Data-parallel probabilistic inference

Abstract

The present invention relates to a probabilistic programming compiler that (a) generates data-parallel inference code to sample from probability distributions in models provided to the compiler; and (b) utilizes a modular framework to allow addition and removal of inference algorithm information based on which the compiler generates the inference code. For a given model, the described compiler can generate inference code that implements any one or more of the inference algorithms that are available to the compiler. The modular compiler framework utilizes an intermediate representation (IR) that symbolically represents features of probability distributions. The compiler then uses the IR as a basis for emitting inference code to sample from the one or more probability distributions represented in the IR. Further, the compiler produces parallelized inference code that facilitates efficient parallel processing of inference computations in order to take advantage of highly data-parallel architectures, such as GPUs.

Description

FIELD OF THE INVENTION

The present invention relates to a probabilistic programming compiler that (a) generates data-parallel inference code to sample from probability distributions in models provided to the compiler; and (b) utilizes a modular framework to allow addition and removal of inference algorithm information, based on which the compiler generates the inference code.

BACKGROUND

Machine learning facilitates complex, statistics-based computations such as speech recognition, image recognition, retrieving information from unorganized text, etc. However, machine learning, and especially probabilistic modeling, can be difficult to apply. To apply machine learning, a user needs to not only design a statistical model, but to also design and implement the right inference procedure to perform inference on the model. There are many different inference algorithms, most of which are conceptually complicated and difficult to implement at scale. Despite the enthusiasm that many people who practice data analysis have for machine learning, this complexity is a barrier to deployment.

Probabilistic programming is a way to simplify the application of machine learning based on Bayesian inference. The key feature of probabilistic programming is separation of concerns: a user specifies what needs to be learned by describing a probabilistic model, while a compiler automatically generates the inference procedure for the described model. In particular, the programmer writes source code that describes one or more probability distributions, and the source code is parsed by a probabilistic programming compiler. The compiler then generates inference code that implements an inference algorithm configured to sample from the one or more described distributions. Using the compiler-generated inference code, the programmer samples from the described distribution(s).

Doing inference on probabilistic programs is computationally intensive and challenging. Most of the algorithms developed to perform inference are conceptually complicated, and it is not clear, especially for non-experts, which algorithm would work best for a given model. It would be beneficial to facilitate selection of an inference algorithm in the context of probabilistic programming.

Furthermore, the amount of data being analyzed via inference code is often voluminous, and a large amount of computation power is required to process such data. In order to produce efficient inference code, it would be beneficial to develop techniques that process large amounts of data efficiently.

The approaches described in this section are approaches that could be pursued, but not necessarily approaches that have been previously conceived or pursued. Therefore, unless otherwise indicated, it should not be assumed that any of the approaches described in this section qualify as prior art merely by virtue of their inclusion in this section.

BRIEF DESCRIPTION OF THE DRAWINGS

In the drawings:

FIG. 1 is a block diagram that depicts an example network arrangement for a probabilistic programming system.

FIG. 2 depicts an example Latent Dirichlet Allocation (LDA) model specification.

FIG. 3 depicts an example regression model specification.

FIG. 4 depicts a procedure for initiating sampling of an LDA model.

FIG. 5 depicts a flowchart for creating an intermediate representation from a provided model specification and producing inference code, implementing a particular inference algorithm, to sample from one or more probability distributions specified in the provided model.

FIGS. 6 and 7 depict mathematical expressions.

FIG. 8 depicts a procedure to sample from K Dirichlet variates.

FIG. 9 is a block diagram of a computer system on which embodiments may be implemented.

DETAILED DESCRIPTION

In the following description, for the purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present invention. It will be apparent, however, that the present invention may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form in order to avoid unnecessarily obscuring the present invention.

General Overview

For probabilistic programming to perform well, an appropriate inference algorithm should be selected for a given statistical model. A statistical model is a description of one or more probability distributions of a particular body of data. Some data models are best sampled using one or more particular inference algorithms. Thus, if a particular compiler only emits inference code for a particular inference algorithm (such as Metropolis-Hastings, Gibbs, or Slice, etc.), then there are models for which the compiler will be of little or no use, even given large amounts of computational power.

Described herein is a probabilistic programming compiler with a modular framework, which facilitates addition and removal of inference algorithm information that the compiler may use to create inference code. Thus, for a given model, the described compiler can generate inference code that implements any one or more of the inference algorithms that are available to the compiler. Such flexibility facilitates experimentation to determine which inference algorithm(s) work best for a given model by allowing a user to create samplers based on any of the inference algorithms available to the compiler.

The modular compiler framework of the described compiler utilizes an intermediate representation (IR) that symbolically represents features of probability distributions. Specifically, when the compiler receives a model specification, the compiler creates an IR that symbolically represents features of one or more probability distributions specified by the received model specification. The compiler then uses the created IR as a basis for emitting inference code to sample from the one or more probability distributions.

Any class of probability distribution may be parsed by the compiler as long as the compiler has information indicating how to create an IR that represents features of the class of probability distribution. Further, the compiler can emit inference code that implements any inference algorithm as long as the compiler has information about how to create inference code, based on probability distribution features represented in an IR, that implements the particular algorithm. Thus, the intermediate representation described herein facilitates scaling application of the compiler to various combinations of inference algorithms being used to sample from various classes of probability distributions.

Furthermore, probabilistic programming performance is enhanced by utilizing advances in parallel computing hardware to perform computations in parallelized inference code. For example, many MCMC inference algorithms are highly data-parallel within a single Markov Chain if advantage is taken of the conditional independence relationships of the input model (e.g., the assumption of independent and identically distributed (i.i.d.) data makes the likelihood independent across data points).

Thus, described herein is a probabilistic programming language and compiler for Bayesian networks designed to make effective use of data-parallel architectures such as Graphics Processing Units (GPUs). A GPU is a processor with a highly parallel structure, having a plurality of execution units called “cores” (e.g., tens, hundreds, or thousands of cores, etc.). Each core of a GPU is configured to process a thread of computation in parallel with the functioning of the other cores of the GPU.

The described compiler produces parallelized inference code configured to be run by a GPU (or other data-parallel architecture such as a distributed system). An example of a type of code configured to be run on a GPU is Compute Unified Device Architecture (CUDA) code, which utilizes a GPU for general processing. Parallelized inference code facilitates parallel processing of inference computations in order to take advantage of the GPU's parallel structure and therefore speed up processing the computations. In embodiments, the compiler can generate data-parallel inference code scalable to thousands of GPU cores by making use of the conditional independence relationships in the Bayesian network.

Probabilistic Programming Architecture

FIG. 1 is a block diagram that depicts an example network arrangement 100 for a probabilistic programming system, according to embodiments. Network arrangement 100 includes a client device 110 and a server device 120 communicatively coupled via a network 130. Example network arrangement 100 may include other devices, including client devices, server devices, and display devices, according to embodiments.

Client device 110 may be implemented by any type of computing device that is communicatively connected to network 130. Example implementations of client device 110 include, without limitation, workstations, personal computers, laptop computers, personal digital assistants (PDAs), tablet computers, cellular telephony devices such as smart phones, and any other type of computing device.

In network arrangement 100, client device 110 is configured with a programming client 112. Programming client 112 may be implemented in any number of ways, including as a stand-alone application running on client device 110, as a plugin to a browser running at client device 110, etc. Programming client 112 may be implemented by one or more logical modules. Client device 110 may be configured with other mechanisms, processes and functionalities, depending upon a particular implementation.

Network 130 may be implemented with any type of medium and/or mechanism that facilitates the exchange of information between client device 110 and server device 120. Furthermore, network 130 may facilitate use of any type of communications protocol, and may be secured or unsecured, depending upon the requirements of a particular embodiment.

Server device 120 may be implemented by any type of computing device that is capable of communicating with client device 110 over network 130. In network arrangement 100, server device 120 is configured with a compiler 122. Compiler 122 has access to two or more inference modules 124A-N, i.e., via a modular compiler framework described in further detail below. Any of the functionality attributed to compiler 122 herein may be performed by another logical entity running on server device 120, or on client device 110, or on other devices that are communicatively coupled to network 130, according to embodiments. Compiler 122 may be implemented by one or more logical modules. Server device 120 may be configured with other mechanisms, processes and functionalities, depending upon a particular implementation.

Programming client 112 and/or compiler 122 may receive and respond to Application Programming Interface (API) calls, Simple Object Access Protocol (SOAP) messages, requests via HyperText Transfer Protocol (HTTP), HyperText Transfer Protocol Secure (HTTPS), Simple Mail Transfer Protocol (SMTP), or any other kind of communication, e.g., from other entities communicatively coupled to network 130. Further, programming client 112 and/or compiler 122 may send one or more communications, using any of the above protocols or any other communication protocol, over network 130 to one of the other entities communicatively coupled to network 130, according to embodiments.

In an embodiment, each of the processes described in connection with programming client 112 and/or compiler 122 are performed automatically and may be implemented using one or more computer programs, other software elements, and/or digital logic in any of a general-purpose computer or a special-purpose computer, while performing data retrieval, transformation, and storage operations that involve interacting with and transforming the physical state of memory of the computer.

Probabilistic Programming—Model Specification

To initiate probabilistic programming to do inference on a particular data model with one or more probability distributions, a user provides, to compiler 122, source code that represents a specification of the data model. Such source code may be developed and provided to compiler 122 via programming client 112 of client device 110. Source code is computer code written in a human-readable programming language such as Scala, C++, Java, etc.

Compiler 122 parses the source code, which transforms the source code to a target computer language, such as a computer-readable language that instructs a GPU or a CPU of a computing device to perform tasks indicated by the source code. As such, compiler 122 parses the source code that describes the data model, and causes server device 120 to create, from the parsed source code, a representation (i.e., an IR) of the data model as described in further detail herein.

FIG. 2 depicts an example Latent Dirichlet Allocation (LDA) model specification 200, which is source code that describes one or more probability distributions of the LDA model. According to embodiments, compiler 122 is a Scala compiler, and the probability distributions in specification 200 are defined for compiler 122 using Scala objects. However, within embodiments, data model specifications may be in any source language (e.g., Scala, Java, C++, etc.) that is capable of conveying information needed to define a data model.

According to the embodiment of FIG. 2, the probability distribution described in specification 200 is defined as a Scala object (object LDA) and is composed of two declarations. First, the support of the probability distribution is declared as a class 202 named “sig”. This support is composed of four arrays, with one array each for the distribution of topics per document (theta), the distribution of words per topic (phi), the topics assigned to the words (z), and the words in the corpus (w). This support is used to store the inferred model parameters. These last two arrays are flat representations of ragged arrays, and so do not require the documents to be of equal length.

The second declaration 204 of specification 200 specifies the Bayesian network associated with LDA and makes use of a domain-specific language (DSL) for Bayesian networks. The DSL is marked by the “bayes” keyword and delimited by the following enclosing brackets. The model first declares the parameters of the model: “K” for the number of topics, “V” for the vocabulary size, “M” for the number of documents, and “N” for the array that associates each document with its size.

In the model itself, the hyper-parameters (values “alpha” and “beta”) are defined for the Dirichlet distributions and draw K Dirichlet samples of dimension V for the distribution of words per topic (phi) and M Dirichlet samples of dimension K for the distribution of topics per document (theta). Then, for each word in each document, a topic z is drawn from theta, and finally a word from phi based on the topic drawn for z.

FIG. 3 depicts a regression model specification 300, which is defined in the same way as the model depicted in FIG. 2, and uses similar language features. In the example of model specification 300, the support comprises the “(x, y)” data points, the weights “w”, the bias “b”, and the noise “tau”. The model uses an additional sum function to sum across the feature vector.

Using a Specified Model

According to embodiments, once a model specification is parsed, it can be used as any other Scala object by writing standard Scala code. For instance, one may use the LDA model specified in FIG. 2 with a training corpus to learn a distribution of words per topic and then use it to learn the per-document topic distribution of a test corpus.

Scala code implementation 400 of FIG. 4 represents such an implementation procedure, according to an embodiment, using the LDA model specified in FIG. 2. The function LDA.model.map returns a maximum a posteriori estimation. The function takes, as arguments, the set of variables to observe (on top of the ones declared as observed in the model specification), the hyperparameters, the initial parameters, the output parameters, the number of iterations, and the inference to use. The parameters are stored in LDA.sig.

The implementation of FIG. 4 allocates the parameter arrays, which contain the inferred values. Then the signature of the model is constructed which encapsulates the parameters. The LDA.model.map command returns the MAP estimate of the parameters given the observed words.

To test the model, a new signature is constructed containing the test documents, and the previously inferred phi values. Then LDA.model.map is called again, but with both the phis and the words observed (by supplying Set(“phi”)). The inferred thetas for the test documents are stored in s_test.theta.

A Modular Compilation Framework

According to embodiments, there are two distinct compilation phases when a model goes from a specification down to inference code. The first compilation phase happens when the programmer compiles a model specification such as specification 200 of FIG. 2. According to an embodiment, compiler 122 compiles the model specification with scalac as follows (where the source code for specification 200 is stored in a file named LDA.scala):

scalac-classpath augur.jar LDA.scala

The file augur.jar is the package containing compiler 122. The first phase of compilation happens statically, during normal scalac compilation. In this phase, the block of code following the bayes keyword in specification 200 is transformed into an intermediate representation (IR) for probability distributions. This IR represents features of the probability distribution specified in specification 200.

The second compilation phase happens at runtime, when the programmer calls the LDA.model.map method. At that point, the IR is transformed, analyzed, and optimized, and finally, inference code is emitted and run.

The framework of compiler 122 is therefore composed of two distinct components that communicate through the IR: a front end of compiler 122, where model specifications in domain-specific languages are converted into IR, and a back end of compiler 122, where the IR is compiled down to inference code that implements various inference algorithms (such as Metropolis-Hastings, Gibbs sampling, Metropolis-Within-Gibbs, etc.).

According to an embodiment, the Scala macro system is used to define a modeling language in the front end. This macro system facilitates defining a set of functions (called “macros”) that will be executed by compiler 122 on the code enclosed by the macro. Example DSLs, for representing classes of probability distributions, used by compiler 122 include Bayesian networks, Markov random fields, etc. Addition of DSLs to the front end of compiler 122 comprises informing the front end of compiler 122 how to convert the DSL to an IR. As such, addition of DSLs to the front end of compiler 122 would not require modification of the back end of compiler 122, which emits inference code based on models represented with IR.

Because the IR is compiled to inference code at runtime, the values of all the hyper-parameters and the size of the dataset are known to compiler 122. This enables compiler 122 to implement better optimization strategies, and also provides information about how to extract parallelism, as described in further detail below. For example, when compiling LDA, the number of topics is known to be much smaller than the number of documents, and thus parallelizing over documents will produce more parallelism than parallelizing over topics.

The back-end of compiler 122 has the capacity to implement those inference algorithms for which the compiler has information, i.e., stored at inference modules 124A-N. Each inference module of inference modules 124A-N has information for creating inference code, to implement a particular inference algorithm, based on IR of a model. In order to expand the capacity of compiler 122 to emit inference code that implements an inference algorithm that was previously unknown to compiler 122, a user creates a new inference module (e.g., inference module 124B) with information indicating how to produce inference code, for a model represented by IR, that implements the new inference algorithm.

Compiler 122 can use one or more of the modular inference algorithms available to compiler 122 to create one or more samplers for statistical models represented with IR. Because compiler 122 creates an intermediate representation of any model that is compiled, no matter how the model is specified (i.e., source language used or class of distribution specified in the model), any of the inference algorithms represented in inference modules 124A-N may be used to create inference code for the specified model, since inference modules 124A-N include information for how to create inference code from the IR of a model. Analysis of an IR that represents a particular model can recover all of the information needed for the inference code generation for that model.

Furthermore, according to an embodiment, compiler 122 has access to a library that defines standard distributions, such as Gaussian, Dirichlet, etc. In addition to these standard distributions, each model specification denotes its own user-defined distribution. According to embodiments, all of these distributions are subtypes of the Dist supertype. Currently, the Dist interface provides two methods: map, which implements maximum a posteriori estimation, and sample, which returns a sequence of samples.

Symbolic Representations of Statistical Models

FIG. 5 depicts a flowchart 500 for creating an IR from a provided model specification and producing inference code, implementing a particular inference algorithm, to sample from one or more probability distributions specified in the provided model. At step 502, source code, representing a model of a body of data, is compiled into an intermediate representation of the model, wherein the model of the body of data specifies one or more probability distributions, and wherein the intermediate representation of the model represents at least one probability distribution, of the one or more probability distributions, with one or more mathematical expressions.

To create a symbolic intermediate representation of a model, compiler 122 parses a model specification represented in source code, and identifies characteristics of the model from the specification. For example, the specification of a particular model instructs compiler 122 to draw a certain number of times from a Gaussian distribution. Upon identifying this instruction, compiler 122 emits a mathematical formula, as part of the symbolic IR of the model, that represents the instruction. To illustrate, compiler 122 emits a mathematical formula that represents a product for every time that the specification indicates to draw from the Gaussian distribution.

As a further example, compiler 122 compiles model specification 300 into an IR of the model, i.e., IR expression 610 of FIG. 6. According to an embodiment, compiler 122 generates IR expressions from the Backus-Naur form grammar 600 of FIG. 6. To create the IR for specification 300, compiler 122 emits IR for a probability density function ƒ that is proportional to the distribution that the user wants to sample from (i.e., the distribution represented in specification 300). The function ƒ is composed of the product of the prior and the likelihood of the model and is extracted automatically by compiler 122 from the provided model specification 300. Thus, applied to the regression example in specification 300, ƒ is defined as shown in IR expression 610 of FIG. 6.

IR expression 610 is a mathematical representation of the probability distribution specified in the regression model represented in specification 300. As such, IR expression 610 symbolically represents features of the specified probability distribution, such as properties of variables for the distribution, and interdependencies between the distribution variables, etc.

The use of such a symbolic (or mathematical) intermediate representation for the model facilitates scaling to large networks. Existing probabilistic programming language implementations, such as JAGS or Stan, utilize a graphical representation of models being processed, rather than symbolic representations of such models. A graphical representation of a model materializes, in memory, values for all (or substantially all) of the data points associated with the model.

A probabilistic programming compiler that utilizes a graphical representation of a model relies on the stored data points to produce inference code that implements a particular inference algorithm to sample from the graphically-represented model. Because all of the data points for the model are materialized in memory, it is relatively easy to pull all of the information, from the materialized graph, that is needed to create a sampler for the represented model. However, graphical representations of models grow linearly as the size of the modeled data grows. Thus, storing and performing analysis on graphical representations of models become unworkable as modeled data grows, resulting in unreasonable memory consumption for graphical representations of models such as LDA.

In contrast with a graphical representation of a model, a symbolic representation of a model is expressed with one or more mathematical expressions. The one or more mathematical expressions summarize classes of data points for the represented model, as well as relationships between the data points in the model. Such a symbolic representation describes one or more probability distributions for a model without instantiating all, or even a substantial portion, of the values of data points of the probability distributions in memory, or in other words, without creating objects that represent the data points of the distribution in memory. The mathematical expressions do not require materialization of a models' data points in memory in order to represent features of the model.

An example that illustrates the difference between symbolic and graphical representations of distributions: compiler 122 parses a specification for a statistical model that models an array of words, where each word in the array is associated with a distinct random variable. A graphical representation of this statistical model would instantiate, in memory, an object that represents each random variable for each word in the array. Thus, if there are a million words in the array, a graphical representation of this model would include a million random variable objects in memory corresponding to the million words, respectively. Further, depending on how the random variables interrelate, a graphical representation of the model would also include instantiated edges between the random variable objects that describe each relationship between the variables.

In contrast, a symbolic representation of this model represents the random variables with a mathematical formula that indicates the product of all of the random variables for the words in the array. The symbolic representation of the model does not instantiate an object for each variable in memory, but includes information in the mathematical formula that symbolically indicates that the model includes a random variable for each word in the array. The symbolic representation describes relationships between the variables using the mathematical formulas rather than with edges that are instantiated in memory.

Generating Metropolis-Hastings Samplers

Returning to flowchart 500, at step 504, a particular inference algorithm, of a plurality of inference algorithms, is identified. To illustrate, in response to an inference procedure (e.g. LDA.model.map) being invoked on a model, compiler 122 initiates the second compilation phase in which compiler 122 emits inference code (such as CUDA code) for the model based on the IR of the model. According to an embodiment, a command invoking an inference procedure on a model includes information about which inference algorithm to use for the inference, such as the reference to “Infer.GIBBS” in the LDA.model.map command of implementation 400 (FIG. 4).

According to an example, a user instructs compiler 122 to use Metropolis-Hastings (MH) inference to emit inference code to sample from a distribution in the model of specification 300, which instruction includes a reference to the MH inference algorithm (such as “Infer.MH”). In this example, compiler 122 has access to information, at inference module 124A, about how to create MH inference code for a provided model based on an IR that represents the model.

At step 506 of flowchart 500, inference code that implements the particular inference algorithm is produced, wherein the inference code is configured to sample from the at least one probability distribution of the one or more probability distributions, and wherein producing the inference code is based, at least in part, on an analysis of the one or more mathematical expressions.

For example, in response to an instruction to use MH inference to sample from a probability distribution represented in IR expression 610, compiler 122 accesses the inference module 124A that corresponds to the requested inference algorithm (MH). Inference module 124A includes information about how to create inference code that implements MH based on the IR of a model, i.e., IR expression 610. Using the symbolic representation (IR 610), compiler 122 identifies characteristics of the data points of the model, which are needed to create inference code for the model.

Compiler 122 uses the information from inference module 124A to create MH inference code based on IR expression 610. To produce MH inference code based on IR expression 610, compiler 122 analyzes IR expression 610 based on information in the library implementation of Metropolis-Hastings, i.e., inference module 124A.

According to an embodiment, inference module 124A includes one or more IR rewrite rules. In the example of MH inference, the rewrite rules are generally-applicable mathematical rules for transforming the IR. As at least part of the analysis of the IR, compiler 122 rewrites IR expression 610 based on the algorithm-specific IR rewrite rules, to produce transformed expression 620 of FIG. 6 (which is equal to expression 610). Compiler 122 produces inference code based on analysis of IR expression 610. According to embodiments, such analysis includes analysis of transformed expression 620 using one or more IR analysis rules.

Identifying Data-Parallelism for the MH Inference Code

It can be challenging to produce inference code that effectively uses the parallelism in highly parallel architectures. Examples herein are given in the context of a GPU and CUDA code, but embodiments are not limited to GPU parallel architecture for implementation of parallelized inference code. Traditional methods of performing inference on a model using a CPU, which generally has one to four cores and is latency oriented, do not always work well with a GPU, which has many more cores than a CPU and is throughput oriented. Programs designed for execution on a CPU do not utilize much parallelism, and would not necessarily run more quickly if implemented on an architecture with more parallelism.

Thus, compiler 122 identifies parallelism in IR, which can be exploited to produce efficient inference code that utilizes the capabilities of a highly parallel architecture. The form of the IR makes the sources of parallelism in the model being processed more explicit. For example, compiler 122 analyzes the probability distributions represented in an IR by applying a set of parallelism rules to the IR to identify portions of the represented probability distribution that may be evaluated in parallel. An example parallelism rule states: a Π indicates that each sub-term of the product symbol can be evaluated in parallel.

Based on such analysis of an IR, such as transformed expression 620, compiler 122 identifies that the distribution factorizes into a large number of terms that are independent and can be evaluated in parallel and then efficiently multiplied together. More specifically, compiler 122 determines that the data is i.i.d. and that it can optimize the inference code accordingly. In this case, each (x, y) of expression 620 contributes to the likelihood independently, and they can be evaluated in parallel. According to an embodiment, compiler 122 works in log-space, and thus performs summations. As such, compiler 122 generates parallelized inference code for the evaluation of ƒ from the IR expression 610, which inference code takes advantage of the parallelism of highly parallel data structures.

Despite the simplicity of this MH parallelization, the code scales reasonably well: there is a large amount of parallelism because it is roughly proportional to the number of data points; uncovering the parallelism in the code does not increase the overall quantity of computation that has to be performed; and the ratio of computation to global memory accesses is high enough to hide memory latency bottlenecks.

Generating Data-Parallel Gibbs Samplers

As a further example, a user instructs compiler 122 to use Gibbs inference to emit inference code to sample from a distribution in a provided model (such as the LDA model of specification 200). In this example, compiler 122 has access to information, at inference module 124B, about how to create Gibbs inference code for a provided model based on an IR that represents the model.

Compiler 122 compiles source code (such as specification 200) representing an LDA model, of a body of data, into an intermediate representation of the model, such as intermediate representation expression 700 of FIG. 7. Compiler 122 receives an instruction to create Gibbs inference code for the provided model, which causes compiler 122 to emit the requested Gibbs inference code (i.e., a Gibbs sampler).

To generate a Gibbs sampler from IR expression 700, compiler 122 analyzes how to sample from each univariate distribution represented in expression 700. As an example, to draw θ_m as part of the (τ+1)th sample, compiler 122 generates code that samples from the distribution indicated in expression 700.

In order to uncover conjugacy relations and independence between variables based on the symbolic representation of the model (i.e., IR expression 700), compiler 122 implements an algebraic rewrite system that attempts to rewrite expression 700 in terms of expressions known to compiler 122 (i.e., the joint distribution of the entire model). This rewriting of the IR is at least part of the analysis that compiler 122 performs on the IR of a model in order to produce inference code to sample from one or more distributions specified in the model. Expressions 710 of FIG. 7 show a few selected rewrite rules (a)-(d) to give a flavor of a rewrite system according to an embodiment.

Rule (a) of expressions 710 states that like terms can be canceled. Rule (b) says that terms that do not depend on the variable of integration can be pulled out of the integral. Rule (c) says that it is possible to partition a product over N-terms into two products, one where a predicate q is satisfied on the indexing variable and one where it is not. Rule (d) is a combination of the product and sum rule.

According to an embodiment, one or more of the rewrite rules applied by the rewrite system of compiler 122 are supplied by the inference module of the inference algorithm being applied to the provided model (e.g., from inference module 124B). According to an embodiment, one or more of the rewrite rules applied by the rewrite system of compiler 122 are not specific to a particular inference algorithm, or in other words, are universally applied by compiler 122 to intermediate representations without regard to the inference algorithm being used to create the inference code. According to an embodiment, such universally-applied rewrite rules embody standard mathematical rules.

Furthermore, the rules used to rewrite an IR may include one or more rewrite rules that are specific to a class of a probability distribution specified by the model represented by the IR. A class of a probability distribution is identified based on one or more of: a type of one or more distributions represented in the model; a pattern present in the model; etc. According to an embodiment, compiler 122 applies such class-specific rules to an IR to analyze the parallelism in the represented model, or in other words, to identify independent variables in the IR.

Going back to the example of IR expression 700, compiler 122 rewrites IR expression 700 into transformed expression 720 (FIG. 7) using one or more rules that are specific to the Gibbs inference algorithm to be applied to the model. As such, expressions 700 and 720 are both symbolic representations of the same model.

From expression 720, compiler 122 extracts, using one or more IR analysis rules, the information needed to create inference code for the represented model, without creating a graphical representation for the model, since expression 720 is rewritten in order to make such information available to compiler 122. For example, based on one or more IR analysis rules, it is clear from expression 720 that each θ₁, . . . , θ_m is independent of the others after conditioning on the other random variables. As a result, they may all be sampled in parallel.

At each step, compiler 122 can test for a conjugacy relation. From the form of expression 720, compiler 122 recognizes (based on one or more of the IR analysis rules) that the z_mj are drawn from a categorical distribution and θ_m is drawn from a Dirichlet, and can exploit the fact that these are conjugate distributions. The posterior distribution for θ_m is:

Dirichlet (α+c_m)

where c_m is a vector whose kth entry is the number of z associated with document m that were assigned topic k. Importantly, compiler 122 now knows that the drawing of each z must include a counting phase.

IR expression 730 (FIG. 7) illustrates a case of the ϕ variables. For compiler 122 to sample from the distribution represented in expression 730, compiler 122 transforms expression 730 into transformed expression 740 of FIG. 7 using one or more rewrite rules that are specific to the Gibbs inference algorithm. Further, compiler 122 simplifies expression 740 to transformed expression 750 of FIG. 7 using one or more rewrite rules for clustering models. Clustering model rewrite rules may be applied to expression 740 based on compiler 122 identifying one or more patterns in expression 740 that indicate that the represented model is a clustering model. In this manner, compiler 122 performs a domain-specific transformation (or rewrite) that allows compiler 122 to generate inference code for clustering models.

An IR analysis rule that compiler 122 takes advantage of to reach this conclusion identifies the fact that the z are distributed according to a categorical distribution and are used to index into the ϕ array. Therefore, they partition the set of words w into K disjoint sets ₁ . . . _k, one for each topic. More concretely, the probability of words drawn from topic k can be rewritten in partitioned form (using rule (c) of expressions 710) as expression 750 of FIG. 7.

Expression 750 expresses the intuition that once a word's topic is fixed, the word depends on only one of the ϕ_k distributions. In this form, compiler 122 recognizes, based on one or more IR analysis rules, that it should draw from:

Dirichlet(β+c_k)

where c_k is the count of words assigned to topic k. In general, compiler 122 detects patterns like the pattern above when compiler 122 determines that samples drawn from categorical distributions are being used to index into arrays, which is an example of an IR analysis rule.

Finally, compiler 122 turns to analyzing the z_ij based on one or more IR analysis rules. In this case, compiler 122 will again detect that they can be sampled in parallel but it will not be able to detect a conjugacy relationship. Compiler 122 will then detect that the z_ij are drawn from discrete distributions, so that the univariate distribution can be calculated exactly and sampled from. In cases where the distributions are continuous, compiler 122 can try to use another approximate sampling method as a subroutine for drawing that variable.

If compiler 122 fails to find a conjugacy relation from the IR of a model, e.g., if the model has a complicated structure, the compiler 122 will use Metropolis-Hastings and therefore exploit the inherent parallelism of these algorithms.

According to an embodiment, the rewrite rules are applied deterministically and the process will always terminate and produce the same result. Overall, the cost of analysis is negligible compared to the sampling time for large data sets. Although the rewrite system is simple, it enables us to use a concise symbolic representation for the model and thereby scale to large networks.

At times, generation of Gibbs sampler inference code involves many random variables that have interdependencies. To generate parallelized code, compiler 122 finds random variables that are independent (which can be treated in parallel). The analysis (such as rewrite and application of IR analysis rules) of the mathematical formula of an IR identifies independence of variables, which reveals whether variables in the IR are actually dependent on one another. For example, prior to a rewrite of a particular IR, it appears from the IR that X is dependent on Y. However, based on a rewritten version of the particular IR, it becomes clear that Y is a constant. As such, X is not dependent on Y. Thus, X and Y can be treated in parallel.

Exploring a Model Using Multiple Inference Algorithms

According to an embodiment, a user provides compiler 122 with information identifying two or more inference algorithms, i.e., from inference modules 124A-N, on which to base two or more samplers for a provided model. For example, a user requests that compiler 122 produce both a Metropolis-Hastings sampler and a Gibbs sampler for a provided model. As a response to this request, compiler 122 produces both inference code for a Metropolis-Hastings sampler and inference code for a Gibbs sampler on the provided model. In this example, each set of inference code explores all dimensions, or all probability distributions, of the provided model with a single inference algorithm, MH and Gibbs, respectively.

According to another embodiment, the user requests a composite sampler that explores different dimensions of a particular model using different inference algorithms. A model is a product of N primitive distributions (Gaussian, Bernouilli, etc. . . . ). Dimensions of a model are the primitive distributions for the model. In this embodiment, compiler 122 produces inference code that explores at least one dimension, or probability distribution, of a provided model using a first inference algorithm and explores at least one other dimension, or probability distribution, of the provided model using a second (distinct) inference algorithm. For example, a user requests that compiler 122 produce a mixed Metropolis-Hastings, Slice, and Gibbs sampler (applied to various dimensions of a model) for a provided model, where the model includes at least three dimensions. In response to this request, compiler 122 produces inference code for a sampler in which a first dimension of the model is explored using Gibbs inference, a second dimension of the model is explored using Metropolis-Hastings inference, and a third dimension of the model is explored using Slice inference.

When compiler 122 produces one or more sets of inference code based on multiple inference algorithms, compiler 122 uses the rewrite rules from each indicated algorithm to rewrite the IR of the provided model into distinct algorithm-specific expressions that represent the model. For example, if both a Gibbs sampler and an MH sampler are requested, compiler 122 rewrites the IR of the provided model based, at least in part, on the rewrite rules for the Gibbs inference algorithm to produce a Gibbs-specific rewritten expression of the model. Compiler 122 produces the Gibbs inference code based, at least in part, on the Gibbs-specific rewritten expression of the model. Compiler 122 also rewrites the IR of the provided model based, at least in part, on the rewrite rules specific to the MH inference algorithm to produce an MH-specific rewritten expression of the model. Compiler 122 produces the MH inference code based, at least in part, on the MH-specific rewritten expression of the model.

Data-Parallel Operations on Distributions

To produce efficient parallel code, compiler 122 needs to uncover parallelism (as described above), and also pulls efficient routines for inference code from a library of data-parallel operations for distributions. The data-parallel operations for distributions in this library are optimized for implementation on a highly parallelized architecture. Thus, using the operations in this library allows compiler 122 to produce CUDA code to use a GPU effectively, i.e., using operations that are optimized for GPU execution.

For instance, in the case of LDA, there are two steps in which compiler 122 emits inference code that draws from many Dirichlet distributions in parallel. In the first case, when drawing the topic distributions for the documents, each thread can draw one of the θ_i by generating K Gamma variates and normalizing them. Since the number of documents is usually very large, this produces enough parallelism to make full use of the GPU's cores.

However, this will not produce sufficient parallelism when drawing the ϕ_k, because the number of topics is usually small compared to the number of cores. Consequently, compiler 122 uses a different procedure (algorithm 800 of FIG. 8) to sample from K Dirichlet variates.

To generate K Dirichlet variates over V categories with concentration parameters α₁₁, . . . , α_KV, compiler 122 emits inference code (e.g., based on a library function) that first generates a matrix A where A_ij˜Gamma(α_ij) and then normalizes each row of this matrix. For sampling the ϕ_i, compiler 122 emits inference code that causes the launch of a thread to generate the gamma variates for each column (rather than for each row), and then separately computes a normalizing constant for each row by multiplying the matrix by an all-ones vector, e.g., using CUDA Basic Linear Algebra Subroutines (cuBLAS). Such an implementation is more effective than creating a thread for each row when the number of columns is larger than the number of rows. This is an instance where the two stage compilation procedure is useful, because compiler 122 is able to use information about the relative sizes of K and V to decide whether algorithm 800 will be more efficient than the simple scheme.

This sort of optimization is not unique to the Dirichlet distribution. For example, when generating a large number of multinormal variates by applying a linear transformation to a vector of normal variates, the strategy for extracting parallelism may change based on the number of variates to generate, the dimension of the multinormal, and the number of GPU cores.

Segmented Data-Parallel Operators

Furthermore, compiler 122 uses segmented data-parallel operators to produce efficient GPU inference code when ragged arrays are included in the specification of a model. Parallelization of ragged arrays can result in inefficient GPU code because of differences in the amounts of data in the various sub-arrays of a ragged array. To illustrate, if the processing of information in sub-arrays of a ragged array is assigned to cores of a GPU on a sub-array by sub-array basis, some of the cores will receive more work (i.e., longer sub-arrays) than others, leaving some GPU cores idle while others are still working. However, using segmented data-parallel operators, compiler 122 divides up the computation work needed to create the inference code based on a ragged array among the cores of a GPU without regard to the boundaries of the sub-arrays of the ragged array, which allows the work to be distributed more evenly among the GPU cores.

Because compiler 122 tries to parallelize operations over arrays, users can maximize the amount of parallelism in their models by structuring them so that data and parameters are stored in large, flattened arrays. In addition, as more options and inference strategies are added to the framework of compiler 122, users will be able to experiment with the tradeoffs of different inference algorithms in a way that would be too time-consuming to do manually.

Hardware Overview

According to one embodiment, the techniques described herein are implemented by one or more special-purpose computing devices. The special-purpose computing devices may be hard-wired to perform the techniques, or may include digital electronic devices such as one or more application-specific integrated circuits (ASICs) or field programmable gate arrays (FPGAs) that are persistently programmed to perform the techniques, or may include one or more general purpose hardware processors programmed to perform the techniques pursuant to program instructions in firmware, memory, other storage, or a combination. Such special-purpose computing devices may also combine custom hard-wired logic, ASICs, or FPGAs with custom programming to accomplish the techniques. The special-purpose computing devices may be desktop computer systems, portable computer systems, handheld devices, networking devices or any other device that incorporates hard-wired and/or program logic to implement the techniques.

For example, FIG. 9 is a block diagram that illustrates a computer system 900 upon which an embodiment of the invention may be implemented. Computer system 900 includes a bus 902 or other communication mechanism for communicating information, and a hardware processor 904 coupled with bus 902 for processing information. Hardware processor 904 may be, for example, a general purpose microprocessor.

Computer system 900 also includes a main memory 906, such as a random access memory (RAM) or other dynamic storage device, coupled to bus 902 for storing information and instructions to be executed by processor 904. Main memory 906 also may be used for storing temporary variables or other intermediate information during execution of instructions to be executed by processor 904. Such instructions, when stored in non-transitory storage media accessible to processor 904, render computer system 900 into a special-purpose machine that is customized to perform the operations specified in the instructions.

Computer system 900 further includes a read only memory (ROM) 908 or other static storage device coupled to bus 902 for storing static information and instructions for processor 904. A storage device 910, such as a magnetic disk, optical disk, or solid-state drive is provided and coupled to bus 902 for storing information and instructions.

Computer system 900 may be coupled via bus 902 to a display 912, such as a cathode ray tube (CRT), for displaying information to a computer user. An input device 914, including alphanumeric and other keys, is coupled to bus 902 for communicating information and command selections to processor 904. Another type of user input device is cursor control 916, such as a mouse, a trackball, or cursor direction keys for communicating direction information and command selections to processor 904 and for controlling cursor movement on display 912. This input device typically has two degrees of freedom in two axes, xkillplane.

Computer system 900 may implement the techniques described herein using customized hard-wired logic, one or more ASICs or FPGAs, firmware and/or program logic which in combination with the computer system causes or programs computer system 900 to be a special-purpose machine. According to one embodiment, the techniques herein are performed by computer system 900 in response to processor 904 executing one or more sequences of one or more instructions contained in main memory 906. Such instructions may be read into main memory 906 from another storage medium, such as storage device 910. Execution of the sequences of instructions contained in main memory 906 causes processor 904 to perform the process steps described herein. In alternative embodiments, hard-wired circuitry may be used in place of or in combination with software instructions.

The term “storage media” as used herein refers to any non-transitory media that store data and/or instructions that cause a machine to operate in a specific fashion. Such storage media may comprise non-volatile media and/or volatile media. Non-volatile media includes, for example, optical disks, magnetic disks, or solid-state drives, such as storage device 910. Volatile media includes dynamic memory, such as main memory 906. Common forms of storage media include, for example, a floppy disk, a flexible disk, hard disk, solid-state drive, magnetic tape, or any other magnetic data storage medium, a CD-ROM, any other optical data storage medium, any physical medium with patterns of holes, a RAM, a PROM, and EPROM, a FLASH-EPROM, NVRAM, any other memory chip or cartridge.

Storage media is distinct from but may be used in conjunction with transmission media. Transmission media participates in transferring information between storage media. For example, transmission media includes coaxial cables, copper wire and fiber optics, including the wires that comprise bus 902. Transmission media can also take the form of acoustic or light waves, such as those generated during radio-wave and infra-red data communications.

Various forms of media may be involved in carrying one or more sequences of one or more instructions to processor 904 for execution. For example, the instructions may initially be carried on a magnetic disk or solid-state drive of a remote computer. The remote computer can load the instructions into its dynamic memory and send the instructions over a telephone line using a modem. A modem local to computer system 900 can receive the data on the telephone line and use an infra-red transmitter to convert the data to an infra-red signal. An infra-red detector can receive the data carried in the infra-red signal and appropriate circuitry can place the data on bus 902. Bus 902 carries the data to main memory 906, from which processor 904 retrieves and executes the instructions. The instructions received by main memory 906 may optionally be stored on storage device 910 either before or after execution by processor 904.

Computer system 900 also includes a communication interface 918 coupled to bus 902. Communication interface 918 provides a two-way data communication coupling to a network link 920 that is connected to a local network 922. For example, communication interface 918 may be an integrated services digital network (ISDN) card, cable modem, satellite modem, or a modem to provide a data communication connection to a corresponding type of telephone line. As another example, communication interface 918 may be a local area network (LAN) card to provide a data communication connection to a compatible LAN. Wireless links may also be implemented. In any such implementation, communication interface 918 sends and receives electrical, electromagnetic or optical signals that carry digital data streams representing various types of information.

Network link 920 typically provides data communication through one or more networks to other data devices. For example, network link 920 may provide a connection through local network 922 to a host computer 924 or to data equipment operated by an Internet Service Provider (ISP) 926. ISP 926 in turn provides data communication services through the world wide packet data communication network now commonly referred to as the “Internet” 928. Local network 922 and Internet 928 both use electrical, electromagnetic or optical signals that carry digital data streams. The signals through the various networks and the signals on network link 920 and through communication interface 918, which carry the digital data to and from computer system 900, are example forms of transmission media.

Computer system 900 can send messages and receive data, including program code, through the network(s), network link 920 and communication interface 918. In the Internet example, a server 930 might transmit a requested code for an application program through Internet 928, ISP 926, local network 922 and communication interface 918.

The received code may be executed by processor 904 as it is received, and/or stored in storage device 910, or other non-volatile storage for later execution.

Cloud Computing Overview

The techniques described herein are implemented using one or more processing solutions, examples of which include distributed systems, clustered computing systems, and cloud computing systems. In an embodiment, compiler 122 is part of a cloud computing system. A cloud computing system implements one or more of: cloud storage, cloud processing, cloud communication, and any other kind of cloud computing service. Further, cloud computing systems may operate under a pay-for-what-you-use-as-you-use-it model, under a fixed subscription model, etc. In this embodiment, any part (or the whole of) the functionality attributed to compiler 122, or to other entities within this description, is controllable via an interface that is exposed at a cloud computing system.

In the foregoing specification, embodiments of the invention have been described with reference to numerous specific details that may vary from implementation to implementation. The specification and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense. The sole and exclusive indicator of the scope of the invention, and what is intended by the applicants to be the scope of the invention, is the literal and equivalent scope of the set of claims that issue from this application, in the specific form in which such claims issue, including any subsequent correction.

Claims

1. A method comprising: compiling source code, representing a model of a body of data, into an intermediate representation of the model;wherein the model of the body of data specifies one or more probability distributions;wherein the intermediate representation of the model represents at least one probability distribution, of the one or more probability distributions, with one or more mathematical expressions;identifying a particular inference algorithm of a plurality of inference algorithms; andproducing inference code that implements the particular inference algorithm;wherein the inference code is configured to sample from the at least one probability distribution of the one or more probability distributions; andwherein producing the inference code is based, at least in part, on an analysis of the one or more mathematical expressions;wherein the method is performed by one or more computing devices.

2. The method of claim 1 further comprising: prior to producing the inference code, performing the analysis of the one or more mathematical expressions;wherein performing the analysis of the one or more mathematical expressions comprises identifying one or more features of the at least one probability distribution that are represented in the one or more mathematical expressions.

3. The method of claim 1, further comprising: prior to identifying the particular inference algorithm, receiving information indicating one or more rules for analyzing intermediate representations of models;wherein the one or more rules are specific to the particular inference algorithm.

4. The method of claim 3 further comprising: prior to producing the inference code, performing the analysis of the one or more mathematical expressions;wherein performing the analysis of the one or more mathematical expressions comprises: rewriting the one or more mathematical expressions based, at least in part, on the one or more rules that are specific to the particular inference algorithm, to produce one or more transformed mathematical expressions, andidentifying features of the at least one probability distribution based, at least in part, on analysis of the one or more transformed mathematical expressions.

5. The method of claim 1, further comprising performing the analysis of the one or more mathematical expressions without materializing, in computer memory, values for data points of a graphical representation of the at least one probability distribution.

6. The method of claim 1 further comprising: identifying a second inference algorithm, of the plurality of inference algorithms; andproducing second inference code that implements the second inference algorithm;wherein the second inference code is configured to sample from the at least one probability distribution of the one or more probability distributions; andwherein producing the second inference code is based, at least in part, on a second analysis of the one or more mathematical expressions.

7. The method of claim 6 wherein: the inference code is first inference code;the first inference code, implementing the particular inference algorithm, explores all of the one or more probability distributions specified in the model of the body of data; andthe second inference code, implementing the second inference algorithm, explores all of the one or more probability distributions specified in the model of the body of data.

8. The method of claim 1 further comprising: identifying a second inference algorithm, of the plurality of inference algorithms; andproducing second inference code that implements the second inference algorithm;wherein the second inference code is configured to sample from a second probability distribution of the one or more probability distributions;wherein the second probability distribution is not a probability distribution of the at least one probability distribution; andwherein producing the second inference code is based, at least in part, on a second analysis of the one or more mathematical expressions.

9. The method of claim 1 further comprising: performing the analysis of the one or more mathematical expressions;wherein performing the analysis of the one or more mathematical expressions comprises identifying independencies between variables in the one or more mathematical expressions; andwherein producing the inference code comprises building parallelism into the inference code based, at least in part, on the identified independencies.

10. The method of claim 9 wherein performing the analysis of the one or more mathematical expressions further comprises rewriting the one or more mathematical expressions based, at least in part, on one or more rewrite rules that are specific to a class of the at least one probability distribution, to produce one or more transformed mathematical expressions.

11. One or more non-transitory computer-readable media storing instructions that, when executed by one or more processors, cause: compiling source code, representing a model of a body of data, into an intermediate representation of the model;wherein the model of the body of data specifies one or more probability distributions;wherein the intermediate representation of the model represents at least one probability distribution, of the one or more probability distributions, with one or more mathematical expressions;identifying a particular inference algorithm of a plurality of inference algorithms; andproducing inference code that implements the particular inference algorithm;wherein the inference code is configured to sample from the at least one probability distribution of the one or more probability distributions; andwherein producing the inference code is based, at least in part, on an analysis of the one or more mathematical expressions.

12. The one or more non-transitory computer-readable media of claim 11 wherein the instructions further comprise instructions that, when executed by one or more processors, cause: prior to producing the inference code, performing the analysis of the one or more mathematical expressions;wherein performing the analysis of the one or more mathematical expressions comprises identifying one or more features of the at least one probability distribution that are represented in the one or more mathematical expressions.

13. The one or more non-transitory computer-readable media of claim 11, wherein the instructions further comprise instructions that, when executed by one or more processors, cause: prior to identifying the particular inference algorithm, receiving information indicating one or more rules for analyzing intermediate representations of models;wherein the one or more rules are specific to the particular inference algorithm.

14. The one or more non-transitory computer-readable media of claim 13 wherein the instructions further comprise instructions that, when executed by one or more processors, cause: prior to producing the inference code, performing the analysis of the one or more mathematical expressions;wherein performing the analysis of the one or more mathematical expressions comprises: rewriting the one or more mathematical expressions based, at least in part, on the one or more rules that are specific to the particular inference algorithm, to produce one or more transformed mathematical expressions, andidentifying features of the at least one probability distribution based, at least in part, on analysis of the one or more transformed mathematical expressions.

15. The one or more non-transitory computer-readable media of claim 11, wherein the instructions further comprise instructions that, when executed by one or more processors, cause performing the analysis of the one or more mathematical expressions without materializing, in computer memory, values for data points of a graphical representation of the at least one probability distribution.

16. The one or more non-transitory computer-readable media of claim 11 wherein the instructions further compromise instructions that, when executed by one or more processors, cause: identifying a second inference algorithm, of the plurality of inference algorithms; andproducing second inference code that implements the second inference algorithm;wherein the second inference code is configured to sample from the at least one probability distribution of the one or more probability distributions; andwherein producing the second inference code is based, at least in part, on a second analysis of the one or more mathematical expressions.

17. The one or more non-transitory computer-readable media of claim 16 wherein: the inference code is first inference code;the first inference code, implementing the particular inference algorithm, explores all of the one or more probability distributions specified in the model of the body of data; andthe second inference code, implementing the second inference algorithm, explores all of the one or more probability distributions specified in the model of the body of data.

18. The one or more non-transitory computer-readable media of claim 11 wherein the instructions further comprise instructions that, when executed by one or more processors, cause: identifying a second inference algorithm, of the plurality of inference algorithms; andproducing second inference code that implements the second inference algorithm;wherein the second inference code is configured to sample from a second probability distribution of the one or more probability distributions;wherein the second probability distribution is not a probability distribution of the at least one probability distribution; andwherein producing the second inference code is based, at least in part, on a second analysis of the one or more mathematical expressions.

19. The one or more non-transitory computer-readable media of claim 11 wherein the instructions further comprise instructions that, when executed by one or more processors, cause: performing the analysis of the one or more mathematical expressions;wherein performing the analysis of the one or more mathematical expressions comprises identifying independencies between variables in the one or more mathematical expressions; andwherein producing the inference code comprises building parallelism into the inference code based, at least in part, on the identified independencies.

20. The one or more non-transitory computer-readable media of claim 19 wherein performing the analysis of the one or more mathematical expressions further comprises rewriting the one or more mathematical expressions based, at least in part, on one or more rewrite rules that are specific to a class of the at least one probability distribution, to produce one or more transformed mathematical expression.