computational psychology

Fundamental observation #1:

— "cognitive psychology" — a term introduced in 1967 by Ulric Neisser of Cornell as a collective label for the faculties of the mind such as perception, memory, decision-making, etc. — has turned out to mean "computational psychology".

Indeed, not just cognitive, but all of psychology is inherently computational.

computational is special

Fundamental observation #2:

— computational (≈ mathematical) understanding is special [consider Eugene Wigner's comments on math].

"I am not here to tell you the news, ..." [Morris Halle, 1975]

[a digression] climate curriculum: the news and the truth

what is computational/cognitive psychology about?

Computational psychology is about how the mind works.

• Developmental psychology is about how the mind gets to work the way it does.
• Personality psychology is about how my mind works differently from yours.
• Clinical psychology is about how the mind sometimes works very differently.
• Social psychology is about how multiple minds work together; also, about how the mind thinks it works (prescientifically).
• [Artificial Intelligence is about how machines can be made to solve some of the problems that minds can also solve.]
• [Machine Learning is about how machines can learn to solve some of the problems that minds can also solve.]

But what problems do minds solve?

What problems do minds confront, in natural situations?

computational psychology is about...

What problems do minds confront, naturally?

• Consider the problems that are the staple of psychology textbooks, such as:
  • perception
  • memory
  • attention
  Are these problems really the ones that minds confront and solve? [Much more about this in Lecture 2.2.]
• And are the kinds of answers/explanations that psychologists typically offer any good? —
  • under "perception": scene reconstruction
  • under "memory": storage and retrieval
  • under "attention": focusing, steering, etc.

converging on a real explanation of how the mind works

• Consider the explanations typically offered by psychologists and neuroscientists, like the one on the right —
• What would a real explanation look like?
• It would have to be intelligently reductionist, by showing how complexity arises from simpler interacting elements.
• It would have to be literal (as opposed to metaphorical) at its core.

---

The real explanation:
Minds are bundles of dynamical, open-ended COMPUTATIONS over REPRESENTATIONS of the world that brains carry out so as to maximize the probability of their continuing existence, by exercising FORETHOUGHT.

is COMPUTATION just a metaphor?

• the clockwork metaphor: the brain is like...
• the switchboard metaphor: the brain is like...
• the computer metaphor: the brain is like...
• the truth: the brain is like...

COMPUTATION is what brains literally do for a living

• the clockwork metaphor: the brain is like...
• the switchboard metaphor: the brain is like...
• the computer metaphor: the brain is like...
• the truth: the brain is  like  literally a kind of computer.

the truth about the brain: it is a kind of computer

How come the brain is a kind of computer?

• [the trivial reason]
  Because everything is. [Consider again Wigner's comments.]

the truth about the brain: it is a kind of computer

How come the brain is a kind of computer?

• [the trivial reason]
  Because everything is. [Consider again Wigner's comments.]

the truth about the brain: it is a kind of computer

How come the brain is a kind of computer?

• [the trivial reason]
  Because everything is. [Consider again Wigner's comments.]
• [the special reason]
  Because the computing that the brain carries out — e.g., in seeking a conspecific in sensory data — is what pays its bills. Evolution judges brains by the degree to which they succeed in computing what they must compute to survive and procreate.

Compare:
— a piece of chalk computing its trajectory as it falls
— a cash register

what determines whether or not an object is a cash register?

Selection (natural or artificial) pushes information-processing systems (natural or artificial) to succeed in processing information.

---

Cash-registerhood cannot be about what it IS MADE OF.

It must be about what it DOES.

what determines whether or not an object is sentient? (= is of interest to psychology)

It cannot be about what it IS MADE OF.

It must be about what it DOES.

[And no, brain scans are not the answer: all they show is brain cells firing in patterns]

minds NECESSARILY CONSIST of NOTHING BUT computations: example #1 (perception)

the lightness perception problem

minds NECESSARILY CONSIST of NOTHING BUT computations: example #2 (thinking)

the supermarket queuing problem

minds NECESSARILY CONSIST of NOTHING BUT computations: example #3 (action)

the motor control problem

minds NECESSARILY CONSIST of NOTHING BUT computations: example #3 (action)

the motor control problem

minds NECESSARILY CONSIST of NOTHING BUT computations: example #3 (action)

the motor control problem

minds NECESSARILY CONSIST of NOTHING BUT computations: example #3 (action)

the motor control problem

[For a perfectly executed cat jump, see here]

minds, computations, and a bet

A standing bet that I offer:

For ANY aspect or faculty of the mind that you would care to name, I can state and motivate a computational formulation.

For certain aspects and faculties of the mind, some actual understanding is also available.

now: a refresher on computation

• Dynamical systems
• How to get complexity from simplicity
• Turing Machines

The reason I make a point of mentioning Turing Machines at this stage (they will not make another appearance this semester):

On the face of it, the human mind gives off the impression of overwhelming complexity. Learning that such complexity can be built up from very simple elementary building blocks (as a TMs can do it) is therapeutic.

a dynamical system: a piece of chalk + the earth, gravitating to each other

An object in free fall — part of a dynamical system^* consisting of the object and the earth, gravitating to each other — computes its instantaneous velocity and location, given the elapsed time (the key factor being the acceleration due to gravity).

On the left: our experience of the world incorporates knowledge of the dynamics of gravitation and other laws of physics.

---

^*A dynamical system is a system that computes the succession of its states over time as prescribed by a function of the current state (and possibly also past states and any inputs).

a dynamical system: the Game of Life

Conway's game of Life is a (digital/discrete) dynamical system^* —

Observe: it possesses a hierarchical structure.

Hierarchical structure is what makes complexity tractable.

complexity from simplicity: a check point

Most behavioral tasks DO NOT reduce to the execution of a set of fixed rules or the application of a fixed mapping from inputs to outputs.

Compare, for example

• playing chess or Go
with
• getting lunch

---

A complex problem is tractable insofar as it can be approached as a hierarchy of simpler problems.

Even then, a closed-form ahead-of-time solution may not exist and active incremental "online" control — a sequence of simpler steps, each depending on present and past interaction with the environment — may be required.

Complexity emerges out of simplicity.

computing the mind HIERARCHICALLY, level by level

A computation that is reducible to a series of simple (= "stupid") steps (perhaps hierarchically) is called effective.

---

[An effective explanation of the mind would need no miracles and no homunculus.]

effective computation

A procedure P for achieving some desired result is called effective or "mechanical" if:

1. P is set out in terms of a finite number of exact instructions (each instruction being expressed by means of a finite number of symbols);
2. P will, if carried out without error, always produce the desired result in a finite number of steps;

---

Because each step in an effective procedure — even in a very complex one — is specified in simple terms, it can be executed by a machine whose components are simple ("stupid").

an epitome of "complexity out of simplicity": the Turing Machine

A key conceptual tool for understanding how complexity can emerge out of simplicity is the Turing Machine.

The Turing Machine is a very general formalism for describing an effective mapping between input and output symbol streams.

TMs are a general formalism for effectively mapping inputs to outputs

A Turing Machine consists of:

1. a table that specifies exactly how the machine's state changes...
2. ...in response to symbols...
3. ...that it reads (and writes)...
4. ...on a memory tape.

the Turing Machine is a very powerful formalism

Amazingly, various apparent enhancements (such as adding more tapes or more read/write heads) do not increase the power of the Turing Machine as originally defined.

Any general-purpose programmable computer has that same power.

how much should we care about Turing Machines?

The TM is a proof of the principle that very complex computations can be broken down into sequences of very simple ones.

---

To the extent that cognitive computations can be expressed as sequences of very simple elementary steps (such as TM operations), cognition can be explained so that there is no "devil in the details" and no recourse to miracles.

---

Turing Machines and dynamical interactive computation

IMPORTANTLY, real behavior does not reduce to completing a computation that maps a fully given input to an output: DYNAMICAL ONGOING CONTROL is typically needed.

how much should we care about Turing Machines?

[The paper, published in Trends in Cognitive Sciences 15:293-300 (2011), can be found here.]

---

The brain can EMULATE (carry on the operations that correspond to the running of) a Turing Machine, but this is not its NATIVE mode of operation. There is always a difference between native computation and emulation (more about this later in the course).

Turing Machines, brains, dynamics, and representation

A Turing Machine and the brain are both examples of dynamical systems with a very rich capacity (1) for stringing together elementary actions and (2) for building up hierarchically structured complex actions out of simple ones.

However, brains are very much unlike — and with regard to what they compute natively, entirely unlike — Turing Machines.

Functionally, brains particularly excel at ongoing control of flexible behavior.
In this, and in everything else that brains do, they rely on the critically important ability to REPRESENT cognitive PROBLEM AND SOLUTION SPACES, including the dynamics of the world.

---

a dynamical system (e.g., a brain, or a computer)	state 1	→	state 2	→	state 3	→	[...]
a dynamical system (in the world at large)	state A	→	state B	→	state C	→	[...]

---

the basic explanatory concepts coming together

a dynamical system (e.g., a brain, or a computer)	state 1	→	state 2	→	state 3	→	[...]
a dynamical system (e.g., a brain, or a computer)	state I	→	state II	→	state III	→	[...]
a dynamical system (in the world at large)	state A	→	state B	→	state C	→	[...]

---

• computation is everywhere: every dynamical system computes its next state;
• some computations represent each other because of their matching causal organization (dynamics);
• recurring causal patterns of events in the world make forethought possible;
• minds evolve and function by making use of all of the above.

MINDS are bundles of dynamical, open-ended COMPUTATIONS over REPRESENTATIONS of the world that brains carry out so as to maximize the probability of their continuing existence, by exercising FORETHOUGHT.

[EXTRA: concerning forethought and prediction]

what next? (the approach that this course takes)

• food for thought — superficial accounts of aspects of cognition
  • perception
  • memory
  • learning
  • decision making and action
  • sequential behavior
  • language
  • reasoning
  • problem solving
  • emotions

---

• tools for thinking —
  in what terms cognition can be understood, and how it should be studied:
  • computational analysis of the problems that arise in cognition
  • experiments motivated by, or shedding light on, computational theories
  • a glimpse of the power and the limitations of neural computation
  • the overarching computational principles that govern brain function

what next? (the content)

administrative details

The team:

• Shimon Edelman
  Office hours: TR 11:30 - 12:00 (immediately after each class), or by appointment via email.
• Astghik Altunyan
  Office hours: WF 12:00-1:00.

---

• Canvas!
• The Cornell Learning Strategies Center (LSC) maintains a page that can help you find study partners: http://lsc.cornell.edu/studying-together/

technical details: navigating the web site

The home page:

      http://shimon-edelman.github.io/Psych-3140

To navigate the slides for a given lecture, use the keyboard arrows (click the help button in the toolbar below for help).

To get a (relatively) printer-friendly version of a lecture slide set, display the slides in a separate tab of the browser and press "A".

readings: the book and the papers

• The textbook is Computing the Mind: How the Mind Really Works (Oxford University Press, 2008); the library has a couple of copies on reserve.
• I maintain an online list of corrections. If you find an error, please let me know.
• For an informal introduction to the material covered in this course, see my slightly newer (and much lighter) book, The Happiness of Pursuit (Basic Books, 2012).
• As an antidote to the positive psychology leanings and neoliberal politics bullshit of the 2012 book, you may want to read my Life, Death, and Other Inconvenient Truths (MIT Press / Penguin, 2020).
• The syllabus has altogether too many references (I can never resist listing a paper that I think is relevant and clever) — but the number of assigned readings per week is really small. You can do it!

make sure to read the syllabus!

• All the information you need on the readings, assignments, etc. is in the syllabus — please give it a close reading and email me with any questions.
• There are no exams in this course. The weekly reading and writing assignments will determine the final grade (see the syllabus).
  • reading: collaborative annotation
  • writing: the prompt for each week's micro-essay
• Up to 3 bonus points can be earned through participation in experiments advertised on SONA. These get added to the final course score before it is converted into a letter grade.

closing remarks for today

There will be math...   \( \bar{L}\left(\tilde{\textbf{x}}\mid\textbf{y}\right) = \int_{\textbf{x}} L\left(\tilde{\textbf{x}},\textbf{x}\right) p\left(\textbf{x}\mid \textbf{y}\right) d\textbf{x} \)

...but it's nothing you can't manage.