beyond-ML.qmd

# [Beyond machine learning]{.midgrey} {#sec-beyond-ML}
{{< include macros.qmd >}}
{{< include macros_prob_inference.qmd >}}
{{< include macros_connection-3.qmd >}}

## Machine learning from a bird's-eye view {#sec-ML-birds-eye}

The last few chapters gave a brief introduction to and overview of popular machine-learning methods, their terminology, and the points of view that they typically adopt. Now let's try to look at them keeping in mind our main goal in this course: [exploring new methods, understanding their foundations, and thinking out of the box](index.html).

In this and the next few chapters we shall focus on the question: [*to what purpose do we use machine-learning algorithms?*]{.yellow}. After answering this question, we shall try to achieve that purpose in the optimal way, according to what our fundamental theory tells us we should do, without considering "machine learning". But we shall keep an eye on where our optimal solution seems to be similar or dissimilar to machine-learning methods.

Then, in the last chapters, we shall examine where the optimal solution and machine-learning methods converge and diverge, try to understand what machine-learning methods do from the point of view of the optimal solution, and think of ways to improve them.

## A task-oriented categorization of some machine-learning problems {#sec-cat-problems}

For our goal, the common machine-learning categorization and terminology discussed in [chapter @sec-ml-introduction] are somewhat inadequate. Distinctions such as "supervised learning" vs "unsupervised learning" are of secondary importance to a data engineer (as opposed to a ["data mechanic"](preface.html)) for several reasons:

- {{< fa shuffle >}}\ \ They group together some types of tasks that are actually quite different from an inferential or decision-making viewpoint; and conversely they separate types of tasks that are quite similar.

- {{< fa bullseye >}}\ \ They focus on procedures rather than on purposes.

The important questions for us, in fact, are: [*What do we wish to infer or choose?*]{.blue} and [*From which kind of information?*]{.green} These questions define the problem we want to solve. The procedure may then be chosen depending on the theory, resources and technologies, other contingent factors, and so on.

<!-- It's somewhat like saying that the difference between car and aeroplane is that the latter has wings. Sure -- but *why?* The focus on wings misses the essential difference between these two means of transportation: they operate through different material media and exploit different kinds of physics; that's why the second has wings. -->

Let's introduce a different categorization that tries to focus on the purpose or task, on the types of desired information and of available information, rather than on the procedure.



The categorization below, of the types of *task* that machine-learning algorithms try to solve, is informal. It only provides a starting point from which to examine a new task. Many tasks will fall in between categories: every data-engineering or data-science problem is unique.

\

We *exclude* from the start all tasks that require an agent to continuously and actively interact with its environment for acquiring information, making choices, getting feedback, and so on. Clearly these tasks are the domain of Decision Theory in its most complex form, with ramified decisions, *strategies*, and possibly the interaction with other decision-making agents. To explore and analyse this kind of tasks is beyond the purpose of this course.

::::{.column-margin}
::: {.callout-tip}
## {{< fa rocket >}} For the extra curious
- [*Decision Analysis*](https://hvl.instructure.com/courses/28605/modules)

- Chapters 16--18 in [*Artificial Intelligence*](https://hvl.instructure.com/courses/28605/modules)

- [*Games and Decisions*](https://hvl.instructure.com/courses/28605/modules)
:::
::::

\

We focus on tasks where multiple "instances" with similar characteristics are involved, and the agent has some task related to a "new instance", possibly to be repeated an indefinite number of times. According to the conceptual framework developed in part [Data II]{.yellow}, we can view these "instances" as *units* of a practically infinite population. The "characteristics" that the agent guesses or observed are *variates* common to all these units.

:::{.column-margin}
Remember that you can adopt any terminology you like. If you prefer "instance" and "characteristics" or some other words to "unit" and "variate", then use them. What's important is that you understand the ideas and methods behind these words
:::


### New unit: given vs generated

A first distinction can be made between

- [{{< fa sign-out-alt >}} {{< fa cube >}}\ \ tasks where an agent must itself *generate* a new unit]{.yellow}

- [{{< fa cube >}} {{< fa question >}}\ \ tasks where a new unit is given to an agent, who must *guess* some of its variates]{.green}

An example of the first type of task is image generation: an algorithm is given a collection of images and is asked to generate a new image based on them.

We shall see that these two types of task are actually quite close to each other, from the point of view of Decision Theory and Probability Theory.

:::{.small .midgrey}
The terms "discriminative" and "generative" are sometimes associated in machine learning with the two types of task. This association, however, is quite loose, because some tasks typically called "generative" actually belong to the first type. We shall therefore avoid these or other terms. It's enough to keep in mind the distinction between the two types of task above.
:::


### Guessing variates: all or some

Focusing on the second type of task (a new unit is given to the agent), we can further divide it into two subtypes:

- [{{< fa regular star >}}\ \ the agent must guess *all* variates of the new unit]{.purple}

- [{{< fa star-half-alt >}}\ \ the agent must guess *some* variates of the new unit, but can observe other variates of the new unit]{.red}

An example of the first subtype of task is the "urgent vs non-urgent" problem of [§@sec-conditional-joint-sim]: having observed incoming patients, some of which where urgent and some non-urgent, the agent must guess whether the next incoming patient will be urgent or not. No other kinds of information (transport, patient characteristics, and so on) are available for any patient.

We shall call [**predictands**]{.blue}^[literally "what has to be predicted] the variates that the agent must guess in a new unit, and [**predictors**]{.blue} those that the agent can observe.^[In machine learning and other fields, the terms "dependent variable", "class" or "label" (for nominal variates) are often used for "predictand"; and the terms "independent variable" or "features" are often used for "predictor".] The first subtype above can be viewed as a special case of the second where all variates are predictands, and there are no predictors.

:::{.small .midgrey}
The terms "unsupervised learning" and "supervised learning" are sometimes associated in machine learning with these two subtypes of task. But also in this case the association is loose and can be misleading. "Clustering" tasks, for example, are usually called "unsupervised" but they are examples of the second subtype above, where the agent has some predictors.
:::

### Information available in previous units

Finally we can further divide the second subtype above into two or three subsubtypes, depending on the information available to the agent about *previous units*:

- [{{< fa star-half-alt >}} {{< fa star-half-alt >}}\ \ all predictors and predictands of previous units are known to the agent]{.blue}

- [{{< fa star-half >}} {{< fa star-half >}}\ \ all predictors of previous units, but not the predictands, are known to the agent]{.lightblue}

- [{{< fa regular star-half >}} {{< fa regular star-half >}}\ \ all predictands of previous units, but not the predictors, are known to the agent]{.midgrey}

\

An example of the first subsubtype of task is image classification. The agent is for example given the following 128 × 128-pixel images and character-labels from the [One Punch Man](https://onepunchman.fandom.com) series:

![](saitama_images.png){width=100%}

and is then given one new 128 × 128-pixel image:

![](saitama_new.png){width=128 fig-align="center"}

of which it must guess the character-label.

In the example just given, the image is the predictor, the character-label is the predictand.

\

A slight modification of the example above gives us a task of the second subsubtype. A different agent is given the images above, but without labels:

![](saitama_images_nolabels.png){width=100%}

and must then guess some kind of "label" or "group" for the new image above -- and possibly even for the images already given. The kind of "group" requested depends on the specific problem.

In the new example above, the image is still the predictor, and the label or group is the predictand.

:::{.small .midgrey}
The term "supervised learning" typically refer to the first subsubtype above.

The term "unsupervised learning" can refer to the second subsubtype, for instance in "clustering" tasks. In a clustering task, the agent tries to guess which group or "cluster" a unit belong to, given a collection of similar units, whose groups are not known either. The cluster effectively is the *predictand* variate. In some cases the agent may want to guess the cluster not only of a new unit, but also of all previous units.

The third subsubtype is very rarely considered in machine learning, yet it is not an unrealistic task.
:::

The types, subtypes, subsubtypes above are obviously not mutually exclusive or comprehensive. We can easily imagine scenarios where an agent has some predictors & predictands available about *some* previous units, but only predictors or only predictands available for other previous units. This scenario falls in between the three subsubtypes above. In machine learning, hybrid situations like these are categorized as "missing data" or "imputation".


## Flexible categorization using probability theory {#sec-categ-probtheory}

We have been speaking about the agent's "guessing" the values of some variates. Guessing means that there's a state of uncertainty; the agent can't just say something like "the value is $42$". Uncertainty means that the most honest thing that the agent can do is to express degrees of belief about each of the possible values. Probability theory must enter the scene.

But it also turns out that the categorization above into subtypes and subsubtypes of task can actually be presented in a more straightforward and flexible way using probability-theory notation.

### Notation

First let's introduce some symbol conventions to be used in the next chapters. We shall denote with $\bZ$ all variates that are of interest to the agent: those to be guessed and those that may be known. The variates to be guessed in a new unit (the predictands) will be collectively denoted with $\bY$. The variates that can be observed in a new unit (the predictors) will be collectively denoted with $\bX$. Therefore we have $\bZ = (\bY \and \bX)$. In cases where there are no predictors, $\bX$ is empty and we have $\bZ = \bY$.

$\bZ_i$, $\bY_i$, $\bX_i$ denote all variates, the predictands, and the predictors for unit [#$i$.]{.m}\ \ As usual we number from $i=1$ to $i=N$ the units that serve for learning, and $i=N+1$ is the new unit of interest to the agent.

Recall ([§@sec-basic-elements-inference]) that in probability notation $\P({\red\boldsymbol{\therefore}}\|{\green\boldsymbol{\because}} \and \yI)$, the [proposal]{.red} contains what the agent's belief is about, and the [conditional]{.green} contains what's supposed to be known to the agent, together with the background [information $\yI$.]{.m}

\

### [{{< fa regular star >}}]{.purple}\ \ The agent must guess *all* variates of the new unit

This kinds of guess is represented by the probability distribution

$$
\P(\red
Z_{N+1}\mo z
\black\|
\green
Z_{N}\mo z_{N}
\and \dotsb \and
Z_{1}\mo z_{1}
\black \and \yI)
$$

for all possible values $\red z$ in the domain of $\bZ$. The specific values $\green z_N, \dotsc, z_1$ of the variate $\bZ$ for the previous units are known to the agent.

\

### [{{< fa star-half-alt >}}]{.red}\ \ The agent must guess *some* variates of the new unit, but can observe other variates of the new unit

This kind of guess is represented by a probability distribution having

$$
\P(\red
Y_{N+1}\mo y
\black\|
\green
X_{N+1}\mo x
\, \and\, 
\dotsb \,
\black \and \yI)
$$

for all possible values $\red y$ in the domain of the predictands $\bY$. The value $\green x$ of the predictors $\bX$ for the new unit is known to the agent.

The remaining information contained in the conditional depends on the subsubtype of task previously discussed:

\

#### [{{< fa star-half-alt >}} {{< fa star-half-alt >}}]{.blue}\ \ All predictors and predictands of previous units are known to the agent

This corresponds to the probability distribution

$$
\P(\red
Y_{N+1}\mo y
\black\|
\green
X_{N+1}\mo x
\, \and\, 
Y_{N}\mo y_{N}
\and
X_{N}\mo x_{N}
\and \dotsb \and
Y_{1}\mo y_{1}
\and
X_{1}\mo x_{1}
\black \and \yI)
$$

for all possible $\red y$. All information about predictands and predictors for previous units appear in the conditional.

In the example with image classification, a pictorial representation of this probability would be

![](saitama_example2.png){width=100%}

where ${\red y} \in \set{\red\cat{Saitama}, \cat{Fubuki}, \cat{Genos}, \cat{MetalBat}, \dotsc \black}$.

\

#### [{{< fa star-half >}} {{< fa star-half >}}]{.lightblue}\ \ All predictors of previous units, but not their predictands, are known to the agent

This corresponds to the probability distribution

$$
\P(\red
Y_{N+1}\mo y
\black\|
\green
X_{N+1}\mo x
\, \and\, 
X_{N}\mo x_{N}
\and \dotsb\and
X_{1}\mo x_{1}
\black \and \yI)
$$

for all possible $\red y$. All information about predictors for the previous units, but not that about their predictands, appear in the conditional.

\

### More general and hybrid tasks

Consider a task that doesn't fit into any of the types discussed above: The agent wants to guess the predictands for a new unit, say #3, after observing that its predictors have value $\green x$. Of two previous units, the agent knows the predictor value $\green x_1$ of the first, and the predictand value $\green y_2$ of the second. This task is expressed by the probability

$$
\P(\red
Y_{3}\mo y
\black\|
\green
X_{3}\mo x
\, \and\, 
Y_{2}\mo y_{2}
\and
X_{1}\mo x_{1}
\black \and \yI)
$$

\

::::{.column-page-inset-right}
:::{.callout-caution}
## {{< fa user-edit >}} Exercises

- Write down the general probability expression for the task of subsubtype "[all predictands of previous units, but not their predictors, are known to the agent]{.midgrey}".

- What kind of task does the following probability express?:

    $$
\P(\red
Y_{N+1}\mo y_{N+1}
\and Y_{N}\mo y_{N}
\and \dotsb \and
Y_{2}\mo y_{2}\and
Y_{1}\mo y_{1}
\black\|
\green
X_{N+1}\mo x_{N+1}
\and X_{N}\mo x_{N}
\and \dotsb\and
X_{2}\mo x_{2}
\and
X_{1}\mo x_{1}
\black \and \yI)
$$
    
	What kind of task could it represent in machine-learning terminology?

:::
::::

### [{{< fa sign-out-alt >}} {{< fa cube >}}]{.yellow}\ \ Tasks where an agent must itself *generate* a new unit

Our first categorical division mentioned the task of generating a new unit, given previous examples. In this kind of task there are possible alternatives that the agent could generate. How should one alternative be chosen? A moment's thought shows that the *probabilities* for the alternatives should enter the choice.

Suppose, as a very simple example, that a unit-generating agent has been shown, in an unsystematic order, 30 copies of the symbol [{{< fa regular circle-up >}}]{.green} and 10 copies of the symbol [{{< fa regular circle-down >}}]{.yellow}, and is asked to generate a new symbol out of these examples. Intuitively we expect it to generate [{{< fa regular circle-up >}}]{.green}, but cannot and don't want to exclude the possibility that it could generate [{{< fa regular circle-down >}}]{.yellow}. These two generation possibilities should simply have different probabilities and, in the long run, appear with different frequencies.

Also in this kind of task, therefore, we have the probability distribution

$$
\P(\red
Z_{N+1}\mo z
\black\|
\green
Z_{N}\mo z_{N}
\and \dotsb \and
Z_{1}\mo z_{1}
\black \and \yI)
$$

the difference from before is that the sentence $\red Z_{N+1}\mo z$ represents not the hypothesis that a *given* new unit has value $\red z$, but the possibility of *generating* a new unit with that value. In other words, the symbol "$\mo$" here means "*must be set to...*" rather than "*would be observed to be...*"; remember the discussion and warnings in [§@sec-sentence-notation]?

\

The general conclusion is that 

:::{.callout-note style="font-size:120%"}
##
::::{style="font-size:120%"}
probability distribution such as those discussed above must intrinsically enter all types of machine-learning algorithms.
::::
:::

## The underlying distribution {#sec-underlying-distribution}

A remarkable feature of all the probabilities discussed in the above task categorization, even of those for "hybrid" types of task, is that they can all be calculated from *one* and the same probability distribution. We briefly discussed and used this feature in [chapter @sec-learning].

A conditional probability such as $\P({\red\boldsymbol{\therefore}}\|{\green\boldsymbol{\because}} \and \yI)$ can always be written, by the `and`-rule, as the ratio of two probabilities:

$$
\P({\red\boldsymbol{\therefore}}\|{\green\boldsymbol{\because}} \and \yI)
=
\frac{
\P({\red\boldsymbol{\therefore}}\and {\green\boldsymbol{\because}} \| \yI)
}{
\P({\green\boldsymbol{\because}} \| \yI)
}
$$

Therefore we have, for the probabilities of some of the tasks above,

:::{.column-page-inset-right}
$$
\begin{aligned}
&\P(\red
Z_{N+1}\mo z
\black\|
\green
Z_{N}\mo z_{N}
\and \dotsb \and
Z_{1}\mo z_{1}
\black \and \yI)
=
\frac{
\P(\red
Z_{N+1}\mo z
\black\and
\green
Z_{N}\mo z_{N}
\and \dotsb \and
Z_{1}\mo z_{1}
\black \| \yI)
}{
\P(
\green
Z_{N}\mo z_{N}
\and \dotsb \and
Z_{1}\mo z_{1}
\black \| \yI)
}
\\[2em]
&\P(\red
Y_{N+1}\mo y
\black\|
\green
X_{N+1}\mo x
\, \and\, 
Y_{N}\mo y_{N}
\and
X_{N}\mo x_{N}
\and \dotsb \and
Y_{1}\mo y_{1}
\and
X_{1}\mo x_{1}
\black \and \yI)
\\[2ex]
&\qquad{}=
\frac{
\P(\red
Y_{N+1}\mo y
\black\and
\green
X_{N+1}\mo x
\, \and\, 
Y_{N}\mo y_{N}
\and
X_{N}\mo x_{N}
\and \dotsb \and
Y_{1}\mo y_{1}
\and
X_{1}\mo x_{1}
\black \| \yI)
}{
\P(
\green
X_{N+1}\mo x
\, \and\, 
Y_{N}\mo y_{N}
\and
X_{N}\mo x_{N}
\and \dotsb \and
Y_{1}\mo y_{1}
\and
X_{1}\mo x_{1}
\black \| \yI)
}
\\[2em]
&\P(\red
Y_{N+1}\mo y
\black\|
\green
X_{N+1}\mo x
\, \and\, 
X_{N}\mo x_{N}
\and \dotsb\and
X_{1}\mo x_{1}
\black \and \yI)
\\[2ex]
&\qquad{}=
\frac{
\P(\red
Y_{N+1}\mo y
\black\and
\green
X_{N+1}\mo x
\, \and\, 
X_{N}\mo x_{N}
\and \dotsb\and
X_{1}\mo x_{1}
\black \| \yI)
}{
\P(
\green
X_{N+1}\mo x
\, \and\, 
X_{N}\mo x_{N}
\and \dotsb\and
X_{1}\mo x_{1}
\black \| \yI)
}
\end{aligned}
$$
:::

\

We also know the marginalization rule ([chapter @sec-marginal-probs]): any quantity $\yellow A$ with values $\yellow a$ can be introduced into the proposal of a probability via the `or`-rule:

$$
\P( {\green\boldsymbol{\dotsb}} \| \yI) =
\sum_{\yellow a}\P({\yellow A\mo a} \and  {\green\boldsymbol{\dotsb}} \| \yI)
$$

Using the marginalization rule we find these final expressions for the probabilities of tasks of various types:

::::{.column-page-right}
:::{.callout-note}
##

- [{{< fa regular star >}}]{.purple} Guess all variates:

$$
\P(\red
Z_{N+1}\mo z
\black\|
\green
Z_{N}\mo z_{N}
\and \dotsb \and
Z_{1}\mo z_{1}
\black \and \yI)
=
\frac{
\P(\red
Z_{N+1}\mo z
\black\and
\green
Z_{N}\mo z_{N}
\and \dotsb \and
Z_{1}\mo z_{1}
\black \| \yI)
}{
\sum_{\purple z}
\P(
\red
Z_{N+1}\mo {\purple z}
\black\and\green
Z_{N}\mo z_{N}
\and \dotsb \and
Z_{1}\mo z_{1}
\black \| \yI)
}
$$

\

- [{{< fa star-half-alt >}} {{< fa star-half-alt >}}]{.blue} All previous predictors and predictands known:

$$
\begin{aligned}
&\P(\red
Y_{N+1}\mo y
\black\|
\green
X_{N+1}\mo x
\, \and\, 
Y_{N}\mo y_{N}
\and
X_{N}\mo x_{N}
\and \dotsb \and
Y_{1}\mo y_{1}
\and
X_{1}\mo x_{1}
\black \and \yI)
\\[2ex]
&\qquad{}=
\frac{
\P(\red
Y_{N+1}\mo y
\black\and
\green
X_{N+1}\mo x
\, \and\, 
Y_{N}\mo y_{N}
\and
X_{N}\mo x_{N}
\and \dotsb \and
Y_{1}\mo y_{1}
\and
X_{1}\mo x_{1}
\black \| \yI)
}{
\sum_{\purple y}
\P(\red
Y_{N+1}\mo {\purple y}
\black\and
\green
X_{N+1}\mo x
\, \and\, 
Y_{N}\mo y_{N}
\and
X_{N}\mo x_{N}
\and \dotsb \and
Y_{1}\mo y_{1}
\and
X_{1}\mo x_{1}
\black \| \yI)
}
\end{aligned}
$$

\

- [{{< fa star-half >}} {{< fa star-half >}}]{.lightblue} Previous predictors known, previous predictands unknown:

$$
\begin{aligned}
&\P(\red
Y_{N+1}\mo y
\black\|
\green
X_{N+1}\mo x
\, \and\, 
X_{N}\mo x_{N}
\and \dotsb\and
X_{1}\mo x_{1}
\black \and \yI)
\\[2ex]
&\quad{}=
\frac{
\sum_{\yellow y_{N}, \dotsc, y_{1}}
\P(\red
Y_{N+1}\mo y
\black\and
\green
X_{N+1}\mo x
\black\, \and\, 
\yellow
Y_{N}\mo y_{N}
\black\and\green
X_{N}\mo x_{N}
\black\and \dotsb\and
\yellow
Y_{1}\mo y_{1}
\black\and\green
X_{1}\mo x_{1}
\black \| \yI)
}{
\sum_{{\purple y}, \yellow y_{N}, \dotsc, y_{1}}
\P(\red
Y_{N+1}\mo {\purple y}
\black\and
\green
X_{N+1}\mo x
\black\, \and\, 
\yellow
Y_{N}\mo y_{N}
\black\and\green
X_{N}\mo x_{N}
\black\and \dotsb\and
\yellow
Y_{1}\mo y_{1}
\black\and\green
X_{1}\mo x_{1}
\black \| \yI)
}
\end{aligned}
$$

<!-- $$ -->
<!-- \begin{aligned} -->
<!-- &\P(\red -->
<!-- Y_{4}\mo y -->
<!-- \black\| -->
<!-- \green -->
<!-- X_{4}\mo x -->
<!-- \, \and\, -->
<!-- Y_{3}\mo y_{3} -->
<!-- \and X_{3}\mo x_{3} -->
<!-- \and \ -->
<!--  X_{2}\mo x_{2} -->
<!-- \and -->
<!-- Y_{1}\mo y_{1} -->
<!-- \and X_{1}\mo x_{1} -->
<!-- \black \and \yI) -->
<!-- \\[2ex] -->
<!-- &\quad{}= -->
<!-- \frac{ -->
<!-- \sum_{\yellow y_{2}} -->
<!-- \P(\red -->
<!-- Y_{4}\mo y -->
<!-- \black\and -->
<!-- \green -->
<!-- X_{4}\mo x -->
<!-- \, \and\, -->
<!-- Y_{3}\mo y_{3} -->
<!-- \and X_{3}\mo x_{3} -->
<!-- \and -->
<!-- {\yellow Y_{2}\mo y_{2}} -->
<!-- \and -->
<!--  X_{2}\mo x_{2} -->
<!-- \and -->
<!-- Y_{1}\mo y_{1} -->
<!-- \and X_{1}\mo x_{1} -->
<!-- \black \| \yI) -->
<!-- }{ -->
<!-- \sum_{\purple y} -->
<!-- \sum_{\yellow y_{2}} -->
<!-- \P(\red -->
<!-- Y_{4}\mo {\purple y} -->
<!-- \black\and -->
<!-- \green -->
<!-- X_{4}\mo x -->
<!-- \, \and\, -->
<!-- Y_{3}\mo y_{3} -->
<!-- \and X_{3}\mo x_{3} -->
<!-- \and -->
<!-- {\yellow Y_{2}\mo y_{2}} -->
<!-- \and -->
<!--  X_{2}\mo x_{2} -->
<!-- \and -->
<!-- Y_{1}\mo y_{1} -->
<!-- \and X_{1}\mo x_{1} -->
<!-- \black \| \yI) -->
<!-- } -->
<!-- \end{aligned} -->
<!-- $$ -->

:::
::::

**All these formulae, even for hybrid tasks, involve sums and ratios of only one distribution:**

$$\boldsymbol{
\P(\blue
Y_{N+1}\mo y_{N+1}
\and
X_{N+1}\mo x_{N+1}
\and \dotsb \and
Y_{1}\mo y_{1}
\and
X_{1}\mo x_{1}
\black \| \yI)
}
$$

Stop for a moment and contemplate some of the consequences of this remarkable fact

- [{{< fa arrows-spin >}}\ \ *An agent that can perform one of the tasks above also can, in principle, perform all other tasks.*]{.blue}

    This is why a perfect agent, working with probability, in principle does not have to worry about "supervised", "unsupervised", "missing data", "imputation", and similar situations. This also shows what was briefly mentioned before: all these task typologies are much closer to one another than it might look like from the perspective of current machine-learning methods.
	
:::{.column-margin}
The acronym [*OPM*]{.green} ![](opm_fist2.png){height=2em} can stand for [*Optimal Predictor Machine*]{.green} or [*Omni-Predictor Machine*]{.green}
:::

- [{{< fa microchip >}}\ \ *The probability distribution above encodes the agent's background knowledge and assumptions; different agents differ only in the values of that distribution.*]{.blue}

    If two agents yield different probability values in the same task, with the same variates and same training data, the difference must come from the joint probability distribution above. And, since the data given to the agents are exactly the same, the difference is in the agents' background [information $\yI$.]{.m}

- [{{< fa user-secret >}}\ \ *Data cannot "speak for themselves"*]{.blue}

    For exactly the same data, we can choose two different distributions above , and therefore get different results. The data alone cannot determine the result: specific background information and assumptions, whether acknowledged or not, always affect the result.

The qualification "in principle" in the first consequence is important. Some of the sums that enter the formulae above are computationally extremely expensive and, with current technologies and maths techniques, cannot be performed within a reasonable time. But *new technologies and new maths discoveries could make these calculations suddenly possible*, this is why a data engineer cannot simply brush them aside and forget them.

As regards the third consequence, we shall see that there are states of knowledge which actually do converge to the same results, as the number of training data increases.



:::{.callout-caution}
## {{< fa user-edit >}} Exercise

In a previous example of "hybrid" task we had the probability distribution

$$
\P(\red
Y_{3}\mo y
\black\|
\green
X_{3}\mo x
\, \and\, 
Y_{2}\mo y_{2}
\and
X_{1}\mo x_{1}
\black \and \yI)
$$

Rewrite it in terms of the underlying joint distribution.
:::

## Plan for the next few chapters

Our goal in building an "Optimal Predictor Machine" is now clear: we must find a way to

:::{.column-margin}
![](optimal_predictor_machine.png){width=100%}
:::


- *assign* the joint probability distribution above, in such a way that it reflects some reasonable background information

- *encode* the distribution in a computationally useful way

The "encode" undertaking sounds quite challenging, because the number $N$ of units can in principle be infinite; we have an infinite probability distribution.

In the next [Inference III] part we shall see that partially solving the "assign" undertaking actually makes the "encode" one feasible.

\

One question arise if we look at machine-learning methods. Some, and not the most popular, of these methods don't give us probabilities about values: they return *one* definite value. How do we reconcile this with the probabilistic point of view above? We shall answer this question in full in the last chapters, but a short answer can be already given now, as it's very intuitive.

If there are several possible correct answers to a given guess, but a machine-learning algorithm is giving us only one answer, then the algorithm must be internally *choosing* one of them. In other words, it's internally doing decision-making. We know from chapters [@sec-framework] and [@sec-basic-decisions] that this process should obey Decision Theory and therefore *must* involve

- [{{< fa scale-unbalanced-flip >}}\ \ the probabilities of the possible correct answers]{.green}

- [{{< fa sack-dollar >}}\ \ the utilities of the possible choices of answer]{.blue}

Non-probabilistic machine-learning algorithms must therefore be approximations of an Optimal Predictor Machine that, after computing probabilities, selects one unique answer by using utilities.