Safety Properties of Inductive Logic Programming Gavin Leech, * 1 Nandi Schoots, * 2 Joar Skalse 3 1 University of Bristol 2 King’s College London and Imperial College London 3 University of Oxford * Equal contribution g.leech@bristol.ac.uk Abstract be a natural fit for the assurance side of safety: often, not just the output model, but also the learning process takes place This paper investigates the safety properties of inductive logic at a relatively high level (that is, at the level of symbolic programming (ILP), particularly as compared to deep learn- ing systems. We consider the following properties: ease of inference). Similarly, ILP plausibly satisfies multiple impor- model specification; robustness to input change; control over tant specification and robustness properties. We assess ILP inductive bias; verification of specifications; post-hoc model on: Specification properties (ease of model specification and editing; and interpretability. We find that ILP could satisfy value loading; ease of adjusting the learned model to satisfy many of these properties in its home domains. Lastly, we pro- specifications; and control over inductive bias); Robustness pose a hybrid system using ILP as a preprocessor to generate properties (robustness to input change and to post-training specifications for other ML systems. model edits); Assurance properties (interpretability and ex- plainability; verification of specifications; and control over the inductive bias). Introduction Many safety properties await formalisation, preventing Symbolic approaches to AI are sometimes considered safer quantitative comparisons. Where a formal metric is lacking, than neural approaches (Condry 2016; Anderson et al. we qualitatively compare ILP to deep learning (DL). 2020). We investigate this by analysing how one symbolic In the following we refer to ‘ILP’ as if it was mono- approach, inductive logic programming (ILP), fares on spe- lithic, but ILP systems differ widely in search strategy, ex- cific safety properties. actness, completeness, target logic (e.g. Prolog, Datalog, ILP is a declarative subfield of ML for learning from ex- ASP), noise-handling, ability to invent predicates, and the amples and encoded “background knowledge” (predicates order of the output theory (Boytcheva 2002). This diversity and constraints), using logic programs to represent both limits the general statements we can make, but some remain. these inputs and the output model (Muggleton 1999). (We use ‘output model’ and ‘ILP hypothesis’ interchangeably.) Safety properties of ILP We find ILP to have potential to satisfy an array of safety properties. To arrive at this, we survey existing work in ILP Model Specification and deep learning in light of the safety properties defined The specification of an ML system serves to define its pur- in the framework of Ortega and Maini (2018). We also for- pose. When this purpose is successfully implemented in hard malise robustness to input change and model editing. We constraints, we may obtain guarantees about the system’s be- suggest a hybrid system, in which ILP is used as a prepro- haviour. A defining feature of ILP systems is user-specified cessing step to generate specifications for other ML systems. background knowledge. This provides a natural way to im- To our knowledge, this is the first analysis of ILP’s safety pose specifications on ILP systems. An ILP problem specifi- potential, and of ILP’s differences from deep learning. Re- cation is a set of positive examples, negative examples, and lated work includes Cropper, Dumančić, and Muggleton background knowledge. (2020)’s recent survey of ILP, the interpretability work of Consider two important properties of classical ILP. Given Muggleton et al. (2018b), and Powell and Thévenod-Fosse a background B, an output model M , and positive examples (2002)’s study of rule-based safety-critical systems. E + , the model M is Consider a machine learning system ‘safe’ when the sys- • weakly consistent if: B ∧ M 6|= False; and tem’s goals are specified correctly, when it acts robustly ac- • strongly consistent if: B ∧ M ∧ E + 6|= False. cording to those goals, when we are assured about these two Weak consistency forbids the generation of models that properties (Ortega and Maini 2018), such that the risk of contradict any clause in B (Muggleton 1999). In general, harm from deploying the system is greatly reduced. ILP may ILP algorithms must satisfy weak consistency (Muggleton Copyright c 2021 for this paper by its authors. Use permitted un- 1999), though probabilistic systems allow its violation; see der Creative Commons License Attribution 4.0 International (CC below. Hence, to guarantee that the learned model M sat- BY 4.0). isfies some specification s, all we need to do is encode s in first-order logic (FOL) and add it to the background B. How- tive parameters, it is not practical to use ordinary constrained ever, there are still some specification challenges for ILP. optimisation methods to impose them (Márquez-Neila, Salz- Not all systems respect strong consistency. Many mod- mann, and Fua 2017). ern implementations of ILP are designed to handle noise in the example set (Srinivasan 2006; Muggleton et al. 2018a). Robustness to Input Change For specifications encoded in the example set, noise han- Robustness concerns smooth output change: If we change dling means that the system is only nudged in the direction the input slightly, will the output (of the learning algorithm of the specification. Furthermore, probabilistic ILP systems or of the learned model) change only slightly? To formalize can specify the background as probabilistic facts (De Raedt this, we define similarity of inputs and output hypotheses. et al. 2015). This means that even weak consistency can be violated. As such, these systems may not offer specification In DL input datasets, the problem description is usually guarantees. very correlated with the semantics of the problem. For ex- ample, Gaussian noise usually does not affect the seman- Incompleteness. Even though a model satisfying our spec- tics of the problem. DL models are often insensitive to small ification exists, an incomplete ILP algorithm might not find changes in the description of the input. However, adversar- it. Some leading implementations of ILP are incomplete, i.e. ial changes induce large changes in output, despite the input a solution may exist even though the system does not find changes being trivial to the human eye (Szegedy et al. 2014). one (Cropper and Tourret 2018) For Horn clauses (a typical form in ILP output hypothe- Specifications may be hard to encode as FOL formulae. In ses), one distance measure is the ‘rewrite distance’ (the min- computer vision, a long tradition of manually encoding vi- imum number of syntactic edits that transform one clause sual concepts was rapidly outperformed by learned represen- into another) (Edelmann and Kunčak 2019). For our pur- tations (Goodfellow, Bengio, and Courville 2016): it proved poses, this is inappropriate, since it neglects the semantic possible to learn these improved concepts, but intractable to distances we are targeting: a negation of the whole clause hand-code them. Insofar as ILP backgrounds must at present would count as a rewrite distance of 1, despite being maxi- be manually encoded (as opposed to learned via predicate mally semantically different. invention), we infer that some specifications are not practi- cally possible to impose on ILP. Definition 1 (Similarity of Datasets) Given two datasets Human values are hard to encode as FOL formulae. A D1 , D2 , let H1 and H2 be the sets of hypotheses compat- particularly interesting kind of specification are those that ible with D1 and D2 respectively. Let the weight of a set of concern norms or values, i.e. specifications that aim to en- hypotheses H be defined as a weighed sum of the hypothe- sure that the output respects ethical considerations. There is ses in H, where more complex hypotheses are given lower precedent for formalizing norms and moral obligations us- weight (so that hypothesis h has weight 0.52c(h) , where c(h) ing logic – deontic logic is an area of philosophical logic is the complexity of h). We then say that D1 and D2 are that aims to formalise and deduce moral claims (McNamara similar if H1 ∩ H2 has a large weight. 2019). This has been used to partially formalise some ethi- cal frameworks (Kroy 1976; Peterson 2014). However, en- Definition 2 (Similarity of Hypotheses) We say that two coding general normative requirements in formal logic is an hypotheses h1 and h2 are similar if the probability that they open problem. Further, we do not have a complete articu- will agree on an instance x sampled from the underlying lation of all such requirements in any formalism. It seems data distribution is high. unlikely that in the near future we will obtain a complete Definition 3 (Robustness to Input Change) Let L: D → encoding, owing to deep inconsistencies across people and M be a learning algorithm. We say that L is robust to input the contextual nature of value (Yudkowsky 2011). Further- change if it is the case that L(D1 ) and L(D2 ) are similar more, it may be impossible to learn a representation of these whenever D1 and D2 are similar. More specifically, we say preferences, in the absence of a strong model of human error that L has robustness parameters rD , rM if: for any D1 and (Armstrong and Mindermann 2018). D2 such that they have similarity rD or higher, the similarity Model specification in DL. Methods exist for limited between L(D1 ) and L(D2 ) is at least rM . model specification in DL (Platt and Barr 1988), many of which focus on specific domains (Kashinath, Marcus et al. We note that, for this notion of similarity between 2019; Zhang et al. 2020). However, if we interpret a spec- datasets, the distance between two ILP problems may be ification as a hard constraint on outputs, then most current very large even if their descriptions are almost the same. For DL methods do not allow specification. Instead they im- example, adding a negation somewhere in the description of pose soft constraints, modifying the loss to discourage out- D1 may completely change its distance to D2 . of-specification behaviour. Imposing hard constraints in DL ILP is robust to syntactic input change. ILP is largely in- amounts to imposing a linear set of constraints on the output variant to how the input problem is represented (in the sense of the model. Soft constraints in the form of subtle alter- of symbol renaming or syntactic substitutions, which do not ations to the loss function or learning algorithm are harder affect the semantics). Two semantically equivalent problems to specify than e.g. a linear set of hard constraints (Pathak, have identical sets of compatible output hypotheses. Krähenbühl, and Darrell 2015). Soft constraints are perva- Examples of trivial syntactic changes to a problem in- sive due to the computational expense of hard constraints in clude: renaming atoms or predicates; substituting a ground neural networks: since networks can have millions of adap- term for a variable; or substituting in a different variable. An ILP problem statement is parsed as an ordered set of log- well with small amounts of data, regardless of its induc- ical sentences, and substitutions within these sentences do tive bias (Vapnik and Chervonenkis 2015). However, with not affect the semantics of the individual examples. Absent a more expressive learning algorithm (such as DL or ILP) complicating implementation details, they thus do not affect this is insufficient to yield good generalisation, and hence the semantics of the output. Another syntactic change to a such learning algorithms need a good inductive bias to work problem is adding or removing copies of examples; these well. ILP’s strong bias allows it to perform well on small changes do not have any effect on what hypothesis is output. datasets, even though hypotheses can also be highly expres- Changing the order of examples could (depending on sive (Tausend 1994). the search algorithm) change the chosen output hypothesis. The two main components of inductive bias are Even though the set of consistent output hypotheses does not • Limits to the hypothesis space: Restricting the hypothesis change when the order of examples changes, the hypothesis space, i.e. the set of possible output models; and that comes up first in the search may change. For example, • Guiding the search: The search order for traversing Metagol depends on the order (Cropper and Morel 2020). through the hypothesis space, as well as heuristics to as- This order dependence is a property of some clause-level sess hypotheses. search methods (Srinivasan 2006). Inductive bias in DL. The hypothesis space in DL is Robustness to semantic input change. Naturally, semantic largely determined by the network architecture (Szymanski, changes to the problem can completely change the output McCane, and Albert 2018), which we have control over. For hypothesis. For example, negating a single example can pre- example, convolutional neural networks hard-code the as- clude finding any appropriate hypothesis. sumption that output classes are invariant to shift transfor- Suppose D1 and D2 are two datasets, with corresponding mation (Goodfellow, Bengio, and Courville 2016). Train- hypothesis spaces respectively H1 and H2 . ILP has a fixed time methods like dropout and learning rate decay also reg- order (which depends on the inductive biases) of traversing ularise networks and so add inductive bias (Srivastava et al. the total set of potential hypotheses for a solution. Say ILP 2014). In addition, neural networks have a broad bias to- outputs hypothesis h1 for problem D1 and hypothesis h2 for wards simplicity, although it is unclear how this bias works D2 . Even if H1 6= H2 , h1 and h2 may be the same. When (Zhang et al. 2017; Poggio, Liao, and Banburski 2020). The h1 6= h2 , we would like to assess their similarity. lack of theoretical understanding of DL’s search bias implies Given an output model. If we change one input example, little explicit control over it. then we may be able to check whether this input example Inductive bias in ILP. We can restrict the hypothesis space is consistent with the output model. We may not be able to in many ways. A critical design decision for an ILP system completely visualise the coverage, but may be able to predict is which fragment of FOL represents the examples, back- whether the output model will be different. ground and output model. The classical choice restricts FOL Empirically assessing robustness to input change. Poten- to definite Horn clauses (Muggleton and de Raedt 1994). tially, sampling can inform us about the robustness to input In addition, a strong ILP language bias stems from user- change of ILP and deep learning. An experiment could work supplied constraints on the structure or type of the hypoth- as follows: Generate ILP problems such that we (approxi- esis, e.g. mode declarations, meta-rules, or program tem- mately) know the distance between the datasets. Then run plates (Payani and Fekri 2019). In some sense these are hy- ILP on each problem and store their output hypotheses. We perparameters, as found in any ML system. However, these then select a distance measure and assess the distance be- constraints can be enormously informative, e.g. specifying: tween each of the output hypotheses. This allows us to (ap- which predicates to use in the head or body of the out- proximately) evaluate the robustness to input change of ILP. put model; the quantifier of each argument in the predicate; A similar sampling process can be used for other learning al- which arguments are to be treated as input and output; and gorithms to compare the robustness of different algorithms. the types of these argument (Evans and Grefenstette 2018). User-supplied constraints can pertain to (Muggleton and Control over Inductive Bias de Raedt 1994) among other things The inductive bias of a learning algorithm is the set of (of- • Syntax, e.g. second-order schema or bounded term depth; ten implicit) assumptions used to generalise a finite input • Semantics (on the level of terms), e.g. hard-coding set to a complete output model (Mitchell 1980; Hüllermeier, the types of predicate arguments, or using determinate Fober, and Mernberger 2013). If several hypotheses fit the clauses; and training data, the inductive bias of the learning algorithm • Bounds on the length of the output model. determines which is selected. Correct behaviour is generally Two elementary ways to order an ILP search over the set under-determined by the training data, so selecting a model of possible output models are top-down (‘from general to with the right behaviour demands inductive bias. It is thus specific’) or bottom-up (‘from specific to general’). At each desirable to adapt the training algorithm through fine con- step of ILP learning, we need a way to score multiple com- trol over the inductive bias. peting hypotheses. This can be done via computing the infor- Informally, a learning algorithm has a low Vap- mation gain of the change to the hypothesis (Quinlan 1990) nik–Chervonenkis (VC) dimension if it can only express or through probabilistic scoring (Muggleton and de Raedt simple models. If a learning algorithm has a low VC- 1994). A further source of search bias involves specifying dimension then it can be shown that it is likely to generalise the order in which we prune candidate hypotheses. Comparing ILP with DL. In Table 1 we compare control can be attempted with arbitrary Prolog specifications, but over inductive bias in ILP and DL. We consider the follow- may not terminate. ing, from Witten et al. (2017): language bias (hypothesis Verification in DL. To quote Katz et al. (2017), “Deep neu- space restriction), search bias (how the search through the ral networks are large, non-linear, and non-convex, and ver- hypothesis space is ordered), and simplicity bias (how over- ifying even simple properties about them is an NP-complete fitting is prevented). problem”. In practice, complete solvers can verify properties of networks with thousands of nodes, but time out for larger Bias ILP DL networks (Wang et al. 2018). Incomplete methods can verify Simplicity Bound on Not well understood properties of networks with ∼100 000 ReLU nodes (Singh program length (besides regularisers et al. 2019; Botoeva et al. 2020). Note that the smallest net- e.g. dropout & LR decay) works that achieve decent results on CIFAR10 have ∼50k Language User constraints, NN architecture nodes. Networks can be trained such that they are easier to target logic verify (Xiao et al. 2019). Search Search order, Local gradient search hypothesis scoring Post-hoc Model Editing Table 1: Realisations of types of inductive bias Definition 5 (Model Editing) Let L: D → M be a learn- ing algorithm. Let M ∈ M be a learned model. Let s be a When is control over inductive bias actually hand-coding specification. We apply model editing to M on specification solutions? The more inductive biases are customised, the s, if we find a model M 0 ∈ M that has property s without more the learning method resembles explicit programming re-applying the learning algorithm L. of a solution class. For example, when doing reinforcement learning it is possible to include information about how the Let d be a distance metric on M. We say that we success- task should be solved in the reward function. As more infor- fully edit M to fit specification s with respect to distance d mation is included, designing the reward function resembles if we find a model M 0 ∈ M that has property s and out of specifying a solution (Sutton and Barto 2018). In ILP, task- all models with property s has minimal distance from M . specific language biases are often unavoidable for perfor- Model Editing in ILP. The symbolic representation could mance reasons, but they risk pruning unexpected solutions, make ILP models easier to manipulate than DL models. ILP involve a good deal of expert human labour, and can lead to output models are very interpretable and it is relatively easy brittle systems which may not learn the problem structure so for humans to write logical sentences, which should make it much as they are passed it to begin with (Payani and Fekri in some cases possible to apply model editing. 2019). This problem could be mitigated by progress in au- The output model of ILP is a conjunction of logical tomating inductive bias choice in ILP. clauses. The model can easily be edited, by removing or adding individual clauses. If we simply add clause s to M , Verification of Specifications then we get a new model M 0 = M ∪ {s}, which satisfies s In cases where model specification does not give hard and has minimal distance to M with respect to the ‘rewrite guarantees about model behaviour, post-hoc verification is distance’. When adding clauses, one needs to ensure the new needed. That is, determining, given program M and specifi- model is still consistent. cation s, whether M ’s behaviour satisfies s. A form of post-hoc model editing has been applied to large neural networks, though only by automating the edits. Definition 4 (Specification) Let L: D → M be a learning The OpenAI Five agent was trained across several differ- algorithm. A specification s is a property such that for all ent architectures, with an automatic process for discovering models M ∈ M, the model satisfies the property or not. weights to copy (Raiman, Zhang, and Dennison 2019). The problem of verifying whether a model satisfies a Model Editing in DL. After training a large neural net- specification is NP-hard (Clarke and Emerson 1981), both work, we (practically speaking) obtain a black-box model. for a neural network (Katz et al. 2017) and for logic pro- This black-box is not easy to manipulate, owing to the num- grams (Madelaine and Martin 2018). ber of parameters and the distributed nature of its repre- Verification in ILP. In practice, verifying properties of an sentation. Through active learning or incremental learning output hypothesis is often easy. Suppose you have a propo- we can update the model - we could add a module that sitional theory and want to verify whether this satisfies the deals with exceptions, or fine-tune on extra training data for specification False. This is equivalent to solving satisfiabil- low-performing subgroups. However, these do not give us ity, and so is at least NP hard. We can verify whether an much control over exactly how the black-box changes (Set- ILP model M satisfies an arbitrary Datalog specification s tles 2009). by running resolution on M ∪ {¬s} to see if it derives False. Because the black-box is difficult to interpret, we do not In fact, this can be done in some cases where s is not in Dat- fully comprehend what function the network has learned and alog. For example, this could be done as long as s is in the so are not able to enhance it. Researchers can override the Bernays-Schönfinkel fragment, albeit in double-exponential output with a different learned module, but there is no low- time (Piskac, de Moura, and Bjørner 2008). The proof search level interactive combination of model and human insight. Transparency practice learning can involve many thousands of steps, and We consider the transparency of learned DL and ILP models, so this can take an impractically long time). and the transparency of the learning algorithms. Attributing the solution to individual training inputs. Transparency of the learned model. In contrast to DL Given an ILP output model and an input example, a human models, ILP outputs are represented in an explicit high-level can usually assess whether they are consistent. So in ILP it language, making them more transparent. We distinguish be- is relatively clear which example or background predicate is tween (Lipton 2018): a globally transparent or ‘simulatable’ causing the ILP algorithm to output a given model. model; and a locally transparent or ‘decomposable’ model. In DL however, it would be very difficult to (for instance) Decomposability. A decomposable model is one in which assess whether an image was a member of the training set of each part of the model - each input, parameter, and compu- a given model. That is, it is difficult to attribute aspects of tation - admits an intuitive explanation, independent of other the output model to individual inputs. instances of the model part (Lipton 2018). The many param- Inductive bias towards interpretability. User-supplied eters of a neural network form a distributed representation program constraints and bounds on program length mean of a nonlinear function, and as such it is unhelpful to reason that we only generate programs of a certain form, which can about individual parameters. be interpretable by construction. Moreover, control over in- An ILP output model is a conjunction of predicates and ductive bias itself can be seen as a form of accessibility. literals. When the background is human-specified, each in- dividual predicate will admit an intuitive explanation. When Discussion predicates are invented by the ILP system, the results can be counter-intuitive or long; however, they are still themselves We have argued that ILP has a number of safety properties encoded as decomposable clauses of intuitive features. that make it attractive compared to DL: Simulatability. A user can simulate a model if they can 1. ILP is convenient for specification, insofar as it is intuitive take input and work out what the model would output. More to encode examples and properties of correct behaviour; precisely, a model M is simulatable in proportion to the 2. ILP is robust to most syntactic changes in inputs; mean accuracy with which, after brief study of M , a pop- 3. The program templates, and bounds on program length ulation of users can reproduce the output of M on new data. give control over the inductive bias in ILP; A small usability study (n=16) found that access to an 4. We can verify whether an ILP model satisfies an arbitrary ILP program did allow users to simulate a concept they were Datalog specification by running resolution; unable to infer themselves (Muggleton et al. 2018b). It also 5. We can edit an ILP model by adding or removing clauses; found, as expected, that increasing the complexity reduced 6. ILP models are interpretable as they are quite transparent simulatability. and are reasonably accessible. Explainability of the learned model. Since ILP models are relatively transparent, explanations are redundant (except in Competitiveness of ILP. It is unlikely that the AI com- very large programs). DL explainability is a highly active munity will adopt ILP if its performance is not competitive. field of research (Gilpin et al. 2018) and has produced many Consider chemistry applications (Srinivasan et al. 1997): post-hoc tools, making use of: visualization, text explana- ILP continues to be applied (Kaalia et al. 2016), but DL ef- tions, feature relevance, explanations by example, model forts are now more extensive (Cova and Pais 2019). One simplification and local explanations (Arrieta et al. 2020). benchmark is suggestive: ILP found success in the early Transparency of the learning algorithm. In DL the learn- years of the Comparative Assessment of protein Structure ing algorithm optimises weights in a model that already has Prediction (Cootes, Muggleton, and Sternberg 2003), but the same architecture as the output model, such that during most submissions now use DL (Senior et al. 2019). These training we see many intermediate models. This implies that results may not be indicative of ILP’s current potential as many of the transparency properties of DL are relevant for far less research is being invested in ILP than in DL. As a its accessibility as well. On the other hand, in ILP we have a suggestive bound on the ratio of investment, compare the distinct training algorithm and output model. 130 researchers (worldwide) listed on the ILP community Individual model updates during learning. In DL, back- hub (Siebers 2019), to the 420 researchers at a single DL propagation attributes changes in the loss to individual lab, Berkeley AI Research, or to the 13,000 attendees of a weights. However, backpropagation can lead to local min- single DL conference, NeurIPS. This relative neglect might ima, and so sometimes weights are updated in a direction allow for performance gains from research into ILP and hy- opposite from the ideal direction. This, along with their (hu- brid ILP-DL systems. manly) incomprehensible representation imply that individ- A deeper concern is the limited domains in which ILP of- ual updates are not interpretable. fers its benefits. ILP generates logic programs, where DL ap- This contrasts with ILP, where each step of the learning proximates continuous functions. We have argued that logic algorithm occurs on a symbolic level (for instance, general- programs are more human interpretable, especially when the ising a candidate hypothesis through dropping one literal). In predicates used in the program represent concepts we know principle a human user could step through ILP learning and and use. Our discussion of ILP’s transparency only applies understand the concept represented at each step, the com- to domains where the data is already available on a symbolic plete effect of each change on the model coverage, and the level. Moreover, a major theme in recent AI, cognitive sci- particular data points that constrain the change (though in ence, and linguistics is that rule approaches are insufficient to express or learn most human-level concepts, where con- Cootes, A. P.; Muggleton, S.; and Sternberg, M. J. 2003. tinuous features and similarity to exemplars appear neces- The Automatic Discovery of Structural Principles Describ- sary (Rouder and Ratcliff 2006; Spivey 2008; Norvig 2012). ing Protein Fold Space. In Mol. Biol. Both of the above suggest a need to unify connection- Cova, T. F.; and Pais, A. A. 2019. Deep Learning for Deep ist and symbolic methods. Recent attempts implement re- Chemistry: Optimizing the Prediction of Chemical Patterns. lational or logical reasoning on neural networks (Garnelo Frontiers in Chemistry 7. and Shanahan 2019; Evans and Grefenstette 2018). How- ever, from a safety perspective, these unifications lose de- Cropper, A.; Dumančić, S.; and Muggleton, S. 2020. Turn- sirable properties. We hope that future versions not only in- ing 30: New Ideas in Inductive Logic Programming. In IJ- crease in performance, but also retain their safety potential. CAI. ILP as specification module. The above suggests a fruit- Cropper, A.; and Morel, R. 2020. Learning programs by ful role for ILP: as a specification generator in a mixed AI learning from failures. arXiv 2005.02259 . system. We may not be able to directly specify safety prop- Cropper, A.; and Tourret, S. 2018. Derivation reduction erties, but may be able to give positive and negative exam- of metarules in meta-interpretive learning. In International ples of safe behaviour. If it is natural to formulate these ex- Conference on Inductive Logic Programming. amples in natural language or logic, then ILP can generate hypotheses based on these partial specifications. Since ILP De Raedt, L.; Dries, A.; Thon, I.; Van den Broeck, G.; and output models are easy to interpret, we may be able to verify Verbeke, M. 2015. Inducing probabilistic relational rules whether they meet our preferences, and perhaps edit them to from probabilistic examples. In IJCAI. account for noisy or missing features. In certain cases (e.g. Edelmann, R.; and Kunčak, V. 2019. Neural-Network Datalog), it may be possible to formally verify the hypoth- Guided Expression Transformation. arXiv:1902.02194 . esis’ correctness. The specification can then be transferred Evans, R.; and Grefenstette, E. 2018. Learning Explanatory losslessly to any other learning system that can handle logi- Rules from Noisy Data. In JAIR. cal expressions (e.g. graph neural networks). Garnelo, M.; and Shanahan, M. 2019. Reconciling deep ILP’s differences with DL suggest solutions to DL’s safety learning with symbolic artificial intelligence: representing shortcomings. We are hopeful that hybrid systems can pro- objects and relations. In Current Opinion in Behavioral Sci- vide safety guarantees. ences. Acknowledgements Supported by UKRI studentships EP/S022937/1 and EP/S023356/1 and the AI Safety Re- Gilpin, L.; Bau, D.; Yuan, B.; Bajwa, A.; Specter, M.; and search Programme. Thanks to Alex Jackson for suggesting Kagal, L. 2018. Explaining Explanations: An Overview of ILP has potential as a specification generator. Interpretability of Machine Learning. In IEEE DSAA. Goodfellow, I.; Bengio, Y.; and Courville, A. 2016. Deep References Learning. MIT Press. http://www.deeplearningbook.org. Anderson, G.; Verma, A.; Dillig, I.; and Chaudhuri, S. 2020. Hüllermeier, E.; Fober, T.; and Mernberger, M. 2013. Induc- Neurosymbolic Reinforcement Learning with Formally Ver- tive Bias, Encyclopedia of Systems Biology. ified Exploration. In NeurIPS. Kaalia, R.; Srinivasan, A.; Kumar, A.; and Ghosh, I. 2016. Armstrong, S.; and Mindermann, S. 2018. Occam’s razor is ILP-assisted de novo drug design. Machine Learning 103. insufficient to infer the preferences of irrational agents. In Kashinath, K.; Marcus, P.; et al. 2019. Enforcing Physi- NeurIPS. cal Constraints in Neural Neural Networks through Differ- Arrieta, A. B.; Rodrı́guez, N. D.; Ser, J. D.; Bennetot, entiable PDE Layer. In ICLR DeepDiffEq workshop. A.; Tabik, S.; Barbado, A.; Garcı́a, S.; Gil-Lopez, S.; Katz, G.; Barrett, C.; Dill, D.; Julian, K.; and Kochenderfer, Molina, D.; Benjamins, R.; Chatila, R.; and Herrera, F. M. 2017. Reluplex: An Efficient SMT Solver for Verifying 2020. Explainable Artificial Intelligence (XAI): Concepts, Deep Neural Networks. Computer Aided Verification. Lec- taxonomies, opportunities and challenges toward responsi- ture Notes in Computer Science 10426. ble AI. Inf. Fusion 58. Kroy, M. 1976. A Partial Formalization of Kant’s Categori- Botoeva, E.; Kouvaros, P.; Kronqvist, J.; Lomuscio, A.; and cal Imperative. An Application of Deontic Logic to Classical Misener, R. 2020. Efficient Verification of ReLU-based Moral Philosophy. Kant-Studien 67. Neural Networks via Dependency Analysis. In AAAI. Lipton, Z. C. 2018. The mythos of model interpretability. Boytcheva, S. 2002. Overview of ILP Systems. In Cyber- Queue 16. netics and Information Technologies. Madelaine, F.; and Martin, B. 2018. On the complexity of Clarke, E. M.; and Emerson, E. A. 1981. The design and the model checking problem. SIAM J. Comput. 47. synthesis of synchronization skeletons using temporal logic. Márquez-Neila, P.; Salzmann, M.; and Fua, P. 2017. Impos- In Workshop on Logics of Programs. ing hard constraints on deep networks: Promises and limita- Condry, N. 2016. Meaningful Models: Utilizing Concep- tions. arXiv:1706.02025 . tual Structure to Improve Machine Learning Interpretability. McNamara, P. 2019. Deontic Logic. https://plato.stanford. arXiv:1607.00279 . edu/archives/sum2019/entries/logic-deontic/. Mitchell, T. M. 1980. The need for biases in learning gen- K.; and Hassabis, D. 2019. Improved protein structure pre- eralizations. Department of Computer Science, Laboratory diction using potentials from deep learning. Nature 577. for Computer Science Research. Settles, B. 2009. Active learning literature survey. Techni- Muggleton, S. 1999. Inductive Logic Programming: Issues, cal report, University of Wisconsin-Madison Department of results and the challenge of Learning Language in Logic. Computer Sciences. Artificial Intelligence 114. Siebers, M. 2019. inductive-programming community site. Muggleton, S.; Dai, W.-Z.; Sammut, C.; Tamaddoni- https://inductive-programming.org/people.html. Nezhad, A.; Wen, J.; and Zhou, Z.-H. 2018a. Meta- Singh, G.; Gehr, T.; Püschel, M.; and Vechev, M. 2019. An interpretive learning from noisy images. Machine Learning Abstract Domain for Certifying Neural Networks. ACM 107. Program. Lang. 3. Muggleton, S.; and de Raedt, L. 1994. Inductive Logic Pro- Spivey, M. 2008. The continuity of mind. Oxford University gramming: Theory and methods. The Journal of Logic Pro- Press. gramming 19/20. Srinivasan, A. 2006. The Aleph System. https://www.cs.ox. Muggleton, S.; Schmid, U.; Zeller, C.; Tamaddoni-Nezhad, ac.uk/activities/programinduction/Aleph/aleph.html. A.; and Besold, T. 2018b. Ultra-Strong Machine Learning: comprehensibility of programs learned with ILP. Machine Srinivasan, A.; King, R. D.; Muggleton, S.; and Sternberg, Learning 107. M. J. E. 1997. Carcinogenesis Predictions Using ILP. In Inductive Logic Programming, 7th International Workshop. Norvig, P. 2012. Chomsky and the two cultures of statistical learning. Significance 9. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; and Salakhutdinov, R. 2014. Dropout: A Simple Way to Prevent Ortega, P. A.; and Maini, V. 2018. Building safe Neural Networks from Overfitting. JMLR 15. artificial intelligence: specification, robustness, and as- surance. https://medium.com/@deepmindsafetyresearch/ Sutton, R. S.; and Barto, A. G. 2018. Reinforcement learn- building-safe-artificial-intelligence-52f5f75058f1. ing: An introduction. MIT press. Pathak, D.; Krähenbühl, P.; and Darrell, T. 2015. Con- Szegedy, C.; Zaremba, W.; Sutskever, I.; Bruna, J.; Erhan, strained Convolutional Neural Networks for Weakly Super- D.; Goodfellow, I.; and Fergus, R. 2014. Intriguing proper- vised Segmentation. In IEEE ICCV. ties of neural networks. In ICLR. Payani, A.; and Fekri, F. 2019. Inductive Logic Program- Szymanski, L.; McCane, B.; and Albert, M. 2018. The effect ming via Differentiable Deep Neural Logic Networks. CoRR of the choice of neural network depth and breadth on the size abs/1906.03523. of its hypothesis space. arXiv:1806.02460 . Peterson, C. 2014. The categorical imperative: Category the- Tausend, B. 1994. Representing biases for inductive logic ory as a foundation for deontic logic. Applied Logic 12. programming. In ECML. Piskac, R.; de Moura, L.; and Bjørner, N. 2008. Deciding ef- Vapnik, V.; and Chervonenkis, A. 2015. On the Uniform fectively propositional logic with equality. Technical report, Convergence of Relative Frequencies of Events to Their MSR-TR-2008-181, Microsoft Research. Probabilities. Springer. Platt, J. C.; and Barr, A. H. 1988. Constrained differential Wang, S.; Pei, K.; Whitehouse, J.; Yang, J.; and Jana, S. optimization. In NeurIPS. 2018. Efficient Formal Safety Analysis of Neural Networks. In NeurIPS. Poggio, T.; Liao, Q.; and Banburski, A. 2020. Complexity control by gradient descent in deep networks. Nature Com- Witten, I.; Frank, E.; Hall, M.; and Pal, C. 2017. Data munications 11. Mining: Practical Machine Learning Tools and Techniques. Morgan Kaufmann. Powell, D.; and Thévenod-Fosse, P. 2002. Dependability is- sues in ai-based autonomous systems for space applications. Xiao, K. Y.; Tjeng, V.; Shafiullah, N. M.; and Madry, A. In IARP-IEEE/RAS joint workshop. 2019. Training for Faster Adversarial Robustness Verifica- tion via Inducing ReLU Stability. In ICLR. Quinlan, J. R. 1990. Learning Logical Definitions from Re- lations. Machine Learning 5. Yudkowsky, E. 2011. Complex Value Systems are Required to Realize Valuable Futures. The Singularity Institute, San Raiman, J.; Zhang, S.; and Dennison, C. 2019. Neural Net- Francisco, CA. work Surgery with Sets. arXiv:1607.00279 . Zhang, C.; Bengio, S.; Hardt, M.; Recht, B.; and Vinyals, Rouder, J. N.; and Ratcliff, R. 2006. Comparing Exemplar- O. 2017. Understanding deep learning requires rethinking and Rule-Based Theories of Categorization. Current Direc- generalization. In ICLR. tions in Psychological Science 15. Zhang, H.; Chen, H.; Xiao, C.; Gowal, S.; Stanforth, R.; Li, Senior, A. W.; Evans, R.; Jumper, J.; Kirkpatrick, J.; Sifre, B.; Boning, D.; and Hsieh, C.-J. 2020. Towards Stable and L.; Green, T.; Qin, C.; Žı́dek, A.; Nelson, A. W. R.; Efficient Training of Verifiably Robust Neural Networks. In Bridgland, A.; Penedones, H.; Petersen, S.; Simonyan, K.; ICLR. Crossan, S.; Kohli, P.; Jones, D. T.; Silver, D.; Kavukcuoglu,