Semisupervised learning

1
Semi-supervised learning
Guido Sanguinetti Department of Computer Science,
University of Sheffield
2
Programme

• Different ways of learning
• Unsupervised learning
• EM algorithm
• Supervised learning
• Different ways of going semi-supervised
• Generative models
• Reweighting
• Discriminative models
• Regularisation clusters and manifolds

3
Disclaimer

• This is going to be a high-level introduction
  rather than a technical talk
• We will spend some time talking about supervised
  and unsupervised learning as well
• We will be unashamedly probabilistic, if not
  fully Bayesian
• Main reference C.Bishops book Pattern
  Recognition and Machine Learning and M. Seegers
  chapter in Semi-supervised learning, Chapelle,
  Schölkopf and Zien, eds.

4
Different ways to learn

• Reinforcement learning learn a strategy to
  optimise a reward.
• Closely related to decision theory and control
  theory.
• Supervised learning learn a map
• Unsupervised learning estimate a density

5
Unsupervised learning

• We are given data x.
• We want to estimate the density that generated
  the data.
• Generally, assume a latent variable y as in
  graphical model.
• A continuous y leads to dimensionality reduction
  (cf previous lecture), a discrete one to
  clustering

?
p
x
y
6
Example mixture of Gaussians

• Data vector measurements xi, i1,...,N.
• Latent variable K-dimensional binary vectors
  (class membership) yi yij1 means point i
  belongs to class j.
• The ? parameters are in this case the covariances
  and the means of each component.

7
Estimating mixtures

• Two objectives estimate the parameters ?, ?j and
  ?j and estimate the posterior probabilities on
  class membership
• ?ij are the responsibilities.
• Could use gradient descent.
• Better to use EM

8
Expectation-Maximization

• Iterative procedure to estimate the maximum
  likelihood values of parameters in models with
  latent variables.
• We want to maximise the log-likelihood of the
  model
• Notice that log of a sum is not nice.
• The key mathematical tool is Jensens inequality

9
Jensens inequality

• For every concave function f, random variable x
  and probability distribution q(x)
• A cartoon of the proof, which relies on the
  centre of mass of a convex polygon lying inside
  the polygon.

10
Bound on the log-likelihood

• Jensens inequality leads to
• Hq is the entropy of the distribution q. Notice
  the absence of the nasty log of a sum.
• If and only if q(c) is the posterior q(cx), we
  get that the bound is saturated

11
EM
E-step

• For a fixed value of the parameters, the
  posterior will saturate the bound.
• In the M-step, optimise the bound with respect to
  the parameters.
• In the E-step, recompute the posterior with the
  new value of the parameters.
• Exercise EM for mixture of Gaussians.

M-step
?
12
Supervised learning

• The data consists of a set of inputs x and a set
  of output (target) values y.
• The goal is to learn the functional relation
• fx? y
• Evaluated using reconstruction error
  (classification accuracy), usually on a separate
  test set.
• Continuous y leads to regression, discrete to
  classification

13
Classification-Generative

• The generative approach starts with modelling
  p(xc) as in unsupervised learning.
• The model parameters are
• estimated (e.g. using Maximum
• Likelihood).
• The assignment is based on the
• posterior probabilities p(cx)
• Requires estimation of many parameters.

?
p
x
c
14
Example discriminant analysis

• Assume class conditional distributions to be
  Gaussian
• Estimate means and covariances using maximum
  likelihood (O(KD2) params).
• Classify novel (test) data using posterior
• Exercise rewrite posterior in terms of sigmoids.

15
Classification-Discriminative

• Also called diagnostic
• Modelling the class-conditional
• distributions involves significant
• overheads.
• Discriminative techniques avoid
• this by modelling directly the posterior
  distribution p(cjx).
• Closely related with the concept of transductive
  learning.

µ
?
x
c
16
Example logistic regression

• Restrict to the two classes case.
• Model posteriors as
• where we have introduced the logistic sigmoid
• Notice similarity with discriminant function in
  generative models.
• Number of parameters to be estimated is D.

17
Estimating logistic regression

• Let
• Let ci be 0 or 1.
• The likelihood for the data (xi,ci) is
• The gradient of the negative log-likelihood is
• Overfitting, may need a prior on w

18
Semi-supervised learning

• In many practical applications, the labelling
  process is expensive and time consuming.
• Plenty of examples of inputs x, but few of the
  corresponding target c
• The goal is still to predict p(cx) and is
  evaluated using classification accuracy
• Semi-supervised learning is, in this sense, a
  special case of supervised learning where we use
  the extra unlabeled data to improve the
  predictive power.

19
Notation

• We have a labeled or complete data set Dl,
  comprising of Nl sets of pairs (x,c)
• We have an unlabeled or incomplete data set Du
  comprising of Nu sets of vectors x
• The interesting (and common) case is when NugtgtNl

20
Baselines

• One attack to semi-supervised learning would be
  to ignore the labels and estimate an unsupervised
  clustering model.
• Another attack is to ignore the unlabeled data
  and just use the complete data.
• No free lunch it is possible to construct data
  distributions for which either of the baselines
  outperforms a given SSL method.
• The art in SSL is designing good models for
  specific applications, rather than theory.

21
Generative SSL

• Generative SSL methods are most intuitive
• The graphical model for
• generative classification and
• unsupervised learning is the
• same
• The likelihoods are combined
• Parameters are estimated by EM

?
p
x
c
22
Discriminant analysis

• McLachlan (1977) considered SSL for a generative
  model with two Gaussian classes
• Assume the covariance to be known, means need to
  be estimated from the data.
• Assume we have NuM1M2ltltNl
• Assume also that the labelled data is split
  evenly among the two classes, so Nl2n

23
A surprising result

• Consider the misclassification expected error R
• Let R be the expected error obtained using only
  the labelled data, and R1 the expected error
  obtained using all data (estimating parameters
  with EM)
• If , the first order expansion
  (in Nu/Nl) of R-R1 is positive only for NultM,
  with M finite
• In other words, unlabelled data helps up to a
  point only.

24
A way out

• Perhaps considering labelled and unlabelled data
  on the same footing is not ideal.
• McLachlan went on to propose to estimate the
  class means as
• He then shows that, for suitable ?, the expected
  error obtained in this way is to first order
  always lower than the one obtained without the
  unlabelled data

25
A hornets nest

• In general, it seems sensible to reweight the
  log-likelihood as
• This is partly because, in the important case
  when Nl is small, we do not want the label
  information to be swamped
• But how do you choose ??
• Cross validation on the labels?
• Unfeasible for small Nl

26
Stability

• An elegant solution proposed by Corduneanu and
  Jaakkola (2002)
• In the limit when all data is labelled the
  likelihood is unimodal
• In the limit when all data is unlabelled the
  likelihood is multimodal (e.g. permutations)
• When moving from ?1 to ?0, we must encounter a
  critical ? when the likelihood becomes
  multimodal
• That is the optimal choice

27
Discriminative SSL

• The alternative paradigm for SSL
• is the discriminative approach
• As in supervised learning, the
• discriminative approach models
• directly p(cx) using the graphical model
  above
• The total likelihood factorises as

µ
?
x
c
28
Surprise!

• Using Bayes theorem we obtain the posterior over
  \theta as
• This is seriously worrying!
• Why?

29
Regularization

• Bayesian methods in a straight discriminative SSL
  cannot be employed
• Some limited success in non-Bayesian methods
  (e.g. Anderson 1978 for logistic regression over
  discrete variables)
• Most promising avenue is to
• modify the graphical model used
• This is known as regularization

µ
?
x
c
30
Discriminative vs Generative

• The idea behind regularization is that
  information about the generative structure of x
  is feeding into the discriminative process
• Does it still make sense to talk about a
  discriminative approach?
• Strictly speaking, perhaps no
• In practice, the hypothesis used for
  regularization are much weaker than modelling the
  full class conditional distribution

31
Cluster assumption

• Perhaps the most widely used form of
  regularization
• It states that the boundary between classes
  should cross areas of low data density
• It is often reasonable particularly in high
  dimensions
• Implemented in a host of Bayesian and
  non-Bayesian methods
• Can be understood in terms of smoothness of the
  discriminant function

32
Manifold assumption

• Another convenient assumption is that the data
  lies on a low-dimensional sub-manifold of a
  high-dimensional space
• Often the starting point is to map the data to a
  high dimensional space via a feature map
• The cluster assumption is then applied in the
  high dimensional space
• Key concept is the Graph Laplacian
• Ideas pioneered by Belkin and Niyogi

33
Manifolds cont.

• We want discriminant functions which vary little
  in regions of high data density
• Mathematically, this is equivalent to
• being very small, where ? is the
  Laplace-Beltrami operator
• The finite sample approximation to ? is given by
  the graph Laplacian
• The objective function for SSL is a combination
  of the error given by f and of its norm under ?