class: top, left, title-slide .title[ # Machine learning ] .subtitle[ ## Overfitting ] .author[ ### Joshua Loftus ] --- class: inverse, middle, center <style type="text/css"> .remark-slide-content { font-size: 1.2rem; padding: 1em 4em 1em 4em; } </style> # What *is* overfitting? ### You've probably heard about it ### Almost halfway through the term... -- ### I've avoided it intentionally I wanted to do justice to this extremely important, central concept of machine learning And I'm dissatisfied with the usual definitions! --- class: inverse, center, middle # Machine learning ## is about algorithms ### that allow us to increase model complexity and optimize on larger datasets over larger sets of parameters and even infinite-dimensional function spaces 📈 🧮 📈 🖥 📈 💻 📈 📲 📈 📡 📈 --- ### With great fitting comes great responsibility ML increases danger of this specific kind of modeling problem "[I know it when I see it](https://www.youtube.com/watch?v=sIlNIVXpIns)" not good enough! .pull-left[  Source: [Wikipedia](https://en.wikipedia.org/wiki/Overfitting) ] .pull-right[  Example in week 1 ] ??? difficult to give one precise definition of overfitting that correctly captures all cases instead, we'll try out a few definitions and consider different types of overfitting --- class: inverse ## Defining/exploring overfitting We'll consider several approaches to defining/using the term - **Validation**: typical machine learning definition - **Generalization**: simple probabilistic definition -- - **Bias-variance**: statistical distinctions - **Causality**: scientific/philosophical applications - **Anthropology**: human learning and overfitting IRL (*non-examinable but brief and useful for life in general?*) ??? ML def: wikipedia, most textbooks, learn anything more than this and a bit about the bias-variance trade off and you'll be above average --- class: inverse, center, middle # Typical ML definition of overfitting Motivating idea: assume the model will be "deployed" -- I.e. Some time after fitting the model will be used on new data > "It is difficult to make predictions, especially about the future" - [Danish saying](https://quoteinvestigator.com/tag/niels-bohr/) --- # Overfitting the "training data" Using a model that is *too complex* Specifically, one where the complexity is larger than the optimal complexity for predicting on a *new observation* or *new sample* of **test/validation data** -- - `\(\hat f_\lambda\)` model fitted/estimated on **training data** - `\(\lambda\)` tuning parameter that *penalizes* complexity - larger `\(\lambda\)`, simpler model -- - `\(\lambda^*\)` optimal param. value for predicting/classifying *new data* - Overfitting: using `\(\hat f_\lambda\)` for some `\(\lambda < \lambda^*\)` --- ## ISLR Figure 2.10 .pull-left[ True model is simple ] .pull-right[ High complexity overfits ]  --- ## ISLR Figure 2.11 .pull-left[ True model is complex ] .pull-right[ Low complexity underfits ]  --- # The end Some discussions of overfitting end there -- .pull-left[ Some go on a little more, relating it to **bias-variance** trade-off Overfitting: *low bias but overwhelmingly high-variance* (we'll do that soon) ] .pull-right[  ] --- class: inverse, center, middle # Generalization and a ## probabilistic definition Motivation: what is the *probability distribution* of the test data? --- # Two kinds of generalization ML/AI books/courses talk about "generalization error" Over-used term, same word / *importantly different* meanings -- ### Generalization to a new observation from... - the same distribution or DGP - a different (but related) distribution and *corresponding reasons for doing poorly* - variance ("random/unstructured error", high entropy) - bias ("systematic/structured error", low entropy) --- ### Think about distributions Suppose the training data is sampled i.i.d. from `\((\mathbf X_1, y_1), (\mathbf X_2, y_2), \ldots, (\mathbf X_n, y_n) \sim F\)` and the test data is sampled i.i.id from `\((\mathbf X'_1, y'_1), (\mathbf X'_2, y'_2), \ldots, (\mathbf X'_{n'}, y'_{n'}) \sim F'\)` -- #### In-distribution (ID) generalization: `\(F = F'\)` Under/overfitting, **variability problem**, larger `\(n\)` allows more complex models to be fit -- #### Out-of-distribution (OOD) generalization: `\(F \neq F'\)` "Covariate/distribution/dataset shift", **bias problem**, larger `\(n\)` may not help. Need modeling **assumptions** like `\(F' \approx F\)` --- ## Optimism and ID generalization Observation: training error generally appears lower than test/validation error. Why? -- Risk vs *empirical* risk minimization `$$R(g) = \mathbb E_F[L(\mathbf X,Y,g)]$$` `$$\hat f = \arg \min_g \hat R(g) = \arg \min_g \frac{1}{n} \sum_{i=1}^n L(\mathbf x_i, y_i, g)$$` -- **Fact**: for some `\(\text{df}(\hat f) > 0\)` (depends on problem/fun. class) `$$\mathbb E_{Y|\mathbf x_1, \ldots, \mathbf x_n}[R(\hat f) - \hat R(\hat f)] = \frac{2\sigma^2_\varepsilon}{n} \text{df}(\hat f) > 0$$` --- ### Optimism, ID gen., and degrees of freedom #### Linear case If `\(\hat f\)` is linear with `\(p\)` predictors (or `\(p\)` basis function transformations of original predictors) then $$ \text{df}(\hat f) = p $$ -- #### Fairly general case For many ML tasks and fitting procedures `$$\text{df}(\hat f) \text{ increases as } \frac{1}{n \sigma^2_\varepsilon} \sum_{i=1}^n \text{Cov}(\hat f (\mathbf x_i), y_i) \text{ increases}$$` --- ### Take-aways about optimism and ID gen. - Empirical risk *underestimates* actual risk (ID generalization error) - The magnitude of this bias is called **optimism** -- - Optimism generally increases with function class complexity - e.g. for linear functions, increases linearly in `\(p\)` - For a fixed function class, optimism decreases linearly in `\(n\)` -- - Too much optimization `\(\to\)` overfitting `\(\to\)` more optimism --- ## Two kinds of overfitting? Many sources identify overfitting as a threat to generalization Typically only apply this to **ID generalization**, and have solution strategies to avoid the **variability problems** due to overfitting -- #### But overfitting is also a threat to **OOD generalization**! This kind of generalization is often what we practically want There are serious **bias problems** due to overfitting -- #### Let's start using new terminology *Overfitting to variation* and *overfitting to bias* --- ### Now let's jump from probability to statistics And talk about why we always need to care about *both kinds of generalization*  (Feel free to take a break while pondering overfitting to variation vs bias) --- class: inverse, center, middle # Statistical aspects of overfitting Motivation: all models are wrong, including `\(F\)` and `\(F'\)` or Motivation: overfitting to noise... what's noise? --- # What *is* noise? The effect of all (causal?) factors not captured by the model Could be different reasons for failing to capture - Measurement issues - Wrong functional relationship - Variables excluded (maybe not even measured or defined) *Does not require physical randomness* (which maybe doesn't exist...) -- Something considered **noise** in one setting, or by one modeler, could be **signal** to a different observer --- ### Noise and residuals in regression (*One of the most "agnostic" or minimal-theory ways of defining regression is as estimation of a conditional expectation function, without assuming any specific functional form like linearity*). The "noise" in regression is defined as $$ \varepsilon = y - \mathbb E_F[y | \mathbf x] $$ -- But this is math, not applied data analysis! Requires assuming a probability distribution `\(F\)` / random variable model Otherwise, how do we define expectation? -- We never observe `\(\varepsilon\)`, only residuals `\(r_i = y_i - \hat f(\mathbf x_i)\)` of some model `\(\hat f\)` *fit with specific assumptions/algorithms* --- # What is bias/variability? Two analysts start with different assumptions e.g. linearity vs flexible non-parametric methods Fit different regression functions Compute different residuals See different patterns (or lack thereof) in residuals -- Something considered **variation** in one setting, or by one modeler, could be **bias** to a different observer --- ## Data science In the "real world" there is a data generation process (DGP) We *assume* this can be modeled as an i.i.d. sample from a probability distribution `\(F\)` Probability model / mathematical justification for our methods #### All models are wrong -- Could model DGP as a mixture of distributions `\(F\)` and `\(F'\)` (heterogeneity), or time-varying `\(F^t\)` Training/test data randomly shuffled? Generalization in/out of distribution? --- ## Two data scientists diverged... Starting with different assumptions about DGP Use different strategies to avoid overfitting e.g. different ways of splitting into training and test data Something considered **ID generalization** in one setting, or by one modeler, could be **OOD generalization** to a different observer --- ## Statistical take-aways Mathematical distinctions between ID and OOD generalization rely on assumptions (as do statistical distinctions between bias and variability) ML methods for avoiding overfitting are motivated by ID generalization, guard against **overfitting to variability** -- In applications, ID/OOD distinctions break down. If we probe them a bit we find it's more gray area / ambiguous Most scientists and decision-makers care about [external validity](https://en.wikipedia.org/wiki/External_validity), conceptually related to OOD generalization **Overfitting to bias** is a serious, widely neglected problem! --- class: inverse, center, middle # Considerations of the ## scientific and philosophic variety ### with respect to overfitting Motivation: does science overfit? Can philosophy of science help us understand how to prevent it? What about causality? --- ### Stability, invariance, and causality **Idea**: causal relationships should persist despite (some) [changes](https://plato.stanford.edu/entries/causation-counterfactual/#ConInv) in "[background](https://static1.squarespace.com/static/54023011e4b00d99c7bb97f9/t/547b5b1fe4b0464e56d755b4/1417370399799/PoS1997.pdf) [conditions](https://static1.squarespace.com/static/54023011e4b00d99c7bb97f9/t/5445ceb4e4b06ae793389dad/1413861044101/BandP2010.pdf)" [Bradford Hill criteria](https://en.wikipedia.org/wiki/Bradford_Hill_criteria) for causation > **Consistency**: Has [the observed association] been repeatedly observed by different persons, in different places, circumstances and times? -- Apparently [people think about causality](https://onlinelibrary.wiley.com/doi/pdf/10.1111/cogs.12605) this way Can use the idea to motivate [statistical methods](https://rss.onlinelibrary.wiley.com/doi/full/10.1111/rssb.12167) for causal inference --- ### Overfitting as a threat to causal inference [Bradford Hill criteria](https://en.wikipedia.org/wiki/Bradford_Hill_criteria) for causation > "the larger the association, the more likely that it is causal." - Wikipedia, not Hill Hill: > the death rate from cancer of the lung in cigarette smokers is nine to ten times the rate in non-smokers -- **Problem**: overfitting can make associations appear stronger e.g. proportion of variation in `lifeExp` explained by `gdpPercap` Increase flexibility, explain higher proportion... stronger evidence of causality? 🤔 --- ### Generalization, novelty, and severity [Philosophy of science](https://plato.stanford.edu/entries/prediction-accommodation/): prediction vs "accommodation" Prediction: happens in time before observation/measurement Accommodation: theory built to explain past observation/data Accurate prediction is better evidence in favor of a scientific theory than mere accommodation ML: What's better evidence in favor of the model? -- Popper and Lakatos: **temporal novelty** Zahar, Gardner, Worrall: **use-novelty** (or problem novelty) [Mayo](http://bactra.org/reviews/error/): novelty is not necessary. **Severity** is necessary --- class: inverse, center, middle # Anthropology? ## ie. overfitting IRL (in real life) Motivation: do *we* overfit? ("Are we the baddies?") Disclaimer: I am not an anthropologist *or* self-help author --- ## How/why are humans different? We seem to better at *learning* than other animals [Human eyes are different](https://en.wikipedia.org/wiki/Cooperative_eye_hypothesis), allowing us to see where others are looking -- ### Social learning #### "Monkey see, monkey do" Lots of animals learn by [imitation](https://en.wikipedia.org/wiki/Imitation#Over-imitation), but [humans specifically](https://psycnet.apa.org/fulltext/1993-44247-001.pdf) take imitation to a [whole](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0159920) [different](https://www.pnas.org/content/104/50/19751) level #### Over-imitation, causal opacity, cultural evolution...