1. PATTERN RECOGNITION AND MACHINE LEARNING CHAPTER 3: LINEAR MODELS FOR REGRESSION

2. What is Supervised- Unsupervised Learning? Regression: Prediction Estimation Classification Detection Hypothesis Testing

3. Supervised Learning Given training data D = {xn,tn}, find t for an unknown x Deterministic: Make a model for t=y(x) Stochastic: Model p(t/x): uncertainity of t for a given x Then Minimize the expecte loss

4. Linear Basis Function Models Example: Polynomial Curve Fitting

5. Linear regression Generally Á j (x): basis functions, linear w.r. To w. Á 0 ( x ) = 1, so that w 0 acts as a bias. Linear basis functions : Á d ( x ) = x d.

6. Polynomial basis functions: These are global; a small change in x affect all basis functions.

7. Gaussian basis functions: These are local; a small change in x only affect nearby basis functions. ¹ j and s control location and scale (width).

8. Sigmoidal basis functions: where Also these are local; a small change in x only affect nearby basis functions. ¹ j and s control location and scale (slope).

9. Two Approaches for supervised learning: 1.Maximum Likelihood Estimate: y(x,w) is a deterministic function with added Gaussian noise: 2. Bayes Estimate w is a stochastic parameter which comes from And conjugate prior is

10. Maximum Likelihood and Least Squares (1) Given observed inputs and targets, D = {xn,tn}, we obtain the likelihood function for is

11. Maximum Likelihood and Least Squares (2) Taking the logarithm, we get where is the sum-of-squares error.

12. Computing the gradient and setting it to zero yields Solving for w, we get where Maximum Likelihood and Least Squares (3) The Moore-Penrose pseudo-inverse,.

13. Maximum Likelihood and Least Squares (4) Maximizing with respect to the bias, w 0, alone, we see that We can also maximize with respect to ¯, giving

14. Geometry of Least Squares Consider t =(t1,...,tN )T S is spanned by. w ML minimizes the distance between t and its orthogonal projection on S, i.e. y. N -dimensional M -dimensional

15. Sequential Learning Data items considered one at a time, use stochastic (sequential) gradient descent: This is known as the least-mean-squares (LMS) algorithm. Issue: how to choose ´ ?

16. Regularized Least Squares Consider the error function: With the sum-of-squares error function and a quadratic regularizer, we get which is minimized by Data term + Regularization term ¸ is called the regularization coefficient.

17. General Regularizer Λ defines the size q defines the shape LassoQuadratic

18. Regularized least squares optimizes the least squares error subject to the regularization constraint min s.to Lagrange Multipliers, form Hamiltonian

19. Sparcity: Some w becomes 0 as λ increases Lasso tends to generate sparser solutions than a quadratic regularizer. E D (w) E w (w)

20. Multiple Outputs (1) Analogously to the single output case we have: Given observed inputs,, and targets,, we obtain the log likelihood function

21. Multiple Outputs (2) Maximizing with respect to W, we obtain If we consider a single target variable, t k, we see that where, which is identical with the single output case.

22. Kırk Katır? Kırk satır? Overfitting? Underfitting?

23. Bias-Variance Trade-off Recall the expected quare error loss Minimize it Optimal value of y(x) is

24. Minimum Expected suqare error loss where The second term of E [L] corresponds to the noise inherent in the random variable t. What about the first term?

25. The Bias-Variance Decomposition Given multiple data sets, each of size N. Any particular data set, D, will give a particular function y(x;D). Let E D [y(x; D)] be average prediction over all data sets. Then,

26. The Bias-Variance Decomposition

27. bias: average prediction over all data sets - the desired regression function. variance: the extent to which the solutions for individual data sets vary around their average, measures the extent to which the function y(x; D) is sensitive to the particular choice of data set.

28. The Bias-Variance Decomposition Thus, our goal is to minimize where

29. But… The Bias-Variance Trade-off From these plots, we note that an over-regularized model (large ¸ ) will have a high bias, while an under- regularized model (small ¸ ) will have a high variance.

30. The Bias-Variance Decomposition (7) Example: 25 data sets from the sinusoidal, varying the degree of regularization, ¸.

31. The Bias-Variance Decomposition (6) Example: 25 data sets from the sinusoidal, varying the degree of regularization, ¸.

32. The Bias-Variance Decomposition Example: 25 data sets from the sinusoidal, varying the degree of regularization, ¸.

33. Recall, for supervised learning: 1.Maximum Likelihood Estimate: y(x,w) is a deterministic function with added Gaussian noise: 2. Bayes Estimate w is a stochastic parameter which comes from And conjugate prior is

34. BAYES POSTERİOR≈ LİKELİHOOD. CONJUGATE PRIOR

35. Bayesian Linear Regression (1) Define a conjugate prior over w Combining this with the likelihood function and using results for marginal and conditional Gaussian distributions, gives the posterior where

36. Bayesian Linear Regression (2) A common choice for the prior is zero-mean isotropic Gaussian governed by a single precision parameter α so that for which Next we consider an example …

37. Bayesian Linear Regression (3) 0 data points observed. Prior is Prior Data Space Alpha is 2

38. Bayesian Linear Regression (4) 1 data point observed from y(x, w)=w 0 + w 1 x Select w 0 = -0.3 and w 1 =0.5 Choose xn from U(x|-1, 1), find f(xn, w), Add Gaussian noise N(e/0, 0.2) to obtain the tn Goal: recover w 0 and w 1 LikelihoodPosterior Data Space

39. Bayesian Linear Regression (5) 2 data points observed LikelihoodPosterior Data Space

40. Bayesian Linear Regression (6) 20 data points observed LikelihoodPosterior Data Space

41. Predictive Distribution (1) Forget about estimating w. Predict t for new values of x by integrating over w : where Data noise Uncertainity of w

42. Example : Sinusoidal data, 9 Gaussian basis functions, 1 data point Sample fonctions drawn from posterior Predictive distribution Pink region: point-wise predictive variance as a function of x. Red curve: mean of the corresponding Gaussian predictive distributiondistribution,

43. Example : Sinusoidal data, 9 Gaussian basis functions, 2 data points Note the point- wise uncertainty in the predictions

44. Example: Sinusoidal data, 9 Gaussian basis functions, 4 data points

45. Example: Sinusoidal data, 9 Gaussian basis functions, 25 data points

46. Equivalent Kernel Recall the Bayesian posterior with isotropic Gaussian prior Where Note that because the posterior distribution is Gaussian, its mode coincides with its mean. Thus the maximum posterior weight vector is

47. Equivalent Kernel (1) The predictive mean can be written This is a weighted sum of the training data target values, t n. Equivalent kernel or smoother matrix.

48. Equivalent kernel k(x, x’) for Gaussian Basis function Weight of t n depends on distance between x and x n nearby x n carry more weight.

49. Examples of equivalent kernels k(x, x ) for x =0 plotted as a function of x Non-local basis functions have local equivalent kernels: PolynomialSigmoidal

50. The kernel as a covariance function: consider recall that the predictive mean at x=x’ is highly correlated, whereas for more distant pairs of points the correlation will be smaller.

51. Properties of Equivalent Kernel for all values of x ; however, the equivalent kernel may be negative for some values of x. Like all kernel functions, the equivalent kernel can be expressed as an inner product: where.

52. An alternative approach to regression Instead of introducing a set of basis functions, which implicitly determines an equivalent kernel, define a localized kernel directly and use this to make predictions for new input vectors x, given the observed training set.

53. Bayesian Model Comparison (1) Model; the underlying pdf. Complexity: The order of the linear regression eq. Note that there are infinetly many different models Among them how do we choose the ‘right’ model?

54. Bayesian Model Comparison (1) Assume we want to compare models M i i=1, …,L, using data D; this requires computing Bayes Factor: ratio of evidence for two models PosteriorPrior Model evidence or marginal likelihood

55. What is Model Evidance? For a model with parameters w, we get the model evidence by marginalizing over w Note that

56. Bayesian Model Comparison (2) Having computed the posterior p(M i /D), and the predictive distribution p(t/x,Mi,D) for each model,we can compute the predictive (mixture) distribution Note: p(t/x,D) is multimodal. A simpler approximation, known as model selection, is to use the model with the highest evidence.

57. Approximate Model Evidance For a given model Mi with a single parameter, w, con- sider the approximation where the posterior is assumed to be sharply peaked.

58. Approximate Model Evidance for a specific Mi Taking logarithms, we obtain With M parameters, all assumed to have the same ratio, we get Negative Negative and linear in M.

59. Evidance with different complexities M1 simplest, M2 medium, M3 most complex

60. The Evidence Approximation (1) The fully Bayesian predictive distribution is given by but this integral is intractable. Approximate with where is the mode of, which is assumed to be sharply peaked;

61. How do you obtain From Bayes’ theorem we have and if we assume p(®,¯) to be flat we see that General results for Gaussian integrals give

62. Approximate Log Evidance

63. Comparison of Approximate Evidance and Root Mean Square Error

64. Maximizing the Evidence Function (1) To maximise w.r.t. ® and ¯, we define the eigenvector equation Let has eigenvalues ¸ i + ®.

65. Maximizing the Evidence Function (2) We can now differentiate w.r.t. ® and ¯, and set the results to zero, to get where depends on both ® and ¯.

66. Algorithm for estimation 1.İnitialize alpha, beta 2.Compute lambda: 3.Compute gamma: 4.Find mN: 5.Re-estimate alpha, beta 6. Re iterate until convergence

67. Effective Number of Parameters (3) Likelihood Prior w 1 is not well determined by the likelihood w 2 is well determined by the likelihood ° is the number of well determined parameters

68. Effective Number of Parameters (2) Example: sinusoidal data, 9 Gaussian basis functions, ¯ = 11.1.

69. Effective Number of Parameters (3) Example: sinusoidal data, 9 Gaussian basis functions, ¯ = 11.1. Test set error

70. Effective Number of Parameters (4) Example: sinusoidal data, 9 Gaussian basis functions, ¯ = 11.1.

71. Effective Number of Parameters (5) In the limit, ° = M and we can consider using the easy-to-compute approximation

72. Limitations of Fixed Basis Functions M basis function along each dimension of a D -dimensional input space requires M D basis functions: the curse of dimensionality. In later chapters, we shall see how we can get away with fewer basis functions, by choosing these using the training data.