Introduction Machine learning The machine learning process An Introduction to Machine Learning Fabio Gonz´ alez, PhD MindLAB Research Group - Universidad Nacional de Colombia Introducci´ on a los Sistemas Inteligentes Fabio Gonz´ alez, PhD An Introduction to Machine Learning
38
Embed
An Introduction to Machine Learning - GitHub Pages · Introduction Machine learning The machine learning process What’s machine learning History Supervised learning Non-supervised
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
IntroductionMachine learning
The machine learning process
An Introduction to Machine Learning
Fabio Gonzalez, PhD
MindLAB Research Group - Universidad Nacional de Colombia
Introduccion a los Sistemas Inteligentes
Fabio Gonzalez, PhD An Introduction to Machine Learning
3 The machine learning processModel learningModel evaluationFeature extractionModel application
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Observation and analysis
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Tycho Brahe
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Tycho Brahe
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Johannes Kepler
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Data and models
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Machine Learning
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Machine Learning with Images
Learning/ Model
Induction
Data Model
outperform the majority of state-of-the-art approaches in mitosisdetection.
The organization of the rest of this paper is as follows.In Sec. 2, we describe motivations of the proposal. In Sec. 3,we describe details of the new methodology. In Sec. 4, wepresent experimental results. Finally, in Sec. 5, we present ourconcluding remarks.
2 Motivation and Previous WorkRecently, the development of computerized systems for auto-mated mitosis detection has become an active area of researchwith the goal of developing decision support systems to be ableto relieve the workload of the pathologist. In a contest held inconjunction with the ICPR 2012 conference5,6 to identify the bestautomated mitosis detection algorithm, a variety of approachescompeted against each other. These approaches can be catego-rized as handcrafted feature based or feature learning based. Thecommonly used handcrafted features include various morpho-logical, shape, statistical, and textural features that attempt tomodel the appearance of the domain and, in particular, theappearance of the mitoses within the digitized images.7–10
Although domain inspired approaches (hand crafted) are use-ful in that they allow for explicit modeling of the kinds of fea-tures that pathologists look for when identifying mitoses, thereis another category of feature generation inspired by convolu-tional neural networks (CNN),11,12 CNN are multilayer neuralnetworks that learns a bank of convolutional filters at eachlayer.13,14 In contrast to handcrafted features, CNN is fully data-driven, therefore, it is more accurate in representing training
samples and is able to find feature patterns that handcraftedfeatures fail to describe. However, CNN is computationallydemanding and sensitive to the scalability of the training data.The winner14 of the ICPR contest used two 11 layers to achievean F-measure of 0.78. However, this approach is not feasible forclinical use since each layer of the CNN model comprised hun-dreds of neurons and required a large amount of time (severalweeks) for both training and testing.
Other methods achieved an F-measure of up to 0.71, basedprimarily on combining various handcrafted features. Althoughhandcrafted feature approaches are faster, drawbacks include(1) the fact that the identification of salient features is highlydependent on the evaluation dataset used and (2) the lack ofa principled approach for combining disparate features. Hence,it stands to reason that a combination of CNN and handcraftedfeatures will allow us to exploit the high accuracy of CNN whilealso reducing the computational burden (in terms of time) oftraining deep CNN models. By employing a light CNN model,the proposed approach is far less demanding computationally,and the cascaded strategy of combining handcrafted featuresand CNN-derived features enables the possibility of maximizingperformance by leveraging the disconnected feature sets.Previous work in this approach includes the Nippon ElectricCompany (NEC) team,13 where an attempt was made to stackthe CNN-learned features and handcrafted features yielded anF-measure of 0.659, suggesting that more intelligent combina-tions of CNN and handcraft features are required.
In this paper, we present a cascaded approach to combiningCNN and handcrafted features for mitosis detection. The work-flow of the new approach is depicted in Fig. 2. The first step is to
(a) (b) (c) (d) (e) (f)
Fig. 1 An illustration of the visual similarity between true mitotic processes and confounding false pos-itives. (a)–(c) True mitoses. (d)–(f) Confounding nonmitotic figures.
Fig. 2 Workflow of our methodology. Blue-ratio thresholding15 is first applied to segment mitosis can-didates. On each segmented blob, handcrafted features are extracted and classified via a random forestsclassifier. Meanwhile, on each segmented 80 × 80 patch, convolutional neural networks (CNN)11 aretrained with a fully connected regression model as part of the classification layer. For those candidatesthat are difficult to classify (ambiguous result from the CNN), we train a second-stage random forestsclassifier on the basis of combining CNN-derived and handcrafted features. Final decision is obtained viaa consensus of the predictions of the three classifiers.
Journal of Medical Imaging 034003-2 Oct–Dec 2014 • Vol. 1(3)
Wang et al.: Mitosis detection in breast cancer pathology images by combining handcrafted and convolutional neural network features
Downloaded From: http://medicalimaging.spiedigitallibrary.org/ on 10/12/2014 Terms of Use: http://spiedl.org/terms
Prediction
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
The fourth paradigm
Fabio Gonzalez, PhD An Introduction to Machine Learning
Some samples may have labels,in that case it is calledsemi-supervised learning
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Model learningModel evaluationFeature extractionModel application
The machine Learning process
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Model learningModel evaluationFeature extractionModel application
Model learning
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Model learningModel evaluationFeature extractionModel application
Model induction from data
Learning is an ill-posed problem (more than one possiblesolution for the same particular problem, solutions aresensitive to small changes on the problem)
It is necessary to make additional assumptions about the kindof pattern that we want to learn
Hypothesis space: set of valid patterns that can be learnt bythe learning algorithm
Occam’s razor: ”All things being equal, the simplest solutiontends to be the best one.”
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Model learningModel evaluationFeature extractionModel application
Approaches to learning
Probabilistic:
Generative models: model P(Y ,X )Discriminative models: model P(Y |X )
Geometrical:
Manifold learning: model the geometry of the space where thedata livesMax margin learning: model the separation between the classes
Optimization:
Energy/loss/risk minimization
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Model learningModel evaluationFeature extractionModel application
Learning as optimization
General optimization problem:
minf ∈H
L(f ,D),
with H:hypothesis space, D:training data, L:loss/error
Example, logistic regression:
Hypothesis space:
y(x) = P(C+|x) = σ(wT x)
Cross-entropy error:
E (w) = − ln p(t|w) = −∑n=1
[tn ln yn + (1− tn) ln(1− yn)]
Fabio Gonzalez, PhD An Introduction to Machine Learning
IntroductionMachine learning
The machine learning process
Model learningModel evaluationFeature extractionModel application
Methods
Supervised generative:
Naıve BayesGraphical modelsMarkov random fieldsHidden markov models
Supervised discriminative:
Logistic regressionRidge regressionConditional random fields