Options

New Free Booklet on Applied Category Theory

Tai-Danae Bradley has a new free "booklet" on applied category theory. It grew out of the workshop Applied Category Theory 2018, and I think it makes a great complement to Fong and Spivak's books Seven Sketches:

Abstract. This is a collection of introductory, expository notes on applied category theory, inspired by the 2018 Applied Category Theory Workshop, and in these notes we take a leisurely stroll through two themes (functorial semantics and compositionality), two constructions (monoidal categories and decorated cospans) and two examples (chemical reaction networks and natural language processing) within the field.

Check it out!

image

Comments

  • 1.

    Woohoo more fun reading! Thanks for the link!

    Comment Source:Woohoo more fun reading! Thanks for the link!
  • 2.

    Looking forward to reading it! Still need to close many blind spots in this class, but this reading will further help :) Happy to see so much material approachable to layman persons! :)

    Comment Source:Looking forward to reading it! Still need to close many blind spots in this class, but this reading will further help :) Happy to see so much material approachable to layman persons! :)
  • 3.

    In page 24 there is helpful insight on the hints of the end paragraph of previous lecture. Efforts like this aimed at a wider categorical outreach out of closed circles shine. Thanks to the author!

    Comment Source:In page 24 there is helpful insight on the hints of the end paragraph of previous [lecture](https://forum.azimuthproject.org/discussion/2342/lecture-75-chapter-4-the-grand-synthesis#latest). Efforts like this aimed at a wider categorical outreach out of closed circles shine. Thanks to the author!
  • 4.
    edited September 20

    The links provided in the further reading section are also great especially this blog which was mentioned during John's lectures as well. Now with more understanding of the subject the blog seems to be extremely enlightening.

    Also, I came across a paper written by John Baez and James Dolan called From Finite Sets to Feynman Diagrams. I found it to be very illuminating as well. I never thought too deeply about \(\mathbf{FinSet}\) but it is a great place to start for newbies and I regret not doing so earlier.

    Comment Source:The links provided in the further reading section are also great especially [this blog](http://www.math3ma.com) which was mentioned during John's lectures as well. Now with more understanding of the subject the blog seems to be extremely enlightening. Also, I came across a paper written by John Baez and James Dolan called [From Finite Sets to Feynman Diagrams](https://arxiv.org/pdf/math/0004133.pdf). I found it to be very illuminating as well. I never thought too deeply about \\(\mathbf{FinSet}\\) but it is a great place to start for newbies and I regret not doing so earlier.
  • 5.

    Great!! This was very helpful in explaining -- in a mathematically entertaining way -- the pith of some nice topics in applied category theory.

    Comment Source:Great!! This was very helpful in explaining -- in a mathematically entertaining way -- the pith of some nice topics in applied category theory.
  • 6.

    =D> !

    Comment Source:=D> !
  • 7.

    To help me digest these ideas, I thought I'd write a summary of Tai-Danae Bradley's booklet. I've added my own reactions and I've tried link the ideas to what we've discussed in this course.

    Two Themes

    Tai-Danae starts with two basic themes in applied category. (1) Functorial semantics and (2) compositionality.

    Functorial Semantics

    Modeling natural human language is a good example (and Tai-Danae discusses this example at length). Roughly speaking, language consists of words ("semantics") and ways to combine to words to create sentences/new words ("grammar"). This distinction can be captured by a functor between two categories!

    $$ F : C \to D $$ The category \(C\) specifies the grammar (how words act, how they can be combined; see compositionality) while the functor assigns meaning to diagrams in \(C\) (where the category \(D\) represents meaning). So assigning individual words their meaning.

    Another good example in John's Lectures and Seven Sketches in Compositionality is Chapter 3: Database Schemas. A database schema was a category presented by a graph. The schema specified objects (data) and some morphisms (relations; functions).

    The database schema tells us the grammar of our database. It says what data is required, what actions can be performed, how different types of data are related, etc. But it doesn't say what the data actually is (or more abstractly, how we should interpret the objects and morphisms). We get an actual database by constructing a functor which sends our database schema to the category \(\mathbf{Set}\) (assigning sets to objects; and functions to morphisms).

    I think the most important idea here (especially for modeling/ theoretical science) is that we can separate how complex systems combine (this is the grammar) and how those complex systems function (this is the semantics). And we can interpret the same grammar with different semantics (which allows us to be somewhat agnostic about the fundamentals of a complex system).

    Electric circuits are a good example of this last point. We can model a electric circuit using a labeled graph, then construct a category where these labeled graphs are morphisms so we can compose them (this is done with cospans). Each edge in the graph is thought of as some basic component (like a resistor). These constructions don't tell us how voltages or currents distribute over the circuit (i.e., they don't tell us what a resistor really means), just how circuits combine. Choosing how components relate currents and voltages (Ohm's law for linear resistors) is then a functor which sends the labeled graphs to some other category that says how the circuit behaves.

    But the grammar of electric circuits constrains how the circuits can behave, because whatever semantics we choose, the semantics has to preserve the functor axioms (intuitively, it has to distribute over our grammar). So if a resistor behaves like a linear relation between two vector spaces, then sticking resistors together still has to be a linear relation on vector spaces.

    Compositionality

    This idea is relatively straight-forward. Essentially, complex things can be constructed out of simpler parts. But as pointed out in Chapter 1 of Seven Sketches, how you combine parts to form a whole matters! The whole is not necessary the sum of its parts (but it might be up to isomorphism!). A system's components might be behave a bit differently when studied separately than when glued together.

    Biology has many, many examples of this phenomenon! Chemical reactions studied in vitro (in a beaker) aren't the same as in vivo (in a cell). The environment/context matters! The cell has many other reactions that influence the reaction under consideration. Making measurements of system components might not agree with measurements of the system as a whole.

    But this problem affects other fields! Quantum mechanics has this very issue with quantum interference (probabilities depend on the whole system; and can't be added up from measuring decoupled components; this is very roughly speaking as I am not really much of a physicist! ). Even the more hands-on case of electric circuits has this same problem. Adding components like resistors changes what you measure elsewhere in the network.

    A goal in applied category theory is to identify the right algebraic structure for systematically building a system out of some component pieces. Using functorial semantics, algebraic structure in our "syntax" category is preserved. So how systems can be combined partially determines how they behave.

    The Yoneda Lemma (bonus theme)

    Tai-Danae mentions the Yoneda Lemma is her discussion of natural language modeling. The Yoneda Lemma gets the heart of why category theory is useful for modeling! I won't give the formal statement, but I will also quote John Firth:

    You shall know a word by the company it keeps.

    Essentially this is what the Yoneda lemma says: "You shall know an object by the company it keeps." Or "You shall know an object by how it interacts with other objects." In other words, you can define/study an object by how it interacts with other objects. And in category theory, "how it interacts with other objects" means "the morphisms between this object and others".

    To quote from the movie Forrest Gump:

    Stupid is as stupid does

    Similarly, "\(X\) is as \(X\) does" is a wonderful encapsulation of the Yoneda Lemma.

    What does this mean for modeling? Sometimes you don't need to know what a thing is, just what it does!

    Two constructions

    Tai-Danae discusses two constructions: monoidal categories and decorated cospans. I won't really summarize these constructions, but discuss how they could be used in modeling.

    Monoidal Categories

    A monoidal category gives us the syntax/grammar to talk about systems where we can talk about composition in series (via morphisms) and in parallel (via tensoring). We've have many, many examples in this course:

    • Monoidal preorders: e.g. integers with order \(\le\) (as morphisms) under addition (as tensoring)
    • Resource monoidal preorder: stuff (like pie ingredients) and mixtures of stuff (tensoring) with processes/reactions turning stuff/mixtures into other stuff/mixtures (morphisms)
    • Database schema: data types with combined data types (tensoring) and relationships/actions between types (morphisms)
    • Co-design diagrams

    along more traditional math examples (Tai-Danae gives some good examples).

    Another way to think about the monoid part of a monoidal category is to notice that in several examples the monoidal product lets us combine or mix together basic stuff. For example, chemical reactions turn mixtures of certain substances to other substances. We can describe reactions as morphisms (morphisms are good for processes) while mixing stuff together is tensoring.

    Decorated Cospans

    Let's imagine a pipe (or a wire). Pipes have no natural distinction between "input" and "output". Water can go in either direction. But morphisms are always directional! If we want to turn "direction agnostic" processes into a morphisms, we have to find a way to make distinction between input and output not really matter. Decorated cospans are the answer (well an answer).

    A cospan is the following diagram: $$ X \rightarrow N \leftarrow Y $$ where the object \(N\) represents our pipe (potentially bidirectional process). What these objects are depends on the category (finite sets is a common choice since we will have to be able to compute colimits for any objects to get a category; and finite sets always have colimits).

    The cospan can be decorated with some extra structure (which might say things about how our pipe can be combined with other pipes). Decorating our cospan is the same idea as the functorial semantics.

    Comment Source:To help me digest these ideas, I thought I'd write a summary of Tai-Danae Bradley's booklet. I've added my own reactions and I've tried link the ideas to what we've discussed in this course. ## Two Themes Tai-Danae starts with two basic themes in applied category. (1) Functorial semantics and (2) compositionality. ### Functorial Semantics Modeling natural human language is a good example (and Tai-Danae discusses this example at length). Roughly speaking, language consists of words ("semantics") and ways to combine to words to create sentences/new words ("grammar"). This distinction can be captured by a functor between two categories! \[ F : C \to D \] The category \\(C\\) specifies the grammar (how words act, how they can be combined; see compositionality) while the functor assigns meaning to diagrams in \\(C\\) (where the category \\(D\\) represents meaning). So assigning individual words their meaning. Another good example in John's Lectures and *Seven Sketches in Compositionality* is Chapter 3: Database Schemas. A database schema was a category presented by a graph. The schema specified objects (data) and some morphisms (relations; functions). The database schema tells us the grammar of our database. It says what data is required, what actions can be performed, how different types of data are related, etc. But it doesn't say what the data actually is (or more abstractly, how we should interpret the objects and morphisms). We get an actual database by constructing a functor which sends our database schema to the category \\(\mathbf{Set}\\) (assigning sets to objects; and functions to morphisms). I think the most important idea here (especially for modeling/ theoretical science) is that we can separate how complex systems combine (this is the grammar) and how those complex systems function (this is the semantics). And we can interpret the same grammar with different semantics (which allows us to be somewhat agnostic about the fundamentals of a complex system). Electric circuits are a good example of this last point. We can model a electric circuit using a labeled graph, then construct a category where these labeled graphs are morphisms so we can compose them (this is done with cospans). Each edge in the graph is thought of as some basic component (like a resistor). These constructions don't tell us how voltages or currents distribute over the circuit (i.e., they don't tell us what a resistor really means), just how circuits combine. Choosing how components relate currents and voltages (Ohm's law for linear resistors) is then a functor which sends the labeled graphs to some other category that says how the circuit behaves. But the grammar of electric circuits constrains how the circuits can behave, because whatever semantics we choose, the semantics has to preserve the functor axioms (intuitively, it has to distribute over our grammar). So if a resistor behaves like a linear relation between two vector spaces, then sticking resistors together still has to be a linear relation on vector spaces. ### Compositionality This idea is relatively straight-forward. Essentially, complex things can be constructed out of simpler parts. But as pointed out in Chapter 1 of *Seven Sketches*, how you combine parts to form a whole matters! The whole is not necessary the sum of its parts (but it might be up to isomorphism!). A system's components might be behave a bit differently when studied separately than when glued together. Biology has many, many examples of this phenomenon! Chemical reactions studied *in vitro* (in a beaker) aren't the same as *in vivo* (in a cell). The environment/context matters! The cell has many other reactions that influence the reaction under consideration. Making measurements of system components might not agree with measurements of the system as a whole. But this problem affects other fields! Quantum mechanics has this very issue with quantum interference (probabilities depend on the whole system; and can't be added up from measuring decoupled components; this is very roughly speaking as I am not really much of a physicist! ). Even the more hands-on case of electric circuits has this same problem. Adding components like resistors changes what you measure elsewhere in the network. A goal in applied category theory is to identify the right algebraic structure for systematically building a system out of some component pieces. Using functorial semantics, algebraic structure in our "syntax" category is preserved. So how systems can be combined partially determines how they behave. ### The Yoneda Lemma (bonus theme) Tai-Danae mentions the Yoneda Lemma is her discussion of natural language modeling. The Yoneda Lemma gets the heart of why category theory is useful for modeling! I won't give the formal statement, but I will also quote John Firth: > You shall know a word by the company it keeps. Essentially this is what the Yoneda lemma says: "You shall know an object by the company it keeps." Or "You shall know an object by how it interacts with other objects." In other words, you can define/study an object by how it interacts with other objects. And in category theory, "how it interacts with other objects" means "the morphisms between this object and others". To quote from the movie Forrest Gump: > Stupid is as stupid does Similarly, "\\(X\\) is as \\(X\\) does" is a wonderful encapsulation of the Yoneda Lemma. What does this mean for modeling? Sometimes you don't need to know what a thing is, just what it does! ## Two constructions Tai-Danae discusses two constructions: monoidal categories and decorated cospans. I won't really summarize these constructions, but discuss how they could be used in modeling. ### Monoidal Categories A monoidal category gives us the syntax/grammar to talk about systems where we can talk about composition in series (via morphisms) and in parallel (via tensoring). We've have many, many examples in this course: + Monoidal preorders: e.g. integers with order \\(\le\\) (as morphisms) under addition (as tensoring) + Resource monoidal preorder: stuff (like pie ingredients) and mixtures of stuff (tensoring) with processes/reactions turning stuff/mixtures into other stuff/mixtures (morphisms) + Database schema: data types with combined data types (tensoring) and relationships/actions between types (morphisms) + Co-design diagrams along more traditional math examples (Tai-Danae gives some good examples). Another way to think about the monoid part of a monoidal category is to notice that in several examples the monoidal product lets us combine or mix together basic stuff. For example, chemical reactions turn mixtures of certain substances to other substances. We can describe reactions as morphisms (morphisms are good for processes) while mixing stuff together is tensoring. ### Decorated Cospans Let's imagine a pipe (or a wire). Pipes have no natural distinction between "input" and "output". Water can go in either direction. But morphisms are always directional! If we want to turn "direction agnostic" processes into a morphisms, we have to find a way to make distinction between input and output not really matter. Decorated cospans are the answer (well an answer). A cospan is the following diagram: \[ X \rightarrow N \leftarrow Y \] where the object \\(N\\) represents our pipe (potentially bidirectional process). What these objects are depends on the category (finite sets is a common choice since we will have to be able to compute colimits for any objects to get a category; and finite sets always have colimits). The cospan can be decorated with some extra structure (which might say things about how our pipe can be combined with other pipes). Decorating our cospan is the same idea as the functorial semantics.
Sign In or Register to comment.