New! Search for patents from more than 100 countries including Australia, Brazil, Sweden and more

WO2014105866A1 - System and method for addressing overfitting in a neural network - Google Patents

System and method for addressing overfitting in a neural network Download PDF

Info

Publication number
WO2014105866A1
WO2014105866A1 PCT/US2013/077608 US2013077608W WO2014105866A1 WO 2014105866 A1 WO2014105866 A1 WO 2014105866A1 US 2013077608 W US2013077608 W US 2013077608W WO 2014105866 A1 WO2014105866 A1 WO 2014105866A1
Authority
WO
WIPO (PCT)
Prior art keywords
training
feature detectors
neural network
layers
feature
Prior art date
Application number
PCT/US2013/077608
Other languages
French (fr)
Inventor
Geoffrey E. HINTON
Alexander KRIZHEVSKY
Ilya SUTSKEVER
Nitish SRIVASTVA
Original Assignee
Google Inc.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority to US201261745711P priority Critical
Priority to US61/745,711 priority
Priority to US14/015,768 priority patent/US9406017B2/en
Priority to US14/015,768 priority
Application filed by Google Inc. filed Critical Google Inc.
Publication of WO2014105866A1 publication Critical patent/WO2014105866A1/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computer systems based on biological models
    • G06N3/02Computer systems based on biological models using neural network models
    • G06N3/08Learning methods
    • G06N3/084Back-propagation
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/36Image preprocessing, i.e. processing the image information without deciding about the identity of the image
    • G06K9/46Extraction of features or characteristics of the image
    • G06K9/4604Detecting partial patterns, e.g. edges or contours, or configurations, e.g. loops, corners, strokes, intersections
    • G06K9/4609Detecting partial patterns, e.g. edges or contours, or configurations, e.g. loops, corners, strokes, intersections by matching or filtering
    • G06K9/4619Biologically-inspired filters, e.g. receptive fields
    • G06K9/4623Biologically-inspired filters, e.g. receptive fields with interaction between the responses of different filters
    • G06K9/4628Integrating the filters into a hierarchical structure
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computer systems based on biological models
    • G06N3/02Computer systems based on biological models using neural network models
    • G06N3/04Architectures, e.g. interconnection topology
    • G06N3/0454Architectures, e.g. interconnection topology using a combination of multiple neural nets
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computer systems based on biological models
    • G06N3/02Computer systems based on biological models using neural network models
    • G06N3/04Architectures, e.g. interconnection topology
    • G06N3/0472Architectures, e.g. interconnection topology using probabilistic elements, e.g. p-rams, stochastic processors
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computer systems based on biological models
    • G06N3/02Computer systems based on biological models using neural network models
    • G06N3/08Learning methods
    • G06N3/082Learning methods modifying the architecture, e.g. adding or deleting nodes or connections, pruning

Abstract

A system for training a neural network. A switch is linked to feature detectors in at least some of the layers of the neural network. For each training case, the switch randomly selectively disables each of the feature detectors in accordance with a preconfigured probability. The weights from each training case are then normalized for applying the neural network to test data.

Description

SYSTEM AND METHOD FOR ADDRESSING OVERFITTING

IN A NEURAL NETWORK

TECHNICAL FIELD

[0001] The following relates generally to neural networks and more specifically to training a neural network.

BACKGROUND

[0002] A feedforward, artificial neural network uses layers of non-linear "hidden" units between its inputs and its outputs. Each unit has a weight that is determined during learning, which can be referred to as a training stage. In the training stage, a training set of data (a training set of inputs each having a known output) is processed by the neural network. Thus, it is intended that the neural network learn how to provide an output for new input data by generalizing the information it learns in the training stage from the training data. Generally, once learning is complete, a validation set is processed by the neural network to validate the results of learning. Finally, test data (i.e., data for which generating an output is desired) can be processed by a validated neural network.

[0003] The purpose of learning is to adapt the weights on the incoming connections of hidden units to learn feature detectors that enable it to predict the correct output when given an input vector. If the relationship between the input and the correct output is complicated and the network has enough hidden units to model it accurately, there will typically be many different settings of the weights that can model the framing set almost perfectly, especially if there is only a limited amount of labeled training data. Each of these weight vectors will make different predictions on held-out test data and almost all of them will do worse on the test data than on the training data because the feature detectors have been tuned to work well together on the training data but not on the test data.

[0004] This occurs because of the overfitting problem, which occurs when the neural network simply memorizes the training data that it is provided, rather than generalizing well to new examples. Generally, the overfitting problem is increasingly likely to occur as the complexity of the neural network increases. [0005] Overfitting can be mitigated by providing the neural network with more training data. However, the collection of training data is a laborious and expensive task.

[0006] One proposed approach to reduce the error on the test set is to average the predictions produced by many separate trained networks and then to apply each of these networks to the test data, but this is computationally expensive during both training and testing.

[0007] It is an object of the following to obviate or mitigate at least one of the foregoing issues.

SUMMARY

[0008] In one aspect, a system for training a neural network is provided, the system comprising a switch linked to a plurality of feature detectors of the neural network, the switch operable to randomly selectively disable each of the plurality of feature detectors for each of a plurality of training cases.

[0009] In another aspect, a method for training a neural network is provided.

DESCRIPTION OF DRAWINGS

[0010] The features of the invention will become more apparent in the following detailed description in which reference is made to the appended drawings wherein:

[0011] Fig. 1 is an architecture diagram of a system for training a neural network; and

[0012] Fig. 2 is a flowchart for training a neural network.

[0013] Like reference symbols in the various drawings indicate like elements.

DETAILED DESCRIPTION

[0014] Embodiments will now be described with reference to the figures. It will be appreciated that for simplicity and clarity of illustration, where considered appropriate, reference numerals may be repeated among the figures to indicate corresponding or analogous elements. In addition, numerous specific details are set forth in order to provide a thorough understanding of the embodiments described herein. However, it will be understood by those of ordinary skill in the art that the embodiments described herein may be practiced without these specific details. In other instances, well-known methods, procedures and components have not been described in detail so as not to obscure the embodiments described herein. Also, the description is not to be considered as limiting the scope of the embodiments described herein.

[0015] It will also be appreciated that any module, unit, component, server, computer, terminal or device exemplified herein that executes instructions may include or otherwise have access to computer readable media such as storage media, computer storage media, or data storage devices (removable and/or non-removable) such as, for example, magnetic disks, optical disks, or tape. Computer storage media may include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information, such as computer readable instructions, data structures, program modules, or other data. Examples of computer storage media include RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by an application, module, or both. Any such computer storage media may be part of the device or accessible or connectable thereto. Any application or module herein described may be implemented using computer readable executable instructions that may be stored or otherwise held by such computer readable media.

[0016] It has been found that overfitting may be reduced by selectively disabling a randomly (or pseudorandomly) selected subset of feature detectors in a neural network during each training case of the training stage, and adapting the weights of each feature detector accordingly during application of the neural network in the test stage. It has been found that the foregoing approach may prevent complex co-adaptations between feature detectors, for example where a particular feature detector is only helpful in the context of several other specific feature detectors. Although it is preferred that the selective disabling of feature detectors be changed for each training case, it is

contemplated herein that disabling of particular feature detectors may be held constant for a plurality of training cases.

[0017] Referring now to Fig. 1 , a feedforward neural network (100) having a plurality of layers (102) is shown. Each layer comprises one or more feature detectors (104), each of which may be associated with activation functions and weights for each parameter input to the respective feature detector (104). Generally, the output of a feature detectors of layer i may be provided as input to one or more feature detector of layer i+1.

[0018] The neural network is implemented by one or more processors. Each feature detector may be considered as a processing "node" and one or more nodes may be implemented by a processor. Further, a memory (106) may be provided for storing activations and learned weights for each feature detector. The memory (106) may further store a training set comprising training data. The training data may, for example, be used for image classification in which case the training data may comprise images with known classifications. The memory (106) may further store a validation set comprising validation data.

[0019] During the training stage, the neural network learns optimal weights for each feature detector. An optimal configuration can then be applied to test data. Exemplary applications of such a neural network include image classification, object recognition and speech recognition.

[0020] A switch (108) is linked to at least a subset of the feature detectors. The switch is operable to selectively disable each feature detector in the neural network to which it is linked, with a learned or preconfigured probability. A random number generator (110) is linked to the switch and provides the switch with a random value that enables the switch to selectively disable each linked feature detector. The possible values generated by the random number generator (110) each correspond to a decision of whether to disable any particular feature detector in accordance with the preconfigured probability.

[0021] In an embodiment, the switch (108) is linked to all feature detectors of the hidden layers. In another embodiment, the switch (108) is linked to all feature detectors of the input layers. In yet another embodiment, the switch (108) may be linked to all feature detectors in both the hidden and input layers. In yet further embodiments, the switch (108) may be linked to the feature detectors of a subset of the input and hidden layers. In another aspect, the switch may be connected to all hidden layers that are fully connected layers.

[0022] Referring now to Fig. 2, during the training stage, a plurality of training cases are input, one at a time, to the neural network in order to train the neural network. For each such training case, the switch selectively disables a subset of the feature detectors to which it is linked (200). In particular embodiments, the switch is configured to disable each such feature detector in accordance with a preconfigured probability. In a more specific embodiment, feature detectors in hidden layers may be selectively disabled with probability 0.5 (that is, on average, each feature detector will be enabled for half of the training cases and disabled for the other half of the training cases) while feature detectors of input layers are disabled with probability 0.2 (that is, on average, these feature detectors will be enabled for 80% of training cases and disabled for 20% of training cases). Therefore, in this example, for each training case, each hidden layer feature detector is randomly omitted from the network with a probability of 0.5 and each input layer feature detector is randomly omitted from the network with a probability 0.2, so each hidden or input feature detector cannot rely on other hidden or input feature detectors being present. Thus, co-adaptation of feature detectors may be reduced.

[0023] Each training case is then processed by the neural network, one at a time (202). For each such training case, the switch may reconfigure the neural network by selectively disabling each linked feature detector.

[0024] Once the training set has been learned by the neural network, the switch may enable all feature detectors and normalize their outgoing weights (204). Normalization comprises reducing the outgoing weights of each feature detector or input by multiplying them by the probability that the feature detector or input was not disabled. In an example, if the feature detectors of each hidden layer were selectively disabled with probability 0.5 in the training stage, the outgoing weights are halved for the test case since approximately twice as many feature detectors will be enabled. A similar approach is applied to the input layers.

[0025] The test set may then be processed by the neural network (206).

[0026] It has been found that the foregoing method provides results similar to performing model averaging with neural networks but is operable to do so with increased efficiency. The described method may enable the training of a plurality of different networks in a reduced time as compared to model averaging.

[0027] In another aspect, a stochastic gradient descent process may be applied for training the neural network on mini-batches of training cases. However, in this example, the penalty term that is normally used to prevent the weights from growing too large may be modified, such that instead of penalizing the squared length (L2 norm) of the whole weight vector, an upper bound may be set on the L2 norm of the incoming weight vector for each individual hidden unit. If a weight-update violates this constraint, the incoming weights of the hidden unit may be renormalized by division. Using a constraint rather than a penalty prevents weights from growing very large regardless of how large the proposed weight-update is. Thus, it may be possible to start with a very large learning rate which decays during learning, thus allowing a thorough search of the weight-space.

[0028] Further, performance may be improved by implementing separate L2 constraints on the mcoming weights of each hidden unit and further improved by also dropping out a random 20% of the pixels.

[0029] In further aspects, performance of the neural network on the test set may be improved by enhancing the training data with transformed images or by wiring knowledge about spatial transformations into a convolutional neural network or by using generative pre-training to extract useful features from the training images without using the labels, or a combination thereof.

[0030] The foregoing method may further be combined with generative pre-training, though it may be beneficial to apply a small learning rate and no weight constraints to avoid losing the feature detectors discovered by the pre-training fine-tuned using standard back-propagation.

[0031] In addition, for speech recognition and object recognition datasets the architecture of the neural network may be adapted by evaluating the performance of a plurality of the more optimal learned neural networks on the validation set and then selecting the architecture that performs best on the validation set as the one to apply to the test set.

[0032] In further examples, for datasets in which the required input-output mapping has several suitably different regimes, performance may be further improved by adapting the preconfigured probabilities to be a learned function of the input.

[0033] In another aspect, the switch may select hidden units in blocks so that all of the hidden units within a block are always either selected or not selected together. The blocks may comprise units in a single hidden layer or may comprise units in more than one hidden layer. [0034] In another aspect, a plurality of switches may be provided to control blocks of hidden units or inputs, including blocks of size 1, and a hidden unit or input is only used in the forward pass if it is selected by all of the switches that are capable of selecting it.

[0035] Example embodiments are now described for applying the foregoing system to particular test data. However, it will be appreciated that alternative neural network configurations may be implemented for such test data, and further neural network configurations may be implemented for other applications. In addition, the type of neural network implemented is not limited merely to feedforward neural networks but can also be applied to any neural networks, including convolutional neural networks, recurrent neural networks, auto-encoders and Boltzmann machines.

What is claimed is:

Claims

1. A computer-implemented method comprising:
obtaining a plurality of training cases; and
training a neural network having a plurality of layers on the plurality of training cases, each of the layers including one or more feature detectors, and each of the feature detectors having a corresponding set of weights, wherein training the neural network on the plurality of training cases comprises, for each of the training cases respectively:
determining one or more feature detectors to disable during processing of the training case,
disabling the one or more feature detectors in accordance with the determining, and
processing the training case using the neural network with the one or more feature detectors disabled to generate a predicted output for the training case.
2. The method of claim 1 , wherein a subset of the feature detectors are associated with respective probabilities of being disabled during processing of each of the training cases, and wherein determining the one or more feature detectors to disable during processing of the training case comprises:
determining whether to disable each of the feature detectors in the subset based on the respective probability associated with the feature detector.
3. The method of claim 2, wherein training the neural network further comprises:
adjusting the weights of each of the plurality of feature detectors to generate trained values for each weight in the set of weights corresponding to the feature detector.
4. The method of claim 3, further comprising:
normalizing the trained weights for each of the feature detectors in the subset, wherein normalizing the trained weights comprises multiplying the trained values of the weights for each of the one or more feature detectors in the subset by a respective probability of the feature detector not being disabled during processing of each of the training cases.
5. The method of claim 2, wherein the subset includes feature detectors in a first layer of the plurality of layers and feature detectors in one or more second layers of the plurality of layers, wherein the feature detectors in the first layer are associated with a first probability and wherein the feature detectors in the one or more second layers are associated with a second, different probability.
6. The method of claim 5, wherein the first layer is an input layer of the neural network and the one or more second layers are hidden layers of the neural network.
7. The method of claim 5, wherein the first layer and the one or more second layers are hidden layers of the neural network.
8. The method of claim 1 , wherein determining one or more feature detectors to disable during processing of the training case comprises:
determining to disable the same feature detectors that were disabled during processing of a preceding training case.
9. The method of claim 1 , wherein the neural network is one of a feedforward neural network, a convolutional neural network, a recurrent neural network, an auto-encoder, or a Boltzmann machine.
10. A system comprising one or more computers and one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising:
obtaining a plurality of training cases; and
training a neural network having a plurality of layers on the plurality of training cases, each of the layers including one or more feature detectors, and each of the feature detectors having a corresponding set of weights, wherein training the neural network on the plurality of training cases comprises, for each of the training cases respectively:
determining one or more feature detectors to disable during processing of the training case,
disabling the one or more feature detectors in accordance with the determining, and
processing the training case using the neural network with the one or more feature detectors disabled to generate a predicted output for the training case.
11. The system of claim 10, wherein a subset of the feature detectors are associated with respective probabilities of being disabled during processing of each of the training cases, and wherein determining the one or more feature detectors to disable during processing of the training case comprises:
determining whether to disable each of the feature detectors in the subset based on the respective probability associated with the feature detector.
12. The system of claim 11 , wherein training the neural network further comprises:
adjusting the weights of each of the plurality of feature detectors to generate trained values for each weight in the set of weights corresponding to the feature detector.
13. The system of claim 12, further comprising:
normalizing the trained weights for each of the feature detectors in the subset, wherein normalizing the trained weights comprises multiplying the trained values of the weights for each of the one or more feature detectors in the subset by a respective probability of the feature detector not being disabled during processing of each of the training cases.
14. The system of claim 11 , wherein the subset includes feature detectors in a first layer of the plurality of layers and feature detectors in one or more second layers of the plurality of layers, wherein the feature detectors in the first layer are associated with a first probability and wherein the feature detectors in the one or more second layers are associated with a second, different probability.
15. The system of claim 14, wherein the first layer is an input layer of the neural network and the one or more second layers are hidden layers of the neural network.
16. The system of claim 14, wherein the first layer and the one or more second layers are hidden layers of the neural network.
17. The system of claim 11, wherein determining one or more feature detectors to disable during processing of the training case comprises: determining to disable the same feature detectors that were disabled during processing of a preceding training case.
18. The system of claim 11 , wherein the neural network is one of a feedforward neural network, a convolutional neural network, a recurrent neural network, an auto-encoder, or a Boltzmann machine.
19. A computer storage medium encoded with a computer program, the program comprising instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:
obtaining a plurality of training cases; and
training a neural network having a plurality of layers on the plurality of training cases, each of the layers including one or more feature detectors, and each of the feature detectors having a corresponding set of weights, wherein training the neural network on the plurality of training cases comprises, for each of the training cases respectively:
determining one or more feature detectors to disable during processing of the training case,
disabling the one or more feature detectors in accordance with the determining, and
processing the training case using the neural network with the one or more feature detectors disabled to generate a predicted output for the training case.
20. The computer storage medium of claim 19, wherein a subset of the feature detectors are associated with respective probabilities of being disabled during processing of each of the training cases, and wherein determining the one or more feature detectors to disable during processing of the training case comprises:
determining whether to disable each of the feature detectors in the subset based on the respective probability associated with the feature detector.
PCT/US2013/077608 2012-12-24 2013-12-23 System and method for addressing overfitting in a neural network WO2014105866A1 (en)

Priority Applications (4)

Application Number Priority Date Filing Date Title
US201261745711P true 2012-12-24 2012-12-24
US61/745,711 2012-12-24
US14/015,768 US9406017B2 (en) 2012-12-24 2013-08-30 System and method for addressing overfitting in a neural network
US14/015,768 2013-08-30

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
AU2013370515A AU2013370515A1 (en) 2012-12-24 2013-12-23 System and method for addressing overfitting in a neural network
BR112015015230A BR112015015230A2 (en) 2012-12-24 2013-12-23 system and process for addressing generalization in a neural network

Publications (1)

Publication Number Publication Date
WO2014105866A1 true WO2014105866A1 (en) 2014-07-03

Family

ID=50975834

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2013/077608 WO2014105866A1 (en) 2012-12-24 2013-12-23 System and method for addressing overfitting in a neural network

Country Status (4)

Country Link
US (2) US9406017B2 (en)
AU (1) AU2013370515A1 (en)
BR (1) BR112015015230A2 (en)
WO (1) WO2014105866A1 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3343456A1 (en) 2016-12-27 2018-07-04 Denso IT Laboratory, Inc. Information estimation apparatus and information estimation method

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20160012537A (en) * 2014-07-24 2016-02-03 삼성전자주식회사 Neural network training method and apparatus, data processing apparatus
US10185803B2 (en) 2015-06-15 2019-01-22 Deep Genomics Incorporated Systems and methods for classifying, prioritizing and interpreting genetic variants and therapies using a deep neural network
US9569696B1 (en) * 2015-08-12 2017-02-14 Yahoo! Inc. Media content analysis system and method
US9965719B2 (en) * 2015-11-04 2018-05-08 Nec Corporation Subcategory-aware convolutional neural networks for object detection
US9858340B1 (en) 2016-04-11 2018-01-02 Digital Reasoning Systems, Inc. Systems and methods for queryable graph representations of videos
WO2018199721A1 (en) * 2017-04-28 2018-11-01 서울대학교 산학협력단 Method and apparatus for accelerating data processing in neural network

Non-Patent Citations (7)

* Cited by examiner, † Cited by third party
Title
A. M. DUDA: "NeuroXyce synapse device with STDP and stochastic transmission reliability", SAND REPORT FROM SANDIA NATIONAL LABORATORIES, 2012, XP055100780, Retrieved from the Internet <URL:http://www.alexduda.com/2012_SAND.pdf> [retrieved on 20140205] *
D. GUO, C. LI: "Stochastic resonance in Hodgkin-Huxley neuron induced by unreliable synaptic transmission", JOURNAL OF THEORETICAL BIOLOGY, vol. 308, 9 June 2012 (2012-06-09), pages 105 - 114, XP028426911, DOI: 10.1016/j.jtbi.2012.05.034 *
G. E. HINTON, N. SRIVASTAVA, A. KRIZHEVSKY, I. SUTSKEVER, R. R. SALAKHUTDINOV: "Improving neural networks by preventing co-adaptation of feature detectors", ARXIV:1207.0580V1, 3 July 2012 (2012-07-03), XP055112910, Retrieved from the Internet <URL:http://arxiv.org/abs/1207.0580v1> [retrieved on 20140407] *
G. HINTON, G. DAHL: "Dropout: a simple and effective way to improve neural networks", THE 26TH ANNUAL CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS (NIPS'25): DECEMBER 3-8, 2012, 6 December 2012 (2012-12-06), XP055112915, Retrieved from the Internet <URL:http://videolectures.net/nips2012_hinton_networks/nips2012_hinton_networks_01.pdf> [retrieved on 20140407] *
N. SRIVASTAVA: "Improving neural networks with dropout", THESIS FOR MASTER OF SCIENCE AT UNIVERSITY OF TORONTO, January 2013 (2013-01-01), XP055112918, Retrieved from the Internet <URL:http://wayback.archive.org/web/20130512061052/http://www.cs.toronto.edu/~nitish/msc_thesis.pdf> [retrieved on 20140407] *
None
S. WANG, C. D. MANNING: "Fast 'dropout' training for logistic regression", WORKSHOP ON LOG-LINEAR MODELS AT THE 26TH ANNUAL CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS, 8 December 2012 (2012-12-08), XP055112919, Retrieved from the Internet <URL:https://docs.google.com/file/d/0B_9blvt8HjXzbGJTZlpWczRyMGM/edit> [retrieved on 20140407] *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3343456A1 (en) 2016-12-27 2018-07-04 Denso IT Laboratory, Inc. Information estimation apparatus and information estimation method

Also Published As

Publication number Publication date
US20140180986A1 (en) 2014-06-26
US9406017B2 (en) 2016-08-02
US20160335540A1 (en) 2016-11-17
AU2013370515A1 (en) 2015-07-16
BR112015015230A2 (en) 2017-07-11

Similar Documents

Publication Publication Date Title
Finn et al. Model-agnostic meta-learning for fast adaptation of deep networks
Le et al. A simple way to initialize recurrent networks of rectified linear units
Chen et al. Median filtering forensics based on convolutional neural networks
Zhang et al. Understanding deep learning requires rethinking generalization
Sakurada et al. Anomaly detection using autoencoders with nonlinear dimensionality reduction
US20130117212A1 (en) Methods and apparatus for unsupervised neural replay, learning refinement, association and memory transfer: neural associative learning, pattern completion, separation, generalization and hierarchical replay
Scardapane et al. Distributed learning for random vector functional-link networks
Henaff et al. Tracking the world state with recurrent entity networks
Györgyi Inference of a rule by a neural network with thermal noise
US20120254086A1 (en) Deep convex network with joint use of nonlinear random projection, restricted boltzmann machine and batch-based parallelizable optimization
de Oliveira et al. A hybrid search method for the vehicle routing problem with time windows
Perez et al. Film: Visual reasoning with a general conditioning layer
Khan et al. Fuzzy cognitive maps with genetic algorithm for goal-oriented decision support
Mourao et al. Learning STRIPS operators from noisy and incomplete observations
Carvalho et al. PROGNOS: predictive situational awareness with probabilistic ontologies
JP2014049118A (en) Convolution neural network classifier system, training method for the same, classifying method, and usage
Muhlbaier et al. Learn++. MT: A new approach to incremental learning
Wierstra et al. Modeling systems with internal state using evolino
US20130266214A1 (en) Training an image processing neural network without human selection of features
Eberhardt et al. N-1 experiments suffice to determine the causal relations among n variables
Noh et al. Training recurrent answering units with joint loss minimization for vqa
Biehl et al. On-line learning of a time-dependent rule
US20160171346A1 (en) Image recognition method and apparatus, image verification method and apparatus, learning method and apparatus to recognize image, and learning method and apparatus to verify image
Tobiyama et al. Malware detection with deep neural network using process behavior
Nordbotten Editing statistical records by neural networks

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 13821374

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase in:

Ref country code: DE

REG Reference to national code

Ref country code: BR

Ref legal event code: B01A

Ref document number: 112015015230

Country of ref document: BR

ENP Entry into the national phase in:

Ref document number: 2013370515

Country of ref document: AU

Date of ref document: 20131223

Kind code of ref document: A

122 Ep: pct application non-entry in european phase

Ref document number: 13821374

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase in:

Ref document number: 112015015230

Country of ref document: BR

Kind code of ref document: A2

Effective date: 20150624