbased HealthMonitoring Prognostic Method Machine Bearings Summary paper study intelligent ultimate technique healthmonitoring prognostic common rotary machine component particularly bearing runtofailure experiment rich unsupervised feature vibration sensory data extracted trained ...
[-0.058137550950050354, 0.04175381734967232, -0.016325121745467186, 0.039435975253582, 0.056581854820251465, 0.0016908240504562855, -0.011529015377163887, 0.050247449427843094, -0.06116427481174469, 0.007793418597429991, 0.06878956407308578, 0.05727796256542206, 0.021545078605413437, 0.04878242686390877, -0.00956003926...
1,079
1,079
['William La Cava', 'Jason H. Moore']
1703.06934v3
Recently we proposed a general, ensemble-based feature engineering wrapper (FEW) that was paired with a number of machine learning methods to solve regression problems. Here, we adapt FEW for supervised classification and perform a thorough analysis of fitness and survival methods within this framework. Our tests demon...
Ensemble representation learning: an analysis of fitness and survival for wrapper-based genetic programming methods
2,017
http://arxiv.org/pdf/1703.06934v3
Title Ensemble representation learning analysis fitness survival wrapperbased genetic programming method Summary Recently proposed general ensemblebased feature engineering wrapper paired number machine learning method solve regression problem adapt supervised classification perform thorough analysis fitness survival m...
[-0.003239601617679, 0.025448648259043694, -0.04197628051042557, -0.03391646221280098, 0.0317525789141655, 0.0031980187632143497, 0.013548492453992367, 0.023654265329241753, -0.007564590312540531, -0.003999680746346712, 0.05780627205967903, -0.0185535941272974, 0.017537806183099747, 0.03596587851643562, -0.017810583114...
1,080
1,080
['Shai Shalev-Shwartz', 'Ohad Shamir', 'Shaked Shammah']
1703.07950v2
In recent years, Deep Learning has become the go-to solution for a broad range of applications, often outperforming state-of-the-art. However, it is important, for both theoreticians and practitioners, to gain a deeper understanding of the difficulties and limitations associated with common approaches and algorithms. W...
Failures of Gradient-Based Deep Learning
2,017
http://arxiv.org/pdf/1703.07950v2
Title Failures GradientBased Deep Learning Summary recent year Deep Learning become goto solution broad range application often outperforming stateoftheart However important theoretician practitioner gain deeper understanding difficulty limitation associated common approach algorithm describe four type simple problem g...
[-0.004683353938162327, 0.049813274294137955, -0.009041251614689827, 0.026926519349217415, 0.002914546988904476, -0.011987301521003246, 0.05571373924612999, 0.013933539390563965, -0.014835190027952194, -0.013096559792757034, 0.030153730884194374, -0.012392301112413406, 0.018408115953207016, 0.0473383292555809, 0.010164...
1,081
1,081
['David Ha', 'Douglas Eck']
1704.03477v4
We present sketch-rnn, a recurrent neural network (RNN) able to construct stroke-based drawings of common objects. The model is trained on thousands of crude human-drawn images representing hundreds of classes. We outline a framework for conditional and unconditional sketch generation, and describe new robust training ...
A Neural Representation of Sketch Drawings
2,017
http://arxiv.org/pdf/1704.03477v4
Title Neural Representation Sketch Drawings Summary present sketchrnn recurrent neural network RNN able construct strokebased drawing common object model trained thousand crude humandrawn image representing hundred class outline framework conditional unconditional sketch generation describe new robust training method g...
[0.002072283998131752, 0.001426812494173646, -0.007050015963613987, 0.02355545572936535, -0.01717188209295273, -0.014916708692908287, 0.010072977282106876, -0.006182495038956404, -0.08580172806978226, 0.010229522362351418, 0.006240080110728741, -0.05838477611541748, 0.0029572630301117897, 0.029146021232008934, 0.008644...
1,082
1,082
['Y. Cem Subakan', 'Paris Smaragdis']
1704.05420v2
In this paper, we propose a new Recurrent Neural Network (RNN) architecture. The novelty is simple: We use diagonal recurrent matrices instead of full. This results in better test likelihood and faster convergence compared to regular full RNNs in most of our experiments. We show the benefits of using diagonal recurrent...
Diagonal RNNs in Symbolic Music Modeling
2,017
http://arxiv.org/pdf/1704.05420v2
Title Diagonal RNNs Symbolic Music Modeling Summary paper propose new Recurrent Neural Network RNN architecture novelty simple use diagonal recurrent matrix instead full result better test likelihood faster convergence compared regular full RNNs experiment show benefit using diagonal recurrent matrix popularly used LST...
[-0.010554762557148933, 0.020659727975726128, -0.0253631379455328, 0.060117144137620926, -0.05250813812017441, -0.022121572867035866, 0.04386484995484352, 0.016779771074652672, -0.04991058260202408, 0.013138765469193459, -0.05485602840781212, 0.007838954217731953, 0.033968374133110046, -0.010870550759136677, 0.02885793...
1,083
1,083
['Davide Bacciu', 'Francesco Crecchi', 'Davide Morelli']
1705.02643v1
The paper presents a novel, principled approach to train recurrent neural networks from the Reservoir Computing family that are robust to missing part of the input features at prediction time. By building on the ensembling properties of Dropout regularization, we propose a methodology, named DropIn, which efficiently t...
DropIn: Making Reservoir Computing Neural Networks Robust to Missing Inputs by Dropout
2,017
http://arxiv.org/pdf/1705.02643v1
Title DropIn Making Reservoir Computing Neural Networks Robust Missing Inputs Dropout Summary paper present novel principled approach train recurrent neural network Reservoir Computing family robust missing part input feature prediction time building ensembling property Dropout regularization propose methodology named ...
[-0.022154180333018303, 0.01173417828977108, -0.008322242647409439, -0.005620358511805534, 0.0121510224416852, -0.017016639932990074, 0.001384788192808628, -0.045691341161727905, -0.007102271541953087, 0.004455183167010546, 0.05254826694726944, 0.020588377490639687, 0.02012738212943077, 0.06040531024336815, 0.015332243...
1,084
1,084
['Tayfun Gokmen', 'O. Murat Onen', 'Wilfried Haensch']
1705.08014v1
In a previous work we have detailed the requirements to obtain a maximal performance benefit by implementing fully connected deep neural networks (DNN) in form of arrays of resistive devices for deep learning. This concept of Resistive Processing Unit (RPU) devices we extend here towards convolutional neural networks (...
Training Deep Convolutional Neural Networks with Resistive Cross-Point Devices
2,017
http://arxiv.org/pdf/1705.08014v1
Title Training Deep Convolutional Neural Networks Resistive CrossPoint Devices Summary previous work detailed requirement obtain maximal performance benefit implementing fully connected deep neural network DNN form array resistive device deep learning concept Resistive Processing Unit RPU device extend towards convolut...
[-0.003952740225940943, -0.004424635320901871, -0.03168492391705513, 0.08012476563453674, 0.02246415615081787, -0.03856043890118599, 0.08240853995084763, -0.04155397787690163, 0.028687527403235435, 0.0005882265395484865, -0.05793306976556778, 0.02462966926395893, 0.021634044125676155, 0.0019402351463213563, 0.026766458...
1,085
1,085
['Friedemann Zenke', 'Surya Ganguli']
1705.11146v2
A vast majority of computation in the brain is performed by spiking neural networks. Despite the ubiquity of such spiking, we currently lack an understanding of how biological spiking neural circuits learn and compute in-vivo, as well as how we can instantiate such capabilities in artificial spiking circuits in-silico....
SuperSpike: Supervised learning in multi-layer spiking neural networks
2,017
http://arxiv.org/pdf/1705.11146v2
Title SuperSpike Supervised learning multilayer spiking neural network Summary vast majority computation brain performed spiking neural network Despite ubiquity spiking currently lack understanding biological spiking neural circuit learn compute invivo well instantiate capability artificial spiking circuit insilico rev...
[-0.02741864323616028, -0.054659098386764526, -0.024781418964266777, 0.014508717693388462, 0.05291152000427246, -0.0021210929844528437, 0.01844301074743271, -0.026208391413092613, -0.0036614658311009407, 0.010912706144154072, -0.027361368760466576, 0.0462682880461216, -0.011503471061587334, 0.07688760757446289, 0.04937...
1,086
1,086
['Adepu Ravi Sankar', 'Vineeth N Balasubramanian']
1706.02052v1
Recent years have seen a growing interest in understanding deep neural networks from an optimization perspective. It is understood now that converging to low-cost local minima is sufficient for such models to become effective in practice. However, in this work, we propose a new hypothesis based on recent theoretical fi...
Are Saddles Good Enough for Deep Learning?
2,017
http://arxiv.org/pdf/1706.02052v1
Title Saddles Good Enough Deep Learning Summary Recent year seen growing interest understanding deep neural network optimization perspective understood converging lowcost local minimum sufficient model become effective practice However work propose new hypothesis based recent theoretical finding empirical study deep ne...
[-0.01636373996734619, 0.03703011944890022, -0.008653861470520496, 0.008967648260295391, 0.003082389011979103, -0.0404864139854908, 0.02635808289051056, -0.00867127813398838, -0.03875653073191643, 0.03163686767220497, -0.018598459661006927, -0.02413289062678814, 0.004278802312910557, 0.033016450703144073, 0.03131746873...
1,087
1,087
['Li Jing', 'Caglar Gulcehre', 'John Peurifoy', 'Yichen Shen', 'Max Tegmark', 'Marin Soljačić', 'Yoshua Bengio']
1706.02761v3
We present a novel recurrent neural network (RNN) based model that combines the remembering ability of unitary RNNs with the ability of gated RNNs to effectively forget redundant/irrelevant information in its memory. We achieve this by extending unitary RNNs with a gating mechanism. Our model is able to outperform LSTM...
Gated Orthogonal Recurrent Units: On Learning to Forget
2,017
http://arxiv.org/pdf/1706.02761v3
Title Gated Orthogonal Recurrent Units Learning Forget Summary present novel recurrent neural network RNN based model combine remembering ability unitary RNNs ability gated RNNs effectively forget redundantirrelevant information memory achieve extending unitary RNNs gating mechanism model able outperform LSTMs GRUs Uni...
[-0.007480069994926453, 0.04770495370030403, 0.023244548588991165, 0.015302649699151516, -0.033076025545597076, -0.005800533574074507, -0.00015856995014473796, -0.01573568396270275, -0.018688447773456573, -0.014689544215798378, -0.02922867238521576, -0.04847094789147377, 0.018516138195991516, -0.01492219790816307, 0.00...
1,088
1,088
['Matus Telgarsky']
1706.03301v1
Neural networks and rational functions efficiently approximate each other. In more detail, it is shown here that for any ReLU network, there exists a rational function of degree $O(\text{polylog}(1/\epsilon))$ which is $\epsilon$-close, and similarly for any rational function there exists a ReLU network of size $O(\tex...
Neural networks and rational functions
2,017
http://arxiv.org/pdf/1706.03301v1
Title Neural network rational function Summary Neural network rational function efficiently approximate detail shown ReLU network exists rational function degree Otextpolylog1epsilon epsilonclose similarly rational function exists ReLU network size Otextpolylog1epsilon epsilonclose contrast polynomial need degree Omega...
[-0.07772735506296158, 0.051522862166166306, 0.012671001255512238, 0.06254872679710388, 0.01926160417497158, -0.020529543980956078, 0.009075514040887356, 0.012158327735960484, -0.009596889838576317, -0.02108626812696457, 0.03329392895102501, 0.009260141290724277, 0.006060292944312096, 0.055068351328372955, 0.0224348120...
1,089
1,089
['Timoleon Moraitis', 'Abu Sebastian', 'Irem Boybat', 'Manuel Le Gallo', 'Tomas Tuma', 'Evangelos Eleftheriou']
1706.05563v1
Spiking neural networks (SNNs) could play a key role in unsupervised machine learning applications, by virtue of strengths related to learning from the fine temporal structure of event-based signals. However, some spike-timing-related strengths of SNNs are hindered by the sensitivity of spike-timing-dependent plasticit...
Fatiguing STDP: Learning from Spike-Timing Codes in the Presence of Rate Codes
2,017
http://arxiv.org/pdf/1706.05563v1
Title Fatiguing STDP Learning SpikeTiming Codes Presence Rate Codes Summary Spiking neural network SNNs could play key role unsupervised machine learning application virtue strength related learning fine temporal structure eventbased signal However spiketimingrelated strength SNNs hindered sensitivity spiketimingdepend...
[-0.08710033446550369, -0.05305293947458267, -0.030626080930233, 0.07477153837680817, 0.029601221904158592, -0.015349716879427433, 0.011585241183638573, -0.0382562018930912, 0.027124693617224693, 0.02741517499089241, -0.010895184241235256, 0.03911459818482399, 0.002717064693570137, 0.08500304073095322, 0.01383508555591...
1,090
1,090
['Alfred Bourely', 'John Patrick Boueri', 'Krzysztof Choromonski']
1706.05683v1
We propose Sparse Neural Network architectures that are based on random or structured bipartite graph topologies. Sparse architectures provide compression of the models learned and speed-ups of computations, they can also surpass their unstructured or fully connected counterparts. As we show, even more compact topologi...
Sparse Neural Networks Topologies
2,017
http://arxiv.org/pdf/1706.05683v1
Title Sparse Neural Networks Topologies Summary propose Sparse Neural Network architecture based random structured bipartite graph topology Sparse architecture provide compression model learned speedup computation also surpass unstructured fully connected counterpart show even compact topology socalled SNN Sparse Neura...
[-0.020375169813632965, 0.01855674386024475, -0.01683023013174534, 0.05763872340321541, -0.032457973808050156, -0.07128623127937317, 3.731029573827982e-05, -0.0336407795548439, 0.013529010117053986, 0.023449791595339775, -0.007047354243695736, -0.017873087897896767, 0.01164446584880352, 0.0521371029317379, 0.0239701885...
1,091
1,091
['Aleksander Madry', 'Aleksandar Makelov', 'Ludwig Schmidt', 'Dimitris Tsipras', 'Adrian Vladu']
1706.06083v3
Recent work has demonstrated that neural networks are vulnerable to adversarial examples, i.e., inputs that are almost indistinguishable from natural data and yet classified incorrectly by the network. In fact, some of the latest findings suggest that the existence of adversarial attacks may be an inherent weakness of ...
Towards Deep Learning Models Resistant to Adversarial Attacks
2,017
http://arxiv.org/pdf/1706.06083v3
Title Towards Deep Learning Models Resistant Adversarial Attacks Summary Recent work demonstrated neural network vulnerable adversarial example ie input almost indistinguishable natural data yet classified incorrectly network fact latest finding suggest existence adversarial attack may inherent weakness deep learning m...
[0.014816870912909508, 0.05214658007025719, -0.018002264201641083, 0.04595234990119934, -0.010541584342718124, -0.02386871539056301, 0.04798188805580139, -0.03780365735292435, -0.000369595451047644, -0.026287145912647247, -0.018617894500494003, 0.0027225359808653593, -0.0018221413483843207, 0.04532580077648163, 0.06297...
1,092
1,092
['Peter Bartlett', 'Dylan J. Foster', 'Matus Telgarsky']
1706.08498v2
This paper presents a margin-based multiclass generalization bound for neural networks that scales with their margin-normalized "spectral complexity": their Lipschitz constant, meaning the product of the spectral norms of the weight matrices, times a certain correction factor. This bound is empirically investigated for...
Spectrally-normalized margin bounds for neural networks
2,017
http://arxiv.org/pdf/1706.08498v2
Title Spectrallynormalized margin bound neural network Summary paper present marginbased multiclass generalization bound neural network scale marginnormalized spectral complexity Lipschitz constant meaning product spectral norm weight matrix time certain correction factor bound empirically investigated standard AlexNet...
[-0.04155050963163376, -0.008558375760912895, -0.033014170825481415, -0.004344085231423378, 0.013449542224407196, -0.02681571990251541, 0.03941807523369789, -0.01090234238654375, -0.0007514691678807139, 0.04652928188443184, -0.0039003698620945215, 0.0069671496748924255, -0.025340158492326736, 0.03138614073395729, -0.00...
1,093
1,093
['Miguel Á. Carreira-Perpiñán']
1707.01209v1
Compressing neural nets is an active research problem, given the large size of state-of-the-art nets for tasks such as object recognition, and the computational limits imposed by mobile devices. We give a general formulation of model compression as constrained optimization. This includes many types of compression: quan...
Model compression as constrained optimization, with application to neural nets. Part I: general framework
2,017
http://arxiv.org/pdf/1707.01209v1
Title Model compression constrained optimization application neural net Part general framework Summary Compressing neural net active research problem given large size stateoftheart net task object recognition computational limit imposed mobile device give general formulation model compression constrained optimization i...
[-0.01731552556157112, 0.04729483649134636, 0.0009190910495817661, 0.031873706728219986, 0.0026739088352769613, -0.020321691408753395, 0.021075960248708725, 0.047936223447322845, -0.05037736892700195, -0.003749072551727295, 0.0034149077255278826, 0.016629399731755257, 0.015772700309753418, -0.025452492758631706, 0.0174...
1,094
1,094
['Miguel Á. Carreira-Perpiñán', 'Yerlan Idelbayev']
1707.04319v1
We consider the problem of deep neural net compression by quantization: given a large, reference net, we want to quantize its real-valued weights using a codebook with $K$ entries so that the training loss of the quantized net is minimal. The codebook can be optimally learned jointly with the net, or fixed, as for bina...
Model compression as constrained optimization, with application to neural nets. Part II: quantization
2,017
http://arxiv.org/pdf/1707.04319v1
Title Model compression constrained optimization application neural net Part II quantization Summary consider problem deep neural net compression quantization given large reference net want quantize realvalued weight using codebook K entry training loss quantized net minimal codebook optimally learned jointly net fixed...
[-0.02916509099304676, 0.08302726596593857, -0.0007826504297554493, 0.006499247159808874, 0.012315145693719387, -0.012033206410706043, 0.007497633341699839, 0.0883617177605629, -0.08171931654214859, 0.012326371856033802, 0.0276323314756155, 0.005068651866167784, -0.025546912103891373, 0.018155327066779137, 0.0245215147...
1,095
1,095
['Andrei Turkin']
1708.02979v1
It is a well-known fact that adding noise to the input data often improves network performance. While the dropout technique may be a cause of memory loss, when it is applied to recurrent connections, Tikhonov regularization, which can be regarded as the training with additive noise, avoids this issue naturally, though ...
Tikhonov Regularization for Long Short-Term Memory Networks
2,017
http://arxiv.org/pdf/1708.02979v1
Title Tikhonov Regularization Long ShortTerm Memory Networks Summary wellknown fact adding noise input data often improves network performance dropout technique may cause memory loss applied recurrent connection Tikhonov regularization regarded training additive noise avoids issue naturally though implies regularizer d...
[0.010209091939032078, 0.004556693136692047, 0.01598581112921238, 0.061207231134176254, 0.05045241490006447, -0.023160258308053017, 0.02056543156504631, -0.011972963809967041, -0.021164556965231895, -0.004597627557814121, -0.024111447855830193, -0.004807803779840469, 0.027682339772582054, 0.03633328154683113, 0.0296173...
1,096
1,096
['Klaus Greff', 'Sjoerd van Steenkiste', 'Jürgen Schmidhuber']
1708.03498v2
Many real world tasks such as reasoning and physical interaction require identification and manipulation of conceptual entities. A first step towards solving these tasks is the automated discovery of distributed symbol-like representations. In this paper, we explicitly formalize this problem as inference in a spatial m...
Neural Expectation Maximization
2,017
http://arxiv.org/pdf/1708.03498v2
Title Neural Expectation Maximization Summary Many real world task reasoning physical interaction require identification manipulation conceptual entity first step towards solving task automated discovery distributed symbollike representation paper explicitly formalize problem inference spatial mixture model component p...
[-0.008433626964688301, 0.016918597742915154, -0.006526081822812557, 0.06968424469232559, 0.005588540807366371, -0.013017911463975906, 0.052101980894804, -0.029498402029275894, 0.0013138457434251904, -0.009275843389332294, 0.005174507852643728, 0.025917036458849907, 0.004535451997071505, -0.0054769571870565414, 0.04130...
1,097
1,097
['Ashley Prater']
1708.07147v1
This work introduces a tensor-based method to perform supervised classification on spatiotemporal data processed in an echo state network. Typically when performing supervised classification tasks on data processed in an echo state network, the entire collection of hidden layer node states from the training dataset is ...
Classification via Tensor Decompositions of Echo State Networks
2,017
http://arxiv.org/pdf/1708.07147v1
Title Classification via Tensor Decompositions Echo State Networks Summary work introduces tensorbased method perform supervised classification spatiotemporal data processed echo state network Typically performing supervised classification task data processed echo state network entire collection hidden layer node state...
[-0.05399998649954796, 0.01863001100718975, -0.032564617693424225, 0.04404335096478462, -0.008139513432979584, 0.033061474561691284, 0.08037623763084412, 0.020166991278529167, -0.023139137774705887, -0.021639293059706688, -0.01467151753604412, 0.001805002335458994, 0.04165005311369896, -0.0005517308600246906, -0.007382...
1,098
1,098
['Robin Tibor Schirrmeister', 'Lukas Gemein', 'Katharina Eggensperger', 'Frank Hutter', 'Tonio Ball']
1708.08012v3
We apply convolutional neural networks (ConvNets) to the task of distinguishing pathological from normal EEG recordings in the Temple University Hospital EEG Abnormal Corpus. We use two basic, shallow and deep ConvNet architectures recently shown to decode task-related information from EEG at least as well as establish...
Deep learning with convolutional neural networks for decoding and visualization of EEG pathology
2,017
http://arxiv.org/pdf/1708.08012v3
Title Deep learning convolutional neural network decoding visualization EEG pathology Summary apply convolutional neural network ConvNets task distinguishing pathological normal EEG recording Temple University Hospital EEG Abnormal Corpus use two basic shallow deep ConvNet architecture recently shown decode taskrelated...
[-0.020881209522485733, -0.0007849079556763172, 0.006975267082452774, 0.028392890468239784, 0.042344383895397186, 0.030427228659391403, 0.016356797888875008, 0.046211127191782, -0.045808255672454834, 0.04334313049912453, -0.025741882622241974, 0.013166657648980618, 0.007784973829984665, 0.07731813192367554, 0.019739378...
1,099
1,099
['Peng-Bo Zhang']
1708.08282v2
In school, a teacher plays an important role in various classroom teaching patterns. Likewise to this human learning activity, the learning using privileged information (LUPI) paradigm provides additional information generated by the teacher to 'teach' learning algorithms during the training stage. Therefore, this nove...
A New Learning Paradigm for Random Vector Functional-Link Network: RVFL+
2,017
http://arxiv.org/pdf/1708.08282v2
Title New Learning Paradigm Random Vector FunctionalLink Network RVFL Summary school teacher play important role various classroom teaching pattern Likewise human learning activity learning using privileged information LUPI paradigm provides additional information generated teacher teach learning algorithm training sta...
[0.010176077485084534, -0.03794378042221069, -0.045590806752443314, 0.03265431895852089, 0.020587285980582237, -0.05481868237257004, 0.002195412525907159, -0.029551386833190918, 0.01054745726287365, -0.02192939631640911, -0.0003719379601534456, 0.0010894577717408538, 0.005972177255898714, 0.023160308599472046, 0.023875...