Bias–variance tradeoff

File:Truen bad prec ok.png|High bias, low variance File:Truen bad prec bad.png|High bias, high variance File:En low bias low variance.png|Low bias, low variance File:Truen ok prec bad.png|Low bias, high variance The bias–variance tradeoff is a central problem in supervised learning. Ideally, one wants to choose a model that both accurately captures the regularities in its training data, but also generalizes well to unseen data. Unfortunately, it is typically impossible to do both simultaneously. High-variance learning methods may be able to represent their training set well but are at risk of overfitting to noisy or unrepresentative training data. In contrast, algorithms with high bias typically produce simpler models that may fail to capture important regularities (i.e. underfit) in the data. It is an often made fallacy to assume that complex models must have high variance. High variance models are "complex" in some sense, but the reverse need not be true. In addition, one has to be careful how to define complexity. In particular, the number of parameters used to describe the model is a poor measure of complexity. The model f_{a,b}(x)=a\sin(bx) has only two parameters (a,b) but it can interpolate any number of points by oscillating with a high enough frequency, resulting in both a high bias and high variance. An analogy can be made to the relationship between accuracy and precision. Accuracy is one way of quantifying bias and can intuitively be improved by selecting from only local information. Consequently, a sample will appear accurate (i.e. have low bias) under the aforementioned selection conditions, but may result in underfitting. In other words, test data may not agree as closely with training data, which would indicate imprecision and therefore inflated variance. A graphical example would be a straight line fit to data exhibiting quadratic behavior overall. Precision is a description of variance and generally can only be improved by selecting information from a comparatively larger space. The option to select many data points over a broad sample space is the ideal condition for any analysis. However, intrinsic constraints (whether physical, theoretical, computational, etc.) will always play a limiting role. The limiting case where only a finite number of data points are selected over a broad sample space may result in improved precision and lower variance overall, but may also result in an overreliance on the training data (overfitting). This means that test data would also not agree as closely with the training data, but in this case the reason is inaccuracy or high bias. To borrow from the previous example, the graphical representation would appear as a high-order polynomial fit to the same data exhibiting quadratic behavior. Note that error in each case is measured the same way, but the reason ascribed to the error is different depending on the balance between bias and variance. To mitigate how much information is used from neighboring observations, a model can be smoothed via explicit regularization, such as shrinkage. ==Bias–variance decomposition of mean squared error==

File:Bias-variance decomposition.png|thumb|Bias–variance decomposition in the case of mean squared loss. The green dots are samples of test label y at a fixed test feature x. Their variance around the mean \mathbb E_{y \sim p(\cdot | x)}[y] is the irreducible error \sigma^2. The red dots are test label predictions f(x | D) as the training set D is randomly sampled. Their variance around the mean \mathbb E_D[f(x | D)] is the variance \operatorname{Var}_D\big[f(x | D)\big] . The difference between the red dash and the green dash is the bias \operatorname{Bias}_D\big[f (x | D)\big] . The bias–variance decomposition is then visually clear: the mean squared error between the red dots and the green dots is the sum of the three components. Suppose that we have a training set consisting of a set of points x_1, \dots, x_n and real-valued labels y_i associated with the points x_i. We assume that the data is generated by a function f(x) such as y = f(x) + \varepsilon, where the noise, \varepsilon, has zero mean and unit variance \sigma^2. That is, y_i = f(x_i) + \varepsilon_i, where \varepsilon_i is a noise sample. We want to find a function \hat{f}\!(x;D), that approximates the true function f(x) as well as possible, by means of some learning algorithm based on a training dataset (sample) D=\{(x_1,y_1) \dots, (x_n, y_n)\}. We make "as well as possible" precise by measuring the mean squared error between y and \hat{f}\!(x;D): we want (y - \hat{f}\!(x;D))^2 to be minimal, both for x_1, \dots, x_n and for points outside of our sample. Of course, we cannot hope to do so perfectly, since the y_i contain noise \varepsilon; this means we must be prepared to accept an irreducible error in any function we come up with. Finding an \hat{f} that generalizes to points outside of the training set can be done with any of the countless algorithms used for supervised learning. It turns out that whichever function \hat{f} we select, we can decompose its expected error on an unseen sample x (i.e. conditional on x) as follows: \mathbb{E}_{D, \varepsilon} \Big[\big(y - \hat{f}\!(x;D)\big)^2\Big] = \Big(\operatorname{Bias}_D\big[\hat{f}\!(x;D)\big] \Big) ^2 + \operatorname{Var}_D \big[\hat{f}\!(x;D)\big] + \sigma^2 where \begin{align} \operatorname{Bias}_D\big[\hat{f}\!(x;D)\big] &\triangleq \mathbb{E}_D\big[\hat{f}\!(x;D)- f(x)\big]\\ &= \mathbb{E}_D\big[\hat{f}\!(x;D)\big] \, - \, f(x)\\ &= \mathbb{E}_D\big[\hat{f}\!(x;D)\big] \, - \, \mathbb{E}_{y|x}\big[y(x)\big] \end{align} and \operatorname{Var}_D\big[\hat{f}\!(x;D)\big] \triangleq \mathbb{E}_D \left[ \left( \mathbb{E}_D[\hat{f}\!(x;D)] - \hat{f}\!(x;D) \right)^2 \right] and \sigma^2 = \operatorname{E}_y \Big[ \big( y - \underbrace{f(x)}_{E_{y|x}[y]} \big)^2 \Big] The expectation ranges over different choices of the training set D=\{(x_1,y_1) \dots, (x_n, y_n)\}, all sampled from the same joint distribution P(x,y) which can for example be done via bootstrapping. The three terms represent: • the square of the bias of the learning method, which can be thought of as the error caused by the simplifying assumptions built into the method. E.g., when approximating a non-linear function f(x) using a learning method for linear models, there will be error in the estimates \hat{f}\!(x) due to this assumption; • the variance of the learning method, or, intuitively, how much the learning method \hat{f}\!(x) will move around its mean; • the irreducible error \sigma^2. Since all three terms are non-negative, the irreducible error forms a lower bound on the expected error on unseen samples. For convenience, we drop the D subscript in the following lines, such that \hat{f}\!(x;D) = \hat{f}\!(x). Let us write the mean-squared error of our model: \begin{align} \text{MSE}(x) &\triangleq \mathbb{E}\Big[\big(y - \hat{f}\!(x)\big)^2\Big]\\ &= \mathbb{E}\Big[\big(f(x) + \varepsilon - \hat{f}\!(x)\big)^2\Big] && \text{since } y \triangleq f(x) + \varepsilon\\ &= \mathbb{E}\Big[\big(f(x) - \hat{f}\!(x)\big)^2\Big] \, + \, 2 \ \mathbb{E}\Big[ \big(f(x) - \hat{f}\!(x)\big) \varepsilon \Big] \, + \, \mathbb{E}[\varepsilon^2] \end{align} We can show that the second term of this equation is null: \begin{align} \mathbb{E}\Big[ \big(f(x) - \hat{f}\!(x)\big) \varepsilon \Big] &= \mathbb{E} \big[ f(x) - \hat{f}\!(x) \big] \ \mathbb{E} \big[ \varepsilon \big] && \text{since } \varepsilon \text{ is independent from } x\\ &= 0 && \text{since } \mathbb{E} \big[ \varepsilon \big] = 0 \end{align} Moreover, the third term of this equation is nothing but \sigma^2, the variance of \varepsilon. Let us now expand the remaining term: \begin{align} & \operatorname\mathbb{E}\left[\left(f(x) - \hat{f}\!(x)\right)^2\right] \\[1ex] &= \operatorname\mathbb{E}\left[\left(f(x) - \operatorname\mathbb{E}[\hat{f}\!(x)] + \operatorname\mathbb{E}[\hat{f}\!(x)] - \hat{f}\!(x)\right)^2\right]\\[1ex] & = {\color{Blue} \operatorname\mathbb{E}\left[ \left( f(x) - \operatorname\mathbb{E}[\hat{f}\!(x)] \right)^2 \right]} \, + \, \operatorname\mathbb{E} \left[ \left( \operatorname\mathbb{E}[\hat{f}\!(x)] - \hat{f}\!(x) \right)^2 \right] \\ &\quad\, + \, 2 \ {\color{PineGreen} \operatorname\mathbb{E} \left[ \left( f(x) - \operatorname\mathbb{E}[\hat{f}\!(x)] \right) \left( \operatorname\mathbb{E}[\hat{f}\!(x)] - \hat{f}\!(x) \right) \right]} \end{align} We show that: \begin{align} {\color{Blue} \mathbb{E}\Big[ \big( f(x) - \mathbb{E} \big[ \hat{f}(x) \big] \big)^2 \Big]} &= \mathbb{E} \big[ f(x) ^2 \big] \, - \, 2 \ \mathbb{E} \Big[ f(x) \ \mathbb{E} \big[ \hat{f}(x) \big] \Big] \, + \, \mathbb{E} \Big[ \mathbb{E} \big[ \hat{f}(x) \big]^2 \Big]\\ &= f(x)^2 \, - \, 2 \ f(x) \ \mathbb{E} \big[ \hat{f}(x) \big] \, + \, \mathbb{E} \big[ \hat{f}(x) \big]^2\\ &= \Big( f(x) - \mathbb{E} \big[ \hat{f}(x) \big] \Big)^2 \end{align} This last series of equalities comes from the fact that f(x) is not a random variable, but a fixed, deterministic function of x. Therefore, \operatorname\mathbb{E}\left[f(x)\right] = f(x). Similarly \operatorname\mathbb{E}\left[ f(x)^2 \right] = f(x)^2, and \operatorname\mathbb{E} \left[ f(x) \, \operatorname\mathbb{E}[\hat{f}\!(x)] \right] = f(x) \, \operatorname\mathbb{E} \left[ \operatorname\mathbb{E}[\hat{f}\!(x)] \right] = f(x) \operatorname\mathbb{E}[\hat{f}\!(x)]. Using the same reasoning, we can expand the second term and show that it is null: \begin{align} &{\color{PineGreen} \operatorname\mathbb{E} \left[ \left( f(x) - \operatorname\mathbb{E}[\hat{f}\!(x)] \right) \left( \operatorname\mathbb{E}[\hat{f}\!(x)] - \hat{f}\!(x) \right) \right]} \\ &= \operatorname\mathbb{E} \left[ f(x)\, \operatorname\mathbb{E}[\hat{f}\!(x)] \, - \, f(x) \hat{f}\!(x) \, - \, \operatorname\mathbb{E}[\hat{f}\!(x)]^2 + \operatorname\mathbb{E}[\hat{f}\!(x)] \, \hat{f}\!(x) \right] \\ &= f(x) \, \operatorname\mathbb{E}[\hat{f}\!(x)] \, - \, f(x) \, \operatorname\mathbb{E}[\hat{f}\!(x)] \, - \, \operatorname\mathbb{E}[\hat{f}\!(x)]^2 \, + \, \operatorname\mathbb{E}[\hat{f}\!(x)]^2\\ &= 0 \end{align} Eventually, we plug our derivations back into the original equation, and identify each term: \begin{align} \text{MSE}(x) &= \left( f(x) - \operatorname\mathbb{E}[\hat{f}\!(x)] \right)^2 + \operatorname\mathbb{E} \left[ \left( \operatorname\mathbb{E}[\hat{f}\!(x)] - \hat{f}\!(x) \right)^2 \right] + \sigma^2\\ &= \operatorname{Bias}\left[\hat{f}\!(x) \right]^2 + \, \operatorname{Var} \left[ \hat{f}\!(x) \right] \, + \, \sigma^2 \end{align} Finally, the MSE loss function (or negative log-likelihood) is obtained by taking the expectation value over x\sim P: \text{MSE} = \operatorname\mathbb{E}_x \left[ \text{MSE}(x) \right] = \operatorname\mathbb{E}_x \left\{\operatorname{Bias}_D\!\left[\hat{f}\!(x;D)\right]^2 + \operatorname{Var}_D\left[\hat{f}\!(x;D)\right]\right\} + \sigma^2. ==Approaches==

Dimensionality reduction and feature selection can decrease variance by simplifying models. Similarly, a larger training set tends to decrease variance. Adding features (predictors) tends to decrease bias, at the expense of introducing additional variance. Learning algorithms typically have some tunable parameters that control bias and variance; for example, • linear and Generalized linear models can be regularized to decrease their variance at the cost of increasing their bias. • In artificial neural networks, the variance increases and the bias decreases as the number of hidden units increase, although this classical assumption has been the subject of recent debate. • In decision trees, the depth of the tree determines the variance. Decision trees are commonly pruned to control variance. For example, boosting combines many "weak" (high bias) models in an ensemble that has lower bias than the individual models, while bagging combines "strong" learners in a way that reduces their variance. Model validation methods such as cross-validation (statistics) can be used to tune models so as to optimize the trade-off. k-nearest neighbors In the case of -nearest neighbors regression, when the expectation is taken over the possible labeling of a fixed training set, a closed-form expression exists that relates the bias–variance decomposition to the parameter : \operatorname\mathbb{E}\left[\left(y - \hat{f}\!(x)\right)^2\mid X=x\right] = \left( f(x) - \frac{1}{k}\sum_{i=1}^k f(N_i(x)) \right)^2 + \frac{\sigma^2}{k} + \sigma^2 where N_1(x), \dots, N_k(x) are the nearest neighbors of in the training set. The bias (first term) is a monotone rising function of , while the variance (second term) drops off as is increased. In fact, under "reasonable assumptions" the bias of the first-nearest neighbor (1-NN) estimator vanishes entirely as the size of the training set approaches infinity. ==Applications==

In regression The bias–variance decomposition forms the conceptual basis for regression regularization methods such as LASSO and ridge regression. Regularization methods introduce bias into the regression solution that can reduce variance considerably relative to the ordinary least squares (OLS) solution. Although the OLS solution provides non-biased regression estimates, the lower variance solutions produced by regularization techniques provide superior MSE performance. In classification The bias–variance decomposition was originally formulated for least-squares regression. For the case of classification under the 0-1 loss (misclassification rate), it is possible to find a similar decomposition, with the caveat that the variance term becomes dependent on the target label. Alternatively, if the classification problem can be phrased as probabilistic classification, then the expected cross-entropy can instead be decomposed to give bias and variance terms with the same semantics but taking a different form. It has been argued that as training data increases, the variance of learned models will tend to decrease, and hence that as training data quantity increases, error is minimised by methods that learn models with lesser bias, and that conversely, for smaller training data quantities it is ever more important to minimise variance. In reinforcement learning Even though the bias–variance decomposition does not directly apply in reinforcement learning, a similar tradeoff can also characterize generalization. When an agent has limited information on its environment, the suboptimality of an RL algorithm can be decomposed into the sum of two terms: a term related to an asymptotic bias and a term due to overfitting. The asymptotic bias is directly related to the learning algorithm (independently of the quantity of data) while the overfitting term comes from the fact that the amount of data is limited. In Monte Carlo methods While in traditional Monte Carlo methods the bias is typically zero, modern approaches, such as Markov chain Monte Carlo are only asymptotically unbiased, at best. Convergence diagnostics can be used to control bias via burn-in removal, but due to a limited computational budget, a bias–variance trade-off arises, leading to a wide range of approaches, in which a controlled bias is accepted, if this allows to dramatically reduce the variance, and hence the overall estimation error. In human learning While widely discussed in the context of machine learning, the bias–variance dilemma has been examined in the context of human cognition, most notably by Gerd Gigerenzer and co-workers in the context of learned heuristics. They have argued (see references below) that the human brain resolves the dilemma in the case of the typically sparse, poorly-characterized training-sets provided by experience by adopting high-bias/low variance heuristics. This reflects the fact that a zero-bias approach has poor generalizability to new situations, and also unreasonably presumes precise knowledge of the true state of the world. The resulting heuristics are relatively simple, but produce better inferences in a wider variety of situations. Geman et al. argue that the bias–variance dilemma implies that abilities such as generic object recognition cannot be learned from scratch, but require a certain degree of "hard wiring" that is later tuned by experience. This is because model-free approaches to inference require impractically large training sets if they are to avoid high variance. ==See also==