Neural Probabilistic Logic Programming in DeepProbLog^{1}^{1}1This is an extended and revised version of work previously published at NeurIPS 2018 (deepproblog).
Abstract
We introduce DeepProbLog, a neural probabilistic logic programming language that incorporates deep learning by means of neural predicates. We show how existing inference and learning techniques of the underlying probabilistic logic programming language ProbLog can be adapted for the new language. We theoretically and experimentally demonstrate that DeepProbLog supports {enumerate*}[label=()]
both symbolic and subsymbolic representations and inference,
program induction,
probabilistic (logic) programming, and
(deep) learning from examples. To the best of our knowledge, this work is the first to propose a framework where generalpurpose neural networks and expressive probabilisticlogical modeling and reasoning are integrated in a way that exploits the full expressiveness and strengths of both worlds and can be trained endtoend based on examples.
keywords:
logic, probability, neural networks, probabilistic logic programming, neurosymbolic integration, learning and reasoning[topline=false,rightline=false,bottomline=false]exampleExample
1 Introduction
Many tasks in AI can be divided into roughly two categories: those that require lowlevel perception, and those that require highlevel reasoning. At the same time, there is a growing consensus that being capable of tackling both types of tasks is essential to achieve true (artificial) intelligence (kahneman). Deep learning is empowering a new generation of intelligent systems that excel at lowlevel perception, where it is used to interpret images, text and speech with unprecedented accuracy. The success of deep learning has caused a lot of excitement and has also created the impression that deep learning can solve any problem in artificial intelligence. However, there is a growing awareness of the limitations of deep learning: deep learning requires large amounts of (the right kind of) data to train the network, it provides neither justifications nor explanations, and the models are blackboxes that can neither be understood nor modified by domain experts. Although there have been attempts to demonstrate reasoninglike behaviour with deep learning (santoro2017simple), their current reasoning abilities are nowhere close to what is possible with typical highlevel reasoning approaches. The two most prominent frameworks for reasoning are logic and probability theory. While in the past, these were studied by separate communities in artificial intelligence, many researchers are working towards their integration, and aim at combining probability with logic and statistical learning; cf. the areas of statistical relational artificial intelligence (DeRaedt16; Getoor07) and probabilistic logic programming (deraedt15).
The abilities of deep learning and statistical relational artificial intelligence approaches are complementary. While deep learning excels at lowlevel perception, probabilistic logics excel at highlevel reasoning. As such, an integration of the two would have very promising properties. Recently, a number of researchers have revisited and modernized ideas originating from the field of neuralsymbolic integration (garcez2012neural), searching for ways to combine the best of both worlds (riedel2017programming; rocktaschel2017end; cohen2018tensorlog; santoro2017simple), for example, by designing neural architectures representing differentiable counterparts of symbolic operations in classical reasoning tools. Yet, joining the full flexibility of highlevel probabilistic and logical reasoning with the representational power of deep neural networks is still an open problem. Elsewhere nesy, we have argued that neurosymbolic integration should: 1) integrate neural networks with the two most prominent methods for reasoning, that is, logic and probability, and 2) that neurosymbolic integrated methods should have the pure neural, logical and probabilistic methods as special cases.
With DeepProbLog, we tackle the neurosymbolic challenge from this perspective. Furthermore, instead of integrating reasoning capabilities into a complex neural network architecture, we proceed the other way round. We start from an existing probabilistic logic programming language, ProbLog (DeRaedt07), and introduce the smallest extension that allows us to integrate neural networks: the neural predicate. The idea is simple: in a probabilistic logic, atomic expressions of the form (aka tuples in a relational database) have a probability . We extend this idea by allowing atomic expressions to be labeled with neural networks whose outputs can be considered probability distributions. This simple idea is appealing as it allows us to retain all the essential components of the ProbLog language: the semantics, the inference mechanism, as well as the implementation.
Therefore, one should not only integrate logic with neural networks in neurosymbolic computation, but also probability.
This effectively leads to an integration of probabilistic logics (hence statistical relational AI) with neural networks and opens up new abilities. Furthermore, although at first sight, this may appear as a complication, it actually can greatly simplify the integration of neural networks with logic. The reason for this is that the probabilistic framework provides a clear optimisation criterion, namely the probability of the training examples. Realvalued probabilistic quantities are also wellsuited for gradientbased training procedures, as opposed to discrete logic quantities.
Example 1
Before going into further detail, the following example illustrates the possibilities of this approach. Consider the predicate , where and are images of digits and is the natural number corresponding to the sum of these digits. The goal is that after training, DeepProbLog allows us to make a probabilistic estimate on the validity of, for example, the example . While such a predicate can be learned directly by a standard neural classifier, such an approach cannot incorporate background knowledge such as the definition of the addition of two natural numbers. In DeepProbLog such knowledge can easily be encoded in rules such as
with is the standard operator of logic programming to evaluate arithmetic expressions. All that needs to be learned in this case is the neural predicate which maps an image of a digit to the corresponding natural number . The trained network can then be reused for arbitrary tasks involving digits. Our experiments show that this leads not only to new capabilities but also to significant performance improvements. An important advantage of this approach compared to standard image classification settings is that it can be extended to multidigit numbers without additional training. We note that the single digit classifier (i.e., the neural predicate) is not explicitly trained by itself: its output can be considered a latent representation, as we only use training data with pairwise sums of digits.
To summarize, we introduce DeepProbLog which has a unique set of features: {enumerate*}[label=()]
it is a programming language that supports neural networks and machine learning and has a welldefined semantics
it integrates logical reasoning with neural networks; so both symbolic and subsymbolic representations and inference;
it integrates probabilistic modeling, programming and reasoning with neural networks (as DeepProbLog extends the probabilistic programming language ProbLog, which can be regarded as a very expressive directed graphical modeling language (DeRaedt16));
it can be used to learn a wide range of probabilistic logical neural models from examples, including inductive programming.
This paper is a significantly extended and completed version of our previous work deepproblog (NeurIPS, spotlight presentation). This extended version now contains the necessary deep learning and probabilistic logic programming background and a more in depth theoretical explanation. It also contains additional experiments (see Section 6): the MNIST addition experiments from the short version are completed with the new experiments T3 and T4, and we designed new experiments (T8 and T9) to further investigate the use of DeepProbLog on combined probabilistic learning and deep learning. The code is available at https://bitbucket.org/problog/deepproblog.
2 Background
2.1 Logic programming concepts
In this section, we briefly summarize basic logic programming concepts; see e.g., lloyd:book89 for more details. Atoms are expressions of the form where is a predicate (of arity , or in shorthand notation) and the are terms. A literal is an atom or the negation of an atom. A term is either a constant , a variable , or a structured term of the form where is a functor and the are terms. We follow the Prolog convention and let constants, functors and predicates start with a lower case character and variables with an upper case. A rule is an expression of the form where is an atom, the are literals, and all variables are universally quantified. Informally, the meaning of such a rule is that holds whenever the conjunction of the holds. Thus represents logical implication (), and the comma () represents conjunction (). Rules with an empty body are called facts. A logic program is a finite set of rules.
A substitution is an assignment of terms to variables . When applying a substitution to an expression we simultaneously replace all occurrences of by and denote the resulting expression as . Expressions that do not contain any variables are called ground. The Herbrand base of a logic program is the set of ground atoms that can be constructed using the predicates, functors and constants occurring in the program.^{4}^{4}4If the program does not contain constants, one arbitrary constant is added. Subsets of the Herbrand base are called Herbrand interpretations. A Herbrand interpretation is a model of a clause if for every substitution such that the conjunction holds in the interpretation, is in the interpretation.
It is a model of a logic program if it is a model of all clauses in the program.
For negationfree programs, the semantics is given by the minimal such model, known as the least Herbrand model, which is unique. General logic programs use the notion of negation as failure, that is, the negation of an atom is true exactly if the atom cannot be derived from the program. These programs are not guaranteed to have a unique minimal Herbrand model, and several ways to define a canonical model have been studied. We follow the wellfounded semantics here (VanGelder91).
The main inference task in logic programming is to determine whether a given atom , also called query (or goal), is true in the canonical model of a logic program , denoted by . If the answer is yes (or no), we also say that the query succeeds (or fails). If such a query is not ground, inference asks for the existence of an answer substitution, that is, a substitution that grounds the query into an atom that is part of the canonical model.
2.2 Deep Learning
The following paragraphs provide a very brief introduction to deep learning, focusing on concepts needed for understanding our work. Extensive further details can be found, e.g., in Goodfellow2016. This section is meant to provide readers with little or no knowledge of deep learning, with a conceptual understanding of the main ideas. In particular, we will focus on the setting of supervised learning, where the model learns to map an input item to a particular output, based on inputoutput examples.
An artificial neural network is a highly parameterized and therefore very flexible nonlinear mathematical function that can be ‘trained’ towards a particular desired behavior, by suitably adjusting its parameters.
During training, the model learns to capture from the input data the most informative ‘features’ for the task at hand. The need for ‘feature engineering’ in classical (or rather, nonneural) machine learning methods has therefore been replaced by ‘architecture engineering’, since a wide variety of neural network components are available to be composed into a suitable model.
Deep neural networks are often designed and trained in an ‘endtoend’ fashion, whereby only the raw input and the final target are known during training, and all components of the model are jointly trained. For example, for the task of handwritten digit recognition, an input instance consists of a pixel image of a handwritten digit, whereas its target denotes the actual digit.
Consider a supervised learning problem, with a training set containing i.i.d. input instances and corresponding outputs . A model represented by a mapping function with parameters , maps an input item to the corresponding predicted output .
To quantify how strongly the predicted output deviates from the target output , a loss function is defined. Training the model then comes down to minimizing the expected loss over the training set. In the specific setting of multiclass classification, each input instance corresponds to one out of a fixed set of output categories. The target vectors are typically represented as onehot vectors: all components are zero, except at index of the corresponding category. The predicted counterpart at the model’s output is often obtained by applying a socalled softmax output function to intermediate realvalued scores obtained at the output of the neural network.
The ’th component of the softmax is defined as
The softmax outputs are wellsuited to model a probability distribution (i.e., and ). The standard corresponding loss function is the crossentropy loss, which quantifies the deviation between the empirical output distribution (i.e., the softmax outputs) and the ground truth distribution (i.e., the onehot target vector ) defined as
The most widely used optimization approaches for neural networks are variations of the gradient descent algorithm, in which the parameters are iteratively updated by taking small steps along the negative gradient of the loss. An estimate at iteration is updated as , in which the step size is controlled by the learning rate . Typically, training is not performed over the entire dataset per iteration, but instead over a smaller ‘minibatch’ of instances. This is computationally more efficient and allows for a better exploration of parameter space. Importantly, the loss gradient can only be calculated if all components of the neural network are differentiable.
A deep neural network typically has a layerwise architecture: the different layers correspond to nested differentiable functions in the overall mapping function . The ‘forward pass’ through the network corresponds to consecutively applying these layer functions to a given input to the network. The intermediate representations obtained by evaluating these layer functions are called hidden states. After a forward pass, the gradient with respect to all parameters can then be calculated by applying the chain rule. This happens during the socalled ‘backward pass’: the gradients are calculated from the output back to the first layer. As an illustration of how the chain rule is applied, consider the network function , which contains a first layer represented by the vector function , and a second layer . For simplicity, say each layer has one trainable parameter, respectively written as and . The derivative with respect to these parameters of a scalar loss function applied to the network output, becomes
in which the individual derivatives are evaluated based on the considered input and current value of the parameters. The entire procedure to calculate the gradients is called the backpropagation algorithm. It requires a forward pass to calculate all intermediate representations up to the value of the loss. After that, in the backward pass, the gradients corresponding to all operations applied during the forward pass, are iteratively calculated, starting at the loss (i.e., with in the example). As such, the gradients with respect to parameters at a given layer can be calculated as soon as the gradients due to all operations further in the network are known, as governed by the chain rule.
To summarize, a single iteration in the optimization happens as follows: 1) A minibatch is sampled from the training data. 2) The output of the neural network is calculated during the forward pass. 3) The loss is calculated based on that output and the target. 4) The gradients for the parameters in the neural network are calculated using backpropagation. 5) The parameters are updated using a gradientbased optimizer.
The most basic neural networks building block is the socalled fullyconnected layer. It consists of a linear transformation with weight matrix and bias vector , followed by applying a componentwise nonlinear function, called the activation function. The input into such a layer can be the vector representation of the actual input to the model, or the output from a previous layer , which is called a hidden representation. Its output is calculated as , in which typical choices for the activation function are the Rectified Linear Unit (ReLU) defined as or the hyperbolic tangent . In other cases, a sigmoid activation can be used, given by . Another important type of neural network layer is the convolutional layer, which convolves the input to pass it to the next layer, by means of a kernel with trainable weights, typically much smaller than the input size. Convolutional layers, followed by a similar activation function, are often used in image recognition models, whereby subsequent layers learn to extract useful features for the given task, from local patterns up to more global and often interpretable patterns. An architecture wellsuited for modeling sequences are the socalled recurrent neural networks (RNN). In short, these define a mapping from an input element in the considered sequence into a hidden representation. Every input element is encoded with the same neural network, called the RNN ‘cell’, such that the model can be applied to variablelength sequences. In order to explicitly model the sequential behavior, when encoding a given item in the sequence, the cell takes as input that item, as well as the hidden state obtained while encoding the previous input item. When training with this recurrent setup, the gradient propagates back through the entire sequence. When encoding long sequences, this may lead to very small gradients. An important type of RNN, wellequipped to deal with this socalled vanishing gradient problem, is the Long Shortterm Memory (LSTM). The same problem is solved by a similar architecture called the GRU.
More technical details, as well as several other popular types of neural network components, are provided in Goodfellow2016.
Deep neural networks can become very expressive, especially when deeper, or with large hidden representations. To avoid overfitting, various regularization approaches have been developed. A widespread technique, also used in some of the presented experiments in this work, is called dropout. For those layers on which dropout is applied, during training a random sample of the layer outputs are set to zero in each iteration, while accordingly compensating the amplitude of the remaining activations. At inference time, i.e., when applying the trained model to heldout data, all activations are kept.
As mentioned, many choices are possible in terms of architecture and training: dimensions, types of layers, learning rates, regularization strength, etc. These are socalled hyperparameters, and are typically ‘tuned’ by evaluating on a validation set, not used explicitly for gradientbased training of the network parameters, and still separate from the final test data.
3 Introducing DeepProbLog
We now recall the basics of probabilistic logic programming using ProbLog (see deraedt15 for more details), and then introduce our new language DeepProbLog.
3.1 ProbLog
Definition 1 (ProbLog program)
A ProbLog program consists of a set of ground probabilistic facts of the form where is a probability and a ground atom and a set of rules .
For instance, the following ProbLog program models a variant of the wellknown alarm Bayesian network (Pearl88:book):
Each probabilistic fact corresponds to an independent Boolean random variable that is true with probability and false with probability . Every subset defines a possible world = and is ground, that is, the world is the canonical model of the logic program obtained by adding to the set of rules , e.g.,
To keep the presentation simple, we focus on the case of finitely many ground probabilistic facts, but note that the semantics is also welldefined for the countably infinite case. The probability of such a possible world is given by the product of the probabilities of the truth values of the probabilistic facts:
(1) 
For instance,
The probability of a ground fact , also called success probability of , is then defined as the sum of the probabilities of all worlds containing , i.e.,
(2) 
The probability of a query is also equal to the weighted model count (WMC) of the worlds where this query is true.
For ease of modeling, ProbLog supports nonground probabilistic facts as a shortcut for introducing a set of ground probabilistic facts, as well as annotated disjunctions (ADs), which are expressions of the form
where the are probabilities that sum to at most one, the are atoms, and the are literals. The meaning of an AD is that whenever all hold, the AD causes one of the to be true, or none of them with probability . Note that several of the may be true at the same time if they also appear as heads of other rules or ADs. This is convenient to model choices between different categorical variables, e.g. different severities of the earthquake:
or without explicitly representing the event of no earthquake:
In which neither earthquake(mild) nor earthquake(severe) will be true with probability . Annotated disjunctions do not change the expressivity of ProbLog, as they can alternatively be modeled through independent facts and logical rules; we refer to deraedt15 for technical details.
3.2 DeepProbLog
In ProbLog, the probabilities of all random choices are explicitly specified as part of probabilistic facts or annotated disjunctions. DeepProbLog extends ProbLog to basic random choices whose probabilities are specified through external functions implemented as neural networks.
Definition 2 (Neural annotated disjunction)
A neural AD is an expression of the form
where is a reserved functor, uniquely identifies a neural network model with inputs and outputs (i.e., its architecture as well as its trainable parameters) defining a probability distribution over classes, is a sequence of input variables, is the output variable, is a sequence of ground terms (the output domain of this neural network) and is a predicate.
A ground neural AD is an expression of the form
where is a sequence of ground terms (the input to the neural network) and are ground terms (the output domain of this neural network).
The term in the definition can be considered a function that returns the probability of class when evaluating the network on input . As such, a ground nAD can be instantiated into a normal AD by evaluating the neural network and replacing the functor with the calculated probability. For instance, in the MNIST addition example, we would specify the nAD
where m_digit is a network that classifies MNIST digits. Grounding this on an input image would result in a ground nAD:
Evaluating this would result in a ground AD:
Where is the output vector of the m_digit network when evaluated on
.
The neural network could take any shape, e.g., a convolutional network for image encoding, a recurrent network for sequence encoding, etc. However, its output layer, which feeds the corresponding neural predicate, needs to be normalized.
We consider an output domain size of two as a special case. Instead of the neural network having two probabilities at the output that sum to one, we can simplify this to a single probability, with the second one the complement of that probability. This difference coincides with the difference between a softmax and singleneuron sigmoid layer in a neural network. We call such an expression a neural fact.
Definition 3 (Neural fact)
A neural fact is an expression of the form
where is a reserved functor, uniquely identifies a neural network model defining a probability distribution over classes, is a sequence of input variables and is a predicate.
A ground neural fact is an expression of the form
where is a sequence of ground terms (the input to the neural network).
To exemplify, we use a neural network that gives a measure of the similarity between two input images. We can encode this with the following neural fact:
Grounding this on the input and would result in the follow ground neural fact:
Evaluating this would result in a ground probabilistic fact:
Where is the output of the network when evaluated on and .
Definition 4 (DeepProbLog Program)
A DeepProbLog program consists of a set of ground probabilistic facts , a set of ground neural ADs and ground neural facts , and a set of rules .
The semantics of a DeepProbLog program is given by the semantics of the ProbLog program obtained by replacing each nAD with the AD obtained by instantiating the probabilities as mentioned above. While the semantics is defined with respect to ground neural ADs and facts, as in ProbLog, we write nonground such expressions if the intended grounding is clear from context.
4 Inference
This section explains how a DeepProbLog model is used for a given query at prediction time. First, we provide more detail on ProbLog inference (Fierens). Next, we describe how ProbLog inference is adapted in DeepProbLog.
4.1 ProbLog Inference
ProbLog inference proceeds in four steps.
The first step is the grounding step, in which the logic program is grounded with respect to the query. This step uses backward reasoning to determine which ground rules are relevant to derive the truth value of the query, and may perform additional logical simplifications that do not affect the query’s probability.
The second step rewrites the ground logic program into a formula in propositional logic that defines the truth value of the query in terms of the truth values of probabilistic facts. We can calculate the query success probability by performing weighted model counting (WMC) on this logic formula (cfr. Fierens). However, performing WMC on this logical formula directly is not efficient.
The third step is knowledge compilation (DarwicheMarquis). During this step, the logic formula is transformed into a form that allows for efficient weighted model counting.
The current ProbLog system uses Sentential Decision Diagrams (SDDs, darwiche2011sdd), the most succinct suitable representation available today.
SDDs, being a subset of dDNNFs allow for polytime model counting (DarwicheMarquis). However, they also support polytime conjunction, disjunction and negation while being more succinct than OBDDs (darwiche2011sdd).
The fourth and final step transforms the SDD into an arithmetic circuit (AC). This is done by putting the probabilities of the probabilistic facts or their negations on the leaves, replacing the OR nodes with addition and the AND nodes by multiplication. The WMC is then calculated with an evaluation of the AC.

Example 2
In Figure 1, we apply the four steps of ProbLog inference on the earthquake example
with query calls(mary).
In the first step, the nonground program (Figure 0(a)) is grounded with respect to the query calls(mary). The result is shown in Figure 0(b): the irrelevant fact hears_alarm(john) is omitted and the variable X in the calls rule is substituted with the constant mary.
The resulting formula in the second step is
The WMC of this formula is shown in Figure 0(c). However, it is not calculated by enumeration as shown here, but an AC is used instead. The AC derived in step four is shown in Figure 0(d), where rounded grey rectangles depict variables corresponding to probabilistic facts, and the rounded red rectangle denotes the query atom defined by the formula. The white rectangles correspond to logical operators applied to their children. The intermediate results are shown in black next to the nodes in Figure 0(d).
4.2 DeepProbLog Inference
The only change required for DeepProbLog inference is that we need to instantiate the ground nADs and neural facts into the corresponding ground ADs and ground facts. This is done in a separate step after grounding, where the parameters for the regular AD are determined by making a forward pass on the relevant neural network with the ground input.
Example 3
We illustrate this by evaluating the MNIST addition example (Figure 1(a)). The DeepProbLog program requires two lines: the first line defining the neural predicate, and the second line defining the addition. We evaluate it on the query . In the first step, the DeepProbLog program is grounded into a ground DeepProbLog Program (Figure 1(b)). Note that the nADs are now all ground. As ProbLog only grounds the relevant part of the program, i.e. the part that can be used to prove the query, only the digits and are retained as the larger digits cannot sum to . The next step is the only difference between ProbLog and DeepProbLog inference: instantiating the ground nADs into regular ground ADs, which could, for instance, produce an AD as shown in Figure 1(c). The probabilities in the instantiated ADs do not sum to one, as the irrelevant terms (, …, and , …, ) have been dropped in the grounding process, although the neural network still assigns probability mass to them. Inference then proceeds identically to that of ProbLog: the ground program is rewritten into a logical formula, this formula is compiled and transformed into an AC. Finally, this AC is evaluated to calculate the query probability.
5 Learning in DeepProbLog
We now introduce our approach to learn the parameters in DeepProbLog programs. The parameters include the learnable parameters of the neural network (which we will call neural parameters from now on) and the learnable parameters in the logic program (which we will refer to as probabilistic parameters). We use the learning from entailment setting (lfe)
Definition 5
Learning from entailment Given a DeepProbLog program with parameters , a set of pairs with a query and its desired success probability, and a loss function , compute:
In most of the experiments, unless mentioned otherwise, we only use positive examples for training (i.e., with desired success probability ). The model then needs to adjust the weights to maximize query probabilities for all training examples. This can be expressed by minimizing the average negative log likelihood of the query, whereby Definition 5 reduces to:
The presented method however works for other choices in the loss function. For example, in experiment T9 (Section 6.3) the mean squared error (MSE) is used.
5.1 Gradient descent in ProbLog
In contrast to the earlier approach for ProbLog parameter learning in this setting by gutmann2008parameter, we use gradient descent rather than EM. This allows for seamless integration with neural network training. The key insight here is that we can use the same AC that ProbLog uses for inference for gradient computations as well. We rely on the automatic differentiation capabilities already available in ProbLog to derive these gradients. More specifically, to compute the gradient with respect to the probabilistic logic program part, we rely on Algebraic ProbLog (aProbLog (kimmig2011algebraic)), a generalization of the ProbLog language and inference to arbitrary commutative semirings, including the gradient semiring (eisner2002parameter). In the following, we provide the necessary background on aProbLog, discuss how to use it to compute gradients with respect to ProbLog parameters and extend the approach to DeepProbLog.
aProbLog and the gradient semiring
ProbLog annotates each probabilistic fact with the probability that is true, which implicitly also defines the probability that its negation is true. It then uses the probability semiring with regular addition and multiplication as operators to compute the probability of a query on the AC constructed for this query, cf. Figure 0(d). The probability semiring is defined as follows:
(3)  
(4)  
(5)  
(6) 
And the accompanying labeling function as:
(7)  
(8) 
This idea is generalized in aProbLog to compute such values based on arbitrary commutative semirings. Instead of probability labels on facts, aProbLog uses a labeling function that explicitly associates values from the chosen semiring with both facts and their negations, and combines these using semiring addition and multiplication on the AC. We use the gradient semiring, whose elements are tuples , where is a probability (as in ProbLog), and is the partial derivative of that probability with respect to a parameter , that is, the probability of a probabilistic fact with learnable probability, written as . This is easily extended to a vector of parameters , the concatenation of all probabilistic parameters in the ground program, as it is easier and faster to process all gradients in one vector. Semiring addition , multiplication and the neutral elements with respect to these operations are defined as follows:
(9)  
(10)  
(11)  
(12) 
Note that the first element of the tuple mimics ProbLog’s probability computation, whereas the second simply computes gradients of these probabilities using derivative rules.
Gradient descent with aProbLog
To use the gradient semiring for gradient descent parameter learning in ProbLog, we first transform the ProbLog program into an aProbLog program by extending the label of each probabilistic fact to include the probability as well as the gradient vector of with respect to the probabilities of all probabilistic facts and ADs in the program, i.e.,
(13)  
(14)  
(15) 
where the vector has a in the th position and in all others. For fixed probabilities, the gradient does not depend on any parameters and thus is 0. Note that after each update step, the probabilistic parameters are clipped to the range, and the parameters of an AD are renormalized to ensure that they sum to one. For the other cases, we use the semiring labels as introduced above.
Example 4
Assume we want to learn the probabilities of earthquake
and burglary
in the example of Figure 1, while keeping those of the other facts fixed. Figure 3 shows the evaluation of the same AC as in Figure 0(d), but with the gradient semiring. The nodes in the AC now also contain the gradient (the second element of the tuple). The result on the top node shows that the partial derivative of the query is and w.r.t. the earthquake and burglary parameters respectively.
5.2 Gradient descent for DeepProbLog
Just as the only difference between inference in ProbLog and DeepProbLog is the evaluation of the nADs, the only difference between gradient descent in ProbLog and DeepProbLog is optimizing the neural parameters alongside the probabilistic parameters. As mentioned in the previous section, the probabilistic parameters in the logic program can be optimized by using the gradient semiring, which allows us to calculate . This gradient is then used to perform the update by using gradient descent. Note that since the outputs of the neural network are used as probabilities in the logic program and can be learned, we can view them as a kind of abstract parameters. However, although we can derive a gradient for these abstract parameters, we cannot optimize them directly, as the logic is unaware of the neural parameters that determine the value of these abstract parameters. Recall from Equation (1) that the gradient of the internal (neural) parameters in standard supervised learning can be derived using the chain rule in backpropagation. Below, we show how we can derive the gradient for these neural parameters of the loss applied to (Definition 5), rather than a loss function defined directly on the output of the neural network.
Specifically, consider the case of a single neural annotated disjunction, with probabilities (i.e., the aforementioned abstract parameters), calculated by evaluating a neural network with softmax output. The predicted probability that the query holds true, based on the current values of the neural and probabilistic parameters, is written . While training, true examples should yield a predicted query probability close to the expected query probability, which is expressed by means of a loss function as introduced in Definition 5.
Application of the chain rule leads to
where the derivative of the loss with respect to any trainable parameter in the neural network is decomposed into the partial derivative of the loss with respect to the predicted output , the latter’s derivative with respect to each component of the annotated disjunction as obtained with the gradient semiring, and finally , the derivative of the neural network’s output components with respect to the considered parameter. The latter is obtained by the standard application of the chain rule in the neural network. The backpropagation procedure in the neural network can thus be started by providing , to systematically obtain the loss gradients for all neural parameters.
Extending this approach to the situation of multiple neural predicates is straightforward. If the same neural network is used for different neural predicates (e.g. in Example 3), the final derivative is obtained by summing over the contributions of each neural predicate.
Then, standard gradientbased optimizers (e.g. SGD, Adam, …) are used to update the parameters of the network. During gradient computation with aProbLog, the probabilities of neural ADs are kept constant. Furthermore, updates on neural ADs come from the neural network part of the model, where the use of a softmax output layer ensures a normalized distribution, hence not requiring the additional normalization as for nonneural ADs.
To extend the gradient semiring to DeepProbLog programs, we define it for nADs and neural facts. The label for the nAD is defined as:
(16) 
Where is the jth domain element, is the jth element of the output of the neural network evaluated on input . The label for a neural fact is defined as:
(17) 
where is the output of the neural network evaluated on input . Since the first element of the tuple for nADs and neural facts is the evaluation of the neural networks as in Section 4.2, this change remains semantically equivalent.
Example 5
To demonstrate the learning pipeline (Figure 5), we will apply it on the MNIST addition example show in Section 4.2 with a small extension: some of the labels have been corrupted and are picked randomly from a uniform distribution over . The goal is to also learn the fraction of noisy examples. The DeepProbLog program is given in Figure 3(a). Grounding on the query results in the ground DeepProbLog program shown in Figure 3(b). The arithmetic circuit corresponding to the ground program is shown in Figure 3(c). As can be seen, the neural networks already have a confident prediction for both images (being 0 and 1 respectively). The top right shows how the different partial derivatives that are calculated: one w.r.t. to the noisy parameter, ten for the evaluation of the neural network on input a and ten for the evaluation on input b.
6 Experimental Evaluation
Number of training examples  

Model  30 000  3 000  300 
Baseline  
DeepProbLog 
We perform three sets of experiments to demonstrate that DeepProbLog supports {enumerate*}[label=()]
logical reasoning and deep learning;
program induction; and
probabilistic inference and combined probabilistic and deep learning.
We provide implementation details at the end of this section and list all programs in A.
6.1 Logical reasoning and deep learning
To show that DeepProbLog supports both logical reasoning and deep learning, we extend the classic learning task on the MNIST dataset (mnist98) to four more complex problems that require reasoning:
 T1:

Instead of using labeled single digits, we train on pairs of images, labeled with the sum of the individual labels. This is the same as Example 3. The DeepProbLog program consists of the clauseand a neural AD for the digit/2 predicate, which classifies an MNIST image. We compare to a CNN baseline ^{5}^{5}5We’d like to thank Paolo Frasconi for the interesting discussion and idea for a new baseline. classifying the two images into the 19 possible sums.
Results
Figure 6 shows the learning curves for the baseline (orange) and DeepProbLog (blue) on the singledigit addition. We evaluated on 3 levels of data availability: 30 000 examples, 3 000 and 300 examples. As can be seen in the figures, DeepProbLog converges faster and achieves a higher accuracy than the baseline. In the case for N = 30 000 (Figure 5(a)), the difference between the baseline and DeepProbLog is significant, but not immense. However, for N = 3000 and especially N = 300, the difference becomes more apparent.
The reason behind this disparity is that the baseline needs to learn making a decision for the combined input digits (and there are a 100 different sums possible), whereas the DeepProbLog’s neural predicate only needs to recognize individual digits (with only 10 possibilities). Table 1 shows the average accuracy on the test set for the different models for different training set sizes.Figure 7: MNIST MultiDigit Addition (T2). The graphs show the accuracy on the validation set during training for different training set sizes. Number of training examples Model 15 000 1 500 150 T1 (30 000) Baseline – DeepProbLog Table 2: The accuracy on the test set for T2.  T2:

The input consists of two lists of images, each element being a digit. Each list represents a multidigit number. The label is the sum of the two numbers. The neural predicate remains the same. Learning the new predicate requires only a small change in the logic program. Because the CNN baseline cannot handle numbers of varying size, we fixed the size of the input to twodigit numbers.Results
First, we perform an experiment where we take the neural network trained in T1 and use it in this model without any further training. Evaluating it on the same test set, we achieve an accuracy that is not significantly different from training on the full dataset of T2. This demonstrates that the approach used in DeepProbLog causes it to generalize well beyond training data. Figure 7 shows the learning curves for the baseline (orange) and DeepProbLog (blue) on the multidigit addition. DeepProbLog achieves a somewhat lower accuracy compared to the single digit problem due to the compounding effect of the classification error on the individual digits, but the model generalizes well. The baseline fails to learn from few examples (150 and 1 500). It is able to learn with 15 000 examples, but converges very slowly. Table 2 shows the average accuracy on the test set for the different models for different training set sizes.  T3:

The input consists of 3 MNIST images such that the last is the sum of the first two. This task demonstrates potential pitfalls of only providing supervision on the logic level. Namely, without any regularization, the neural network quickly learns to predict for all digits, i.e., the model collapses to always predicting , as it is a valid logical solution. To avoid this, we add a regularisation term based on entropy maximization (Equation 18, Section 6.4). The intuition behind this regularisation term is that it penalizes mode collapse by requiring the entropy of the average output distribution per batch to be high. As such, this term encourages exploration, but is only necessary to start the training of the neural networks. If they are sufficiently trained, this term can be dropped. We call this additional loss term infoloss. This additional regularization loss is multiplied by a factor and added to the crossentropy loss. We run the experiment for different values of .Results
Figure 8 shows the accuracy of the neural predicate on classifying single digits for different levels of the regularization parameter. As can be seen, for , the neural predicate converges on the trivial solution. For , the neural predicate sometimes converges on the correct solution, but can also converge on the wrong solution. For , the neural network consistently converges on the correct solution.
 T4:

This experiment is the example shown in Figure 4. It’s the same as T1, but with noise introduced in the labels. Namely, a fraction of the labels is replaced by uniformly selected number between and . We compare three models: the CNN baseline from T1, the DeepProbLog model from T1, and a DeepProbLog model where the noise is explicitly modeled as in Figure 4.Results
Table 3 shows the accuracy on the test set which has no noise. The baseline is not tolerant to noisy labels, quickly dropping in accuracy as the fraction of noisy labels increases. The DeepProbLog model from T1 is more tolerant, but also drops noticeably in accuracy as the fraction of noise goes over . Explicitly modeling the noise makes the model very noise tolerant, even retaining an accuracy of with noisy labels. As shown in the last row, it is also able to learn the fraction of noisy labels in the data. This shows that the model is able to recognize which examples have noisy labels.
Fraction of noise  

0.0  0.2  0.4  0.6  0.8  1.0  
Baseline  
DeepProbLog  
DeepProbLog w/ explicit noise  
Learned fraction of noise 
.
6.2 Program Induction
The second set of problems demonstrates that DeepProbLog can perform program induction. We follow the program sketching (sketching) setting of differentiable Forth () (riedel2017programming), where holes in given programs need to be filled by neural networks trained on inputoutput examples for the entire program. As in their work, we consider three tasks: addition, sorting (reed2015neural) and word algebra problems (WAPs) (roy2016solving).
 T5:

The input consists of two numbers, represented as lists of digits, and a carry. The output is the sum of the numbers and the carry. The program specifies the basic addition algorithm in which we go from right to left over all digits, calculating the sum of two digits and taking the carry over to the next pair. The hole in this program corresponds to calculating the resulting digit (result/4) and carry (carry/4), given two digits and the previous carry.Results
The results are shown in Table 4. Similarly to , DeepProbLog achieves 100% on all training sizes.  T6:

The input consists of a list of numbers, and the output is the sorted list. The program implements bubble sort, but leaves open what to do on each step in a bubble (i.e. whether to swap or not: swap/2).Results
The results are shown in Table 5. Similarly to , DeepProbLog achieves 100% on training sizes 2 and 3. However, whereas fails to converge on training sizes larger than 3, DeepProbLog stills achieves 100% accuracy. As riedel2017programming mention, the failure of is due to computational issues arising from the long program trace resulting from sorting long lists. DeepProbLog does not suffer from these issues. As shown in Table 6, DeepProbLog runs faster and scales better with increasing training length.  T7:

wap(‘Robert has 12 books . … How many does he have now ?’,12,3,1,10)
The input to the WAPs consists of a natural language sentence describing a simple mathematical problem. These WAPs always contain three numbers, which are extracted from the string and are given as part of the input. The output is the solution to the question. Every WAP can be solved by chaining the following 4 steps: permuting the three numbers (permute/2), applying an operation on the first two numbers (addition, subtraction or product operation_1/2), potentially swapping the intermediate result and the last digit (swap/2), and performing a last operation (operation_2/2). The hole in the program is in deciding which of the alternatives should happen on each step.Results
DeepProbLog reaches an accuracy of up to 96.5%, similar to the results for reported by riedel2017programming (96%).
Training length  

Test length  2  4  8  
(riedel2017programming)  8  100.0  100.0  100.0 
64  100.0  100.0  100.0  
DeepProbLog  8  100.0  100.0  100.0 
64  100.0  100.0  100.0 
Training length  

Test length  2  3  4  5  6  
(riedel2017programming)  8  100.0  100.0  49.22  –  – 
64  100.0  100.0  20.65  –  –  
DeepProbLog  8  100.0  100.0  100.0  100.0  100.0 
64  100.0  100.0  100.0  100.0  100.0 
Training length  

2  3  4  5  6  
on GPU  42 s  160 s  –  –  – 
on CPU  61 s  390 s  –  –  – 
DeepProbLog  11 s  14 s  32 s  114 s  245 s 
6.3 Probabilistic programming and deep learning
In this section we introduce two final experiments that show the intricacies involved in combining probabilistic logic programming and deep learning.
 T8:

Coin classification and comparison
In this experiment we train two neural networks using distant supervision. The input consists of a synthetic image containing two coins (an example is shown in Figure 9. They are either heads or tails. The image is labeled either with same or different. We train a neural network for each coin to predict either heads or tails. Solving this task requires solving two problems. On the one hand, the neural networks have to learn to recognize and separate the two different coins; on the other hand, they also have to each classify a different coin as heads or tails. The first question we ask is whether the neural networks can recover the latent structure imposed by the logic program. We expect the two neural networks to agree on which side of the coin is heads and which is tails, however, this might be the inverse of what is generally considered heads and tails. Furthermore, we expect the two neural networks to each pick one coin to label, but which network classifies which coin will vary between runs. As such, there are four possible solutions that the neural networks can converge on. The second question we ask is how many additionally labeled examples (with both the label for same/different and heads/tails of one of the coins given) we need for the neural network to recover the desired latent representation.Results
We ran each experiment 100 times. The fraction of runs that converged on either no solution, the expected solution or a logically equivalent solution is shown in Table 7. We see that with no additionally labeled examples, DeepProbLog doesn’t converge on a satisfactory solution in about half of all runs. When it does converge on a solution, it converges on the expected solution 25% of the time, and on different solutions 75% of the time, which is conform with our expectations. We can also see that as the number of additionally labeled examples increases, DeepProbLog converges more reliably, and more on the expected solution. Starting with 10 additionally labeled examples, DeepProbLog reliably converges on the desired solution. Beyond 20 additionally labeled examples, we don’t see any further improvements. Labeled examples Not solved Expected solution Other solution 0 56% 11% 33% 5 39% 40% 21% 10 7% 92% 1% 20 4% 96% 0% 50 3% 97% 0% 100 4% 96% 0%  T9:

0.8::poker([Q, Q, A, K],loss).
In this last experiment we demonstrate that DeepProbLog can perform combined probabilistic reasoning, probabilistic learning and deep learning. We do this by playing a simplified poker game: there are two players that are dealt two cards from several decks. There is also one community card. Each player then makes a poker hand (e.g. pair, straight, …) with their two cards and the community card.For simplicity, we only use the jack, queen, king and ace. We also do not consider the suits of the cards.
The input consists of 4 images that show the cards dealt to the two players. Additionally, every example is labeled with the chance that the game is won, lost or ended in a draw, e.g.:
We expect DeepProbLog to:

train the neural network to recognize the four cards

reason probabilistically about the nonobserved card

learn the distribution of the unlabeled community card
To make DeepProbLog converge more reliably, we add some examples with additional supervision. Namely, in 10% of the examples we additionally specify the community card, i.e.
This also showcases one of the strengths of DeepProbLog, namely, it can make use of examples that have different levels of observability. The loss function used in this experiment is the MSE between the predicted and target probabilities.
Results
We ran the experiment 10 times. Out of these 10 runs, 4 didn’t converge on the correct solution. The average values of the learned parameters for the remaining 6 runs are shown Table 8. As can be seen, DeepProbLog is able to correctly learn the probabilistic parameters. In these 6 runs, the neural network also correctly learned to classify all card types, achieving a 100% accuracy. The other runs did not converge because some of the classes were permuted (i.e., queens predicted as aces and vice versa) or multiple classes mapped onto the same one (queens and kings were both predicted as kings).
Distribution Jack Queen King Ace Actual 0.2 0.4 0.15 0.25 Learned Table 8: The results for the Poker experiment (T9). 
6.4 Implementation details
For the implementation we integrate ProbLog2 (dries2015problog2) with PyTorch (paszke2017automatic). All programs are listed in A. In experiments T1T8 we optimize the crossentropy loss between the predicted and target query probabilities, as we found that this works better to learn the probabilistic parameters. In experiment T9 we optimize the MSE between the predicted and target query probabilities. We use Adam (kingma2014adam) optimization for the neural networks, and SGD for the logic parameters. For T9, we add random rotations (max 10 degrees) and shift the colours in the HSV by up to 5% to the images of the cards.
The neural network architectures are summarized in Table 11. Conv(o,k) denotes a convolutional layer with o output channels and kernel size k. Lin(n) denotes a fully connected layer of size n. BiGRU(h) denotes a singlelayer bidirectional GRU with a hidden size h. (layer) means that there are two identical layers in parallel that are concatenated. A layer in bold means it is followed by a ReLU activation function. All neural networks end with a Softmax layer, unless otherwise specified. The hyperparameters used in the experiments are shown in Table 9. The sizes of the datasets used are specified in Table 10.
The regularisation term used in T3 is calculated per network and per batch on the average of the neural network output. It is calculated as
(18) 
where is the ith output of the neural network and is the nary entropy (i.e. entropy using the basen logarithm).
6.5 Computation time
Due to the nature of the exact inference used in DeepProbLog, which it inherits from aProbLog kimmig2011algebraic, the grounding and compilation steps can become expensive as the problem size grows. For example, in T1, which is the smallest problem we consider, grounding and compilation takes on average 0.01 seconds, while evaluating the NN and AC takes on average 0.002 seconds per example. In T9, which has the largest logic program out of all experiments, grounding and compilation takes on average 1.3 seconds, while NN and AC evaluation takes on average 0.05 seconds per example.
It is important to note that when we evaluate an example a second time, the structure of the AC, which is determined by the grounding and compilation, remains the same. Only the learned probabilities in the nAD change. We make use of this to improve the performance by caching the arithmetic circuits so that we only have to perform the (potentially expensive) grounding and compilation steps once. During evaluation, we only reevaluate the neural networks and evaluate the AC with the updated probabilities.
Note that this optimization can also be applied to, for example, the queries and , as both have the same structure. To do this, we introduce placeholder constants and change both queries to the single query , which reduces all queries in T1 to 19 unique queries, one for each different label. During the evaluation of the neural networks, we replace the constants with the correct input and use the resulting probabilities in the cached ACs. We apply these optimizations to all experiments.
Task  Batch size  Learning rate  Parameter learning rate  Infoloss 

T1T3  2  1e3  
T4  2  1e3  1e3  2, 4, 8 
T5  50  0.02  
T6  16  1.0  
T7  100  0.005  
T8  5  1e4  0.25  
T9  50  1e4  1e3  0.5 
Task  Training set  Validation Set  Test set 

T1  29 500, 3 000, 300  500  5 000 
T2  14750, 1 500, 150  250  2 500 
T3  16 000  2 000  3 000 
T4  29 500  500  5 000 
T5  512  256  1 024 
T6  256  32  32 
T7  300  100  200 
T8  100  –  20 
T9  500  –  25 
Task  Network  Architecture 
T1T4  digit/2  MNISTConv, Lin(120), Lin(84), Lin(10) 
T1,T3  baseline  MNISTConv, Lin(120), Lin(84, Lin(19) 
T2  baseline  (MNISTConv, Lin(100)),Lin(128, Lin(199) 
T5  result/4  Lin(50), TanH, Lin(10) 
carry/4  Lin(10), TanH, Lin(2)  
T6  swap/3  Lin(20), Lin(10) 
T7  RNN  Embedding(256), BiGRU(512), Dropout(0.5)* 
perm/2  Lin(6)  
op1/2  Lin(4)  
swap/2  Lin(2)  
op2/2  Lin(4)  
T8  coin1/2  AlexNetConv, Lin(100), Lin(2) 
coin2/2  AlexNetConv, Lin(100), Lin(2)  
T9  rank/2  AlexNetConv, Lin(100), Lin(4) 
MNISTConv: Conv(6,5), MP(2,2), Conv(16,5), MP(2,2)*
AlexNetConv: Conv(64, 11, 2,2), MP(3,2), Conv(192, 5, 2), MP(3,2),
Conv(384, 3, 1), Conv(256, 3, 1), Conv(256, 3, 1), MP(3,2)*
* Does not end with a Softmax layer.
7 Related Work
Most of the work on combining neural networks and logical reasoning comes from the neurosymbolic reasoning literature (garcez2012neural; hammer2007perspectives). These approaches typically focus on approximating logical reasoning with neural networks by encoding logical terms in Euclidean space. However, they neither support probabilistic reasoning nor perception, and are often limited to nonrecursive and acyclic logic programs (Holldobler1999). DeepProbLog takes a different approach and integrates neural networks into a probabilistic logic framework, retaining the full power of both logical and probabilistic reasoning and deep learning.
At the same time, DeepProblog also integrates probability in neurosymbolic computation. Although this may appear as a complication, our work actually shows that it can greatly simplify the integration of neural networks with logic. The reason for this is that the probabilistic framework provides a clear optimisation criterion, namely the probability of the training examples. Realvalued probabilistic quantities are also wellsuited for gradientbased training procedures, as opposed to discrete logic quantities.
The prominent recent lines of related work focus on three main branches: pushing the logic as regularisation, templating neural networks, and neural program induction.
7.1 Logic as regularisation
The main idea behind this line of research is that logic is included as a regularizer during the optimization of the neural network, or the learning of the embeddings. The goal is to encode the logic into the weights so that after training, when the logic is no longer explicitly present, the evaluation still shows the characteristics of the logic. rocktaschel2015injecting; Demeester2016; minervini2017adversarial; diligenti2017semantic; donadello2017logic; xu2018semantic all center around including logical background knowledge as a regularizer during training. rocktaschel2015injecting inject background knowledge into a matrix factorization model for relation extraction, by adding differentiable loss terms for propositionalized firstorder rules. Demeester2016 propose a more efficient alternative by inducing order relations in embedding space, effectively leading to a lifted application of the rules. This is further generalized by minervini2017adversarial, who investigate injecting rules by minimizing an inconsistency loss on adversariallygenerated examples. diligenti2017semantic use FOL to specify constraints on the output of the neural network. They use fuzzy logic to create a differentiable way of measuring how much the output of the neural networks violates these constraints. This is then added as an additional loss term that acts as a regularizer. More recent work by xu2018semantic introduces a similar method that uses probabilistic logic instead of fuzzy logic, and is thus more similar to DeepProbLog. They also compile the formulas to an SDD for efficiency.
However, whereas DeepProbLog is based on (probabilistic) logic programming, these methods use first order logic instead. This is reminiscent to the difference between ProbLog and Markov Logic richardson2006markov or PSL bach2015hinge. donadello2017logic, though at first sight related to diligenti2017semantic, work slightly differently. They learn functions that map numerical properties of logical constants onto truth values, which are then combined using fuzzy logic.
7.2 Templating neural networks
This line of work uses the logic as a template for constructing the architecture of neural networks. This is reminiscent of the knoweldge base construction approach of statistical relational artificial intelligence DeRaedt16.
rocktaschel2017end introduce a differentiable framework for theorem proving. They reimplemented Prolog’s theorem proving procedure in a differentiable manner and enhanced it with learning subsymbolic representation of the existing symbols, which are used to handle noise in data. Whereas rocktaschel2017end use logic only to construct a neural network and focus on learning subsymbolic representations, DeepProblog focuses on tight interactions between the two and parameter learning for both the neural and the logic components. In this way, DeepProbLog retains the best abilities of both worlds. Recently, weber2019nlprolog extend the notion of soft unification towards structured textual knowledge, i.e., unification can be performed between sentences, not only symbols. In contrast to rocktaschel2017end, weber2019nlprolog retain the full ability of logical reasoning, and as such is closer to DeepProblog, but it is specialised for NLP tasks.
cohen2018tensorlog introduce a framework to compile a tractable subset of logic programs into differentiable functions and to execute it with neural networks. It provides an alternative probabilistic logic but it has a different and less developed semantics. Furthermore, to the best of our knowledge it has not been applied to the kind of tasks tackled in the present paper. An idea similar in spirit to ours is that of andreas2016neural, who introduce a neural network for visual question answering composed out of smaller modules responsible for individual tasks, such as object detection. Whereas the composition of modules is determined by the linguistic structure of the questions, DeepProbLog uses probabilistic logic programs to connect the neural networks.
7.3 Neural program induction.
The third line of work has focused on learning programs from data by combining neural and symbolic approaches.
Neural execution
The first category captures a program behaviour with neural networks and therefore focuses on program execution. The approach most similar to ours is that of riedel2017programming, where neural networks are used to fill in holes in a partially defined Forth program. DeepProblog differs in that it uses ProbLog as the host language which results in native support for both logical and probabilistic reasoning, while differentiable Forth uses a procedural language. Differentiable Forth has been applied to tasks T57, but it is unclear whether it could be applied to the remaining ones. Finally, evans2018learning introduce a differentiable framework for rule induction, that does not focus on the integration of the two approaches like DeepProblog.
Neurally guided search
The second line of research enhances the search procedures of the symbolic program induction techniques by incorporating neural components in the search itself. The key principle these techniques employ is to perform the search over programs in a systematic symbolic way, but guide the search with a heuristic learned by a deep neural network. For instance, kalyan2018 train a neural network to predict the scores of branches during the branchandbound search procedure, zhang2018neural train a neural network to choose which candidate program to expand next while exploiting the constraints on the inputoutput examples, while Ellis2018 use a neural network to efficiently search over a welldesigned DSL.
Neural program construction
The final category involves techniques that decompose a problem into independent parts that can be individually solved by either neural or symbolic components and synchronize the individual components to solve the main problem. For instance, Yi2018; mao2018 develop a neurosymbolic approach towards visual question answering by using a neural network to generate a program computing the answer to the question and executing the program symbolically. Ellis2018a generate a LaTeX code from a handdrawn sketch by using a neural network to recognise basic shapes within a sketch and symbolically inducing the program describing the sketch. In contrast, dong2018neural induce programs in a purely neural way and demonstrate favourable generalization; however, they do not induce symbolic programs, but rather express a program through a neural network.
7.4 Symbolic deep learning
The success of neural deep learning has inspired has inspired several works introducing symbolic deep learning methods which, instead of representing the logical aspects in a vector space, retain the logical data representation in the latent representation. These include the symbolic versions of deep neural networks: LRNN treat symbolic rules expressed in firstorder logic as a template for constructing a neural network, while KazemiAAAI18 compose a relational neural network by adding hidden layers to relational logistic regression Kazemi:2014:RLR:2908339.2908347. Another research direction focuses on taskagnostic discovery of relational (symbolic) latent representations by exploiting approximate symmetries DumancicIjcai2017, a symbolic extension of the autoencoding principle DumancicIjcai2019, or selfplay CropperIjcai2019.
8 Conclusion
We introduced DeepProbLog, a framework where neural networks and probabilistic logic programming are integrated in a way that exploits the full expressiveness and strengths of both worlds and can be trained endtoend based on examples. This was accomplished by extending an existing probabilistic logic programming language, ProbLog, with neural predicates. Learning is performed by using aProbLog to calculate the gradient of the loss which is then used in standard gradientdescent based methods to optimize parameters in both the probabilistic logic program and the neural networks. We evaluated our framework on experiments that demonstrate its capabilities in combined symbolic and subsymbolic reasoning, program induction, and probabilistic logic programming.
Although we have shown promising results, DeepProbLog is currently using only exact inference. As exact inference does not always scale well and can be prohibitively expensive for large problems, the DeepProbLog implementation cannot yet be applied to problems such as KB completion. Future work will be concerned with incorporating approximate inference algorithms to speedup the grounding and compilation process.
Acknowledgements
RM is a SB PhD fellow at FWO (1S61718N). SD is supported by the Research Fund KU Leuven (GOA/13/010) and Research Foundation  Flanders (G079416N). LDR is partially supported by the European Research Council Advanced Grant project SYNTH (ERCAdG694980).
References
Appendix A DeepProbLog Programs
In Listing 1, digit/2 is the neural predicate that classifies an MNIST image into the integers 0 to 9. The addition/3 predicate’s first two arguments are MNIST digits, and the last is the sum. It classifies both images using digit/2 and calculates the sum of the two results.
In Listing 2, the only difference with Listing 1 is that the multi_addition/3 predicate now uses the number/2 predicate instead of the digit/2 predicate. The number/3 predicate’s first argument is a list of MNIST images. It uses the digit/2 neural predicate on each image in the list, summing and multiplying by ten to calculate the number represented by the list of images (e.g. number([],38)).
In Listing 3, the only difference with Listing 1 is that all 3 inputs X,Y,Z are images. As such, the digit/2 predicate is also used on the third input. The sum is also redefined as Z2 is X2+Y2.
In Listing 4, an additional probabilistic fact (noisy/2) is added that encodes the chance of an example being noisy. The addition/3 predicate is split into two cases: when the noisy is true or when noisy is false. The latter is the same as in Listing 1. If noisy is true, Z is considered to be drawn from the uniform distribution (uniform/3).
In Listing 5, there are two neural predicates: result/4 and carry/4. These are used in the slot/4 predicate that corresponds to the slot in the Forth program. The first three arguments are the two digits and the previous carry to be summed. The next two arguments are the new carry and the new resulting digit. The add/5 predicate’s arguments are: the two list of input digits, the input carry, the resulting carry and the resulting sum. It recursively calls itself to loop over both lists, calling the slot/5 predicate on each position, using the carry from the previous step.
In Listing 6, there’s a single neural predicate: swap/3. Its first two arguments are the numbers that are compared, the last argument is an indicator whether to swap or not. The bubble/3 predicate performs a single step of bubble sort on its first argument using the hole/4 predicate. The second argument is the resulting list after the bubble step, but without its last element, which is the third argument. The bubblesort/3 predicate uses the bubble/3 predicate, and recursively calls itself on the remaining list, adding the last element on each step to the front of the sorted list.
In Listing 7, there are four neural predicates: net1/2 to net4/2. Their first argument is the input question, and the second argument are indicator variables for the choice of respectively: one of six permutations, one of 4 operations, swapping and one of 4 operations. These are implemented in the permute/7, swap/5 and operator/4 predicates. The wap/5 predicate then sequences these steps to calculate the result.
In Listing 8, there are two neural predicates: coin1/2 and coin2/2. Their input is the image of the two coins (e.g. Figure 9). The output is heads or tails. The coins/2 classifies both coins using these two predicates and then performs the comparison of the classes with the compare/3 predicate.
In Listing 9, there’s a single neural predicate rank/2 that takes as input the image of a card and classifies it as either a jack, queen, king or ace. There’s also an AD with learnable parameters that represents the distribution of the unseen community card (house_rank/1). The hand/2 predicate’s first argument is a list of 3 cards. It unifies the output with any of the valid hands that these cards contain. The valid hands are: high card, pair (two cards have the same rank), three of a kind (three cards have the same rank), low straight (jack, queen king) and high straight(queen, king, ace). Each hand is assigned a rank with the hand_rank/2 predicate. The best_hand_rank/2 predicate takes as input a list of cards, and unifies the second argument with the highest hand rank that is possible with the three given cards. The outcome/3 predicate determines the outcome by comparing the two ranks of the best hand. The game/3 predicate’s first argument is a list of the 4 input images. It’s second input is the labeled community card. It classifies the cards using the neural predicates, determines the best rank, and then unifies the last argument with the outcome. The game/2 determines the community card from the learned distribution house_rank/1, and then determines the outcome using the game/3 predicate. The member/2 and select/3 predicates are predicates from the lists library. member/2 is true if it’s second argument is a list and the first argument appears in that list. select/3 is true if it’s second argument is a list and the first argument appears in that list. It also unifies the last argument with the list that is the same as it’s second argument, but with the first argument removed.