Least Squares Estimation via Projection

Part Of: Machine Learning sequence
Content Summary: 800 words, 8 min read

Projection as Geometric Approximation

If we have a vector b and a line determined by vector a, how do we find the point on the line that is closest to b?

OLS- Geometry of Projection

The closest point p is at the intersection formed by a line through b that is orthogonal to a. If we think of p as an approximation to b, then the length of e = b - p is the error of that approximation.

a^T e = a^T (b - ax) = 0

This formula captures projection onto a vector. But what if you want to project to a higher dimensional surface?

Imagine a plane, whose basis vectors are a_1 and a_2. This plane can be described with a matrix, by mapping the basis vectors onto its column space:

A = \begin{bmatrix} a_1 & a_2 \end{bmatrix}

Suppose we want to project vector b onto this plane. We can use the same orthogonality principle as before:

A^Te = A^T(b-Ax) = 0

A^TAx = A^Tb

Matrices like A^TA are self-transpositions. We have shown that such matrices are square symmetric, and thereby contain positive, real eigenvalues.

We shall assume that the columns of A^TA are independent, and it thereby is invertible. The inverse thereby allows us to solve for x:

(A^TA)^{-1}(A^TA)x = (A^TA)^{-1}A^Tb

x = (A^TA)^{-1}A^Tb

Recall that,

p = Ax = A(A^TA)^{-1}A^Tb

Since matrices are linear transformations (functions that operate on vectors), it is natural to express the problem in terms of a projection matrix P, that accepts a vector b, and outputs the approximating vector p:

p = Pb

By combining these two formula, we solve for P:

P = A(A^TA)^{-1}A^T

Thus, we have two perspectives on the same underlying formula:

OLS- Regression Functions via Matrices (1)

Linear Regression via Projection

We have previously noted that machine learning attempts to approximate the shape of the data. Prediction functions include classification (discrete output) and regression (continuous output).

Consider an example with three data points. Can we predict the price of the next item, given its size?

OLS- Regression Function Setup

For these data, a linear regression function will take the following form:

\psi : Size \rightarrow Price

\psi(Size) = \beta_0 + \beta_1 Size

We can thus interpret linear regression as an attempt to solve Ax=b:

OLS- Linear Algebra Regression Matrix

In this example, we have more data than parameters (3 vs 2). In real-world problems, it is an extremely common predicament. It yields matrices with may more equations than unknowns. This means that Ax=b has no solution (unless all data happen to fall on a straight line).

If exact solutions are impossible, we can still hope for an approximating solution. Perhaps we can find a vector p that best approximates b. More formally, we desire some p = A\bar{x} such that the error e = b-p is minimized.

Since projection is a form of approximation, we can use a projection matrix to construct our linear prediction function \psi : Size \rightarrow Price.

OLS- Least Squares Fundamental Spaces

A Worked Example

The solution is to make the error b-Ax as small as possible. Since Ax can never leave the column space, choose the closest point to b in that subspace. This point is the projection p. Then the error vector e = b-p has minimal length.

To repeat, the best combination p = Ax is the projection of b onto the column space. The error is perpendicular to that subspace. Therefore e = b-p is in the left nullspace:

Ax = b

A^TA = \begin{bmatrix} 1 & 1 & 1 \\ 1 & 2 & 3 \\ \end{bmatrix} \begin{bmatrix} 1 & 1 \\ 1 & 2 \\ 1 & 3 \\ \end{bmatrix} = \begin{bmatrix} 3 & 6 \\ 6 & 14 \\ \end{bmatrix}

We can use Guass-Jordan Elimination to compute the inversion:

(A^TA)^{-1} = \begin{bmatrix} 7/3 & -1 \\ -1 & 1/2 \\ \end{bmatrix}

A useful intermediate quantity is as follows:

(A^TA)^{-1}A^T = \begin{bmatrix} 7/3 & -1 \\ -1 & 1/2 \\ \end{bmatrix} \begin{bmatrix} 1 & 1 & 1 \\ 1 & 2 & 3 \\ \end{bmatrix} = \begin{bmatrix} 4/3 & 1/3 & -2/3 \\ -1/2 & 0 & 1/2 \\ \end{bmatrix}

We are now able to compute the parameters of our model, \bar{x}:

\bar{x} = \left[ (A^TA)^{-1}A^T \right] b = \begin{bmatrix} 4/3 & 1/3 & -2/3 \\ -1/2 & 0 & 1/2 \\ \end{bmatrix} \begin{bmatrix} 1 \\ 2 \\ 2 \\ \end{bmatrix} = \begin{bmatrix} 2/3 \\ 1/2 \\ \end{bmatrix}

These parameters generate a predictive function with the following structure:

\psi : Size \rightarrow Price

\psi(Size) = \frac{2}{3} + \frac{1}{2}Size

These values correspond with the line that best fits our original data!



  • In linear algebra, projection approximates a high-dimensional surface in a lower-dimensional space. The projection error can be measured.
  • In linear regression, we usually cannot solve Ax=b, because there tends to be more data than parameters (b is not in the column space)
  • We can find the closest vector in the column space by projecting onto b, and minimizing the projection error.
  • Thus, the operation of projection can be used to perform parameter estimation, and produce a model that best approximates the training data.

The Argumentative Theory of Reason

Part Of: Demystifying Language sequence
Content Summary: 1200 words, 12 min read.

The Structure of Reason

Learning is the construction of beliefs from experience. Conversely, inference predicts experience given those beliefs.

Reasoning refers to the linguistic production and evaluation of an argument. Learning and inference are ubiquitous across all animal species. But only one species are capable of reasoning: human beings.

Argument can be understood by the lens of deductive logic. Logical syllogisms are a calculus that maps premises to conclusions. An argument is valid if the conclusions follow from the premises. An argument is sound if it is valid, and its premises are true.

Premises can be evaluated directly via intuition. The relationship between argument structure and intuition parallels decision trees versus evaluative functions.

Two Theories of Reason

Why did reasoning evolve? What is its biological purpose? Consider the following theories:

  1. Epistemic theoryreasoning is an extension of our individual cognitive powers. 
  2. Argumentative theory: reasoning is a device for social communication.

One way to adjudicate these rival theories is to examine domain gradients. Roughly, a biological mechanism performs optimally when situated in contexts for which they were originally designed. Our cravings for sugars and fats mislead us today, but encourage optimal foraging in the Pleistocene epoch.

Reasoning is used in both individual and social contexts. But our theories disagree on which is the original domain. Thus, they generate opponent predictions as to which context will elicit the most robust performance.

Argumentative Reason- Domain Gradients (1)

Here we see our first direct confirmation of the argumentative theory: in practice, people are terrible at reasoning in individual contexts. Their reasoning skills become vibrant only when placed in social contexts. It’s a bit like Kevin Malone doing mental math. 🙂

Structure of Argumentative Reason

All languages ever discovered contain both nouns and verbs. This universal distinction reflects the brain’s perception-action dichotomy. Nouns express perceptual concepts, and verbs express action concepts.

Recall that natural language has two processes: speech production & speech comprehension. These functions both accept nouns and verbs as arguments. Thus, we can express the cybernetics of language as follows:
Argumentative Reason- Cybernetics of Language

Argumentative reasoning is a social extension of the faculty of language. It consists of two processes:

  1. Persuasion deals with arguments to support beliefs. 
  2. Justification deals with reasons to justify our actions.

Persuasion and justification draw on perceptual and action concepts, respectively. Thus, the persuasion-justification distinction mirrors the noun-verb distinction, but at a higher level of abstraction. Here is our cybernetics of reasoning diagram.

Argumentative Reason- Cybernetics of Reason

We return to phylogeny. Why did reasoning-as-argumentation evolve?

For communication to persist, it must benefit both senders and receivers. But stability is often threatened by senders who seek to manipulate receivers. We know that humans are gullible by default. Nevertheless, our species does possess lie detection devices. 

The evolution of argumentative reason was shaped by a similar set of ecological pressures as that of language. Let me cover these hypotheses in another post.

For now, it helps to think of belief as clothes, serving both pragmatic and social functions. A wide swathe of biases stems from persuasive arguments performing social rather than epistemic ends.  This is not to say that truth is irrelevant to reasoning. It is simply not always the dominant factor.

On Persuasion

Persuasion processes involve arguments about beliefs. It has two subprocesses: argument production (listener persuasion) and argument evaluation (argument quality inspection). These two processes are locked in an evolutionary arms race, developing ever more sophisticated mechanisms to defeat the other.

Argument production is responsible for the two most damning biases in the human repertoire. There is extensive evidence that we are subject to confirmation bias: the attentional habit to preferentially examine evidence that helps our case. We are also victim to motivated reasoning, which biases our judgments towards our self-interest. We often describe instances of motivated reasoning as hypocrisy.

Consider the following example:

There are two tasks one short & pleasant, the other long & unpleasant. Selectors are asked to select their task, knowing that the other task is giving to another participant (the Receiver). Once they are done with the task, each participant states how fair the Selector has been. It is then possible to compare the fairness ratings of Selectors versus those of the Receivers.

Selectors rate their decisions as more fair than the Receivers, on the average. However, if participants are distracted when they asked their fairness judgments, the ratings were identical and showed no hint of hypocrisy. If reasoning were not the cause of motivated reasoning but the cure for it, the opposite would be expected.

In contrast to production, argument evaluation involves two subprocesses: trust calibration and coherence checking. The ability to distrust malevolent informants has been shown to develop in stages between the ages of 3 and 6.

Coherence checking is less self-serving than production mechanism. In fact, it is responsible for the phenomenon of truth wins. For example, in group puzzles the person whoever stumbles on the solution will successfully persuade her peers, regardless of her social standing. In practice, good arguments tend to be more persuasive than bad arguments. 

On Justification

Justification processes involve reasons about behavior. This is not to be confused with motivations for behavior, which happen at the subconscious level. In fact, there is evidence to suggest that the reasons we acquire by introspection are not true. It has been consistently observed that attitudes based on reasons are much less predictive of future behaviors (and often not predictive at all) than were attitudes stated without recourse to reasons.

The justification module produces reason-based choice; that is, we tend to choose behaviors that are easy to justify to our peers. Reason-based choice explains an impressive number of documented human biases. For example,

The sunk cost fallacy is the tendency to continue an endeavor once an investment has been made. It doesn’t occur in children or non-human animals. If reasoning were not the cause of this phenomenon but the cure for it, the opposite would be expected.

The disjunction effect, endowment effect, and decoy effect can similarly be explained in terms of reason-based choice.

This is not to say that justification is insensitive to the truth. Better decisions are usually easier to justify. But when a more easily justifiable decision is not a good one, reasoning still drives us towards ease of justification.

Theory Evaluation

I was initially skeptical of the argumentative theory because it felt “fashionable” in precisely the wrong sense, underwritten by postmodern connotations of narrative-is-everything and epistemic nihilism. Another warning flag is that the theory draws from the field of social psychology, which has been quite vulnerable to the replication crisis.

However, the evidential weight in favor of the argumentative theory has recently persuaded me. For a comphrehensive view of that evidence, see [MS11]. I no longer believe argumentative reason entails epistemic nihilism, and I predict its evidential basis will not erode substantially in coming decades.

I am also attracted to the theory because it helps tie together several other theories into a comprehensive meta-theory: The Tripartite Mind. Let me sketch just one of example of this appeal.

The heuristics and biases literature has uncovered a bewildering variety of errors, shortcuts, and idiosyncrasies in human cognition.  Responses to this literature vary widely. But too many voices take such biases as “conceptual atoms”, or fundamental facts of the human brain. Neuroscience can and must identify the mechanisms underlying these phenomena.

The argumentative theory is attractive in that it explains a wide swathe of the zoo.

Argumentative Reason- Bias Explanation (1)


Reason is not a profoundly flawed general mechanism. Instead, it is an efficient linguistic device adapted to a certain type of social interaction.


[MS11]. Mercer & Sperber (2011). Why do humans reason? Arguments for an argumentative theory.

Prediction Functions: Regression vs Classification

Part Of: Principles of Machine Learning sequence
Content Summary: 500 words, 5 min read


Data scientists are in the business of answering questions with data. To do this, data is fed into prediction functions, which learn from the data, and use this knowledge to produce inferences.

Today we take an intuitive, non-mathematical look at two genres of prediction machine: regression and classification. Whereas these approaches may seem unrelated, we shall discover a deep symmetry lurking below the surface.

Introducing Regression

Consider a supermarket that has made five purchases of sugar from its supplier in the past. We therefore have access to five data points:

Regression Classification- Regression Data (3)

One of our competitors intends to buy 40kg of sugar. Can we predict the price they will pay?

This question can be interpreted visually as follows:

Regression Classification- Regression Prediction Visualization

But there is another, more systematic way to interpret this request. We can differentiate training data (the five observations where we know the answer) versus test data (where we are given a subset of the relevant information, and asked to generate the rest):

Regression Classification- Regression Prediction Schema

A regression prediction machine will for any hypothetical x-value, predicts the corresponding y-value. Sound familiar? This is just a function. There are in fact many possible regression functions, of varying complexity:

Regression Classification- Simple vs Complex Regression Outputs

Despite their simple appearance, each line represents a complete prediction machine. Each one can, for any order size, generate a corresponding prediction of the price of sugar.

Introducing Classification

To illustrate classification, consider another example.

Suppose we are an animal shelter, responsible for rescuing stray dogs and cats. We have saved two hundred animals; for each, we record their height, weight, and species:
Regression Classification- Classification Data (1)

Suppose we are left a note that reads as follows:

I will be dropping off a stray tomorrow that is 19 lbs and about a foot tall.

A classification question might be: is this animal more likely to be a dog or a cat?

Visually, we can interpret the challenge as follows:

Regression Classification- Classification Prediction Interpretation (2)

As before, we can understand this prediction problem as taking information gained from training data, to generate “missing” factors” from test data:

Regression Classification- Classification Prediction Schema

To actually build a classification machine, we must specify a region-color map, such as the following:

Regression Classification- Simple Classification Output


Indeed, the above solution is complete: we can produce a color (species) label for any new observation, based on whether it lies above or below our line. 

But other solutions exist. Consider, for example, a rather different kind of map:Regression Classification- Complex Classification Boundary

We could use either map to generate predictions. Which one is better? We will explore such questions next time.

Comparing Prediction Methods

Let’s compare our classification and regression models. In what sense are they the same?

Regresssion Classification- Comparing Models

If you’re like me, it is hard to identify similarities. But insight is obtained when you compare the underlying schemas:

Regresssion Classification- Comparing Schema

Here we see that our regression example was 2D, but our classification example was 3D. It would be easier to compare these models if we removed a dimension from the classification example.

Regresssion Classification- 1D Classification (1)

With this simplification, we can directly compare regression and classification:

Regresssion Classification- More Direct Comparison

Thus, the only real difference between regression and classification is whether the prediction (the dependent variable) is continuous or discrete.


Until next time.

Kagan: Structure of Ethical Theories

Part Of: Demystifying Ethics sequence
Followup ToAn Introduction To Ethical Theories
See Also: Shelly Kagan (1994). The Structure of Normative Ethics
Content Summary: 700 words, 7 min read

Are Ethical Theories Incompatible?

Last time, we introduced five major ethical theories:

Ethical Theories- Summary

At first glance, we might consider these theories as rivals competing for the status of a ground for morality. However, when discussing these theories, one has a distinct sense that they are simple addressing different concerns.

Perhaps these theories are compatible with one another. But it is hard to see how, because we lack a map of the major conceptual regions of normative ethics, and how they relate to one another.

Let’s try to construct such a map.

Identifying Morally Relevant Factors

There are two major activities in the philosophical discourse about normative ethics: factorial analysis, and foundational theories.

Factorial analysis involves getting clear on which variables affect in our moral judgments. This is the goal of moral thought experiments. By constructing maps from situations to moral judgment, we seek to understand situational factors that contribute to (and compete for control over) our final moral appraisals.

We can discern four categories of factors which bear on moral judgments: Goodness of Outcome, General Constraints, Special Obligations, and Options. We might call these categories factorial genres. Here are some example factors from each genre.

Ethics Taxonomy- Morally Relevant Factors (1)

While conducting factorial analysis, we typically ask questions about:

  1. Relative Strength. Does Don’t Harm always outweigh factors related to outcome?
  2. Explanatory Parsimony. Is Keep Your Promises redundant with Don’t Be Unfair?
  3. Subfactor Elaboration. What does Maximize Overall Happiness mean, exactly?

Constructing Foundational Theories

A foundational mechanism is a conceptual apparatus designed to generate the right set of morally relevant factors. An example of such a theory is contractarianism, which roughly states that:

Morally relevant factors are those which would be agreed to by a social community, if they were placed in an Original Position (imagine you are designing a social community from scratch), and subject to the Veil of Ignorance (you don’t know the details of what your particular role will be).

Thus, our two philosophic activities relate as follows:

Ethical Structure- Factors vs Foundations (2)

These two activities are fueled by different sets of intuitions.

  • Factorial intuitions are identified by appeal to concrete ethical dilemmas.
  • Foundational intuitions are often related to one’s metaethical dispositions.

Let us examine other accounts of foundational mechanisms. These claim that we should accept only morally relevant factors that…

  • …if everyone followed such rules, total well-being would be maximized (rule utilitarianism).
  •  … if the factor was universalized, became like a law of nature, no contradictions would emerge (Kantian universalization).
  • … can be attributed to a being acting purely in self-interest (egoism).

Localizing Ethical Theories in our Map

We can now use this scheme to better understand the space of ethical theories.

Proposition 1. Ethical theories can be decomposed into their foundational and factorial components.

Three of our five ethical theories have the following decomposition:

Ethical Structure- Deconstructing Ethical Theories (1)

Proposition 2. Factorial pluralism is compatible with foundational monism.

Certain flavors of consequentialists, deontologists, consequentialists insist on factorial monism, that only one kind of moral factor really matters.

But as a descriptive matter, it seems that human morality is sensitive to many different kinds of factors. Outcome valence, action constraint, role-based obligations all seem to play in real moral decisions.

Factorial monism has the unpleasant implication of demonstrating some of these factors as misguided. But philosophers are perfectly free to affirm factorial pluralism: that each intuition “genre” are prescriptively justified.

Some examples of one foundational device generating a plurality of genres:

  1. Rule Utilitarianism (rules that maximize societal well-being) could easily generate rules to keep one’s promises.
  2. Kantian Universalization might generate outcome-sensitive moral factors that are immune to contradiction.
  3. People in the Original Position might enter into a contract of general constraints (e.g., human rights).


Are ethical theories truly competitors? One might suspect that the answer is no. Ethical theories seem to address different concerns.

We can give flesh to this intuition by analyzing the structure of ethical theories. They can be decomposed into two parts: factorial analysis, and foundational mechanisms.

  • Factorial analysis provide the list of factors relevant to moral judgments.
  • Foundational mechanisms are hypothesized to generate these moral factors.

Ethical Structure- Factors vs Foundations (2)

Most defenses of foundational mechanisms have them generating a single factorial genre. However, it is possible to endorse factorial pluralism. There is nothing incoherent in the view that e.g., both event outcome and general constraints bear on morality.

This taxonomy allows us to contrast ethical theories in a new way. Utilitarianism can be seen as a theory about the normative factors, contractarianism is a foundational mechanism. Far from being rival views, one could in fact endorse both!