Ing. Martin Svatoš

Neural Markov Logic networks are a statistical relational model capable of generating relational structures. In this paper, we investigate how this particular model behaves in the setup of few-shot learning and show that Neural Markov Logic Networks are able to learn to generate small molecules from a handful of training examples without any pre-training.

Relational learning for knowledge base completion has been receiving considerable attention. Intuitively, rule-based strategies are clearly appealing, given their transparency and their ability to capture complex relational dependencies. In practice, however, pure rule-based strategies are currently not competitive with state-of-the-art methods, which is a reflection of the fact that (i) learning high-quality rules is challenging, and (ii) classical entailment is too brittle to cope with the noisy nature of the learned rules and the given knowledge base. In this paper, we introduce STRiKE, a new approach for relational learning in knowledge bases which addresses these concerns. Our contribution is three-fold. First, we introduce a new method for learning stratified rule bases from relational data. Second, to use these rules in a noise-tolerant way, we propose a strategy which extends k-entailment, a recently introduced cautious entailment relation, to stratified rule bases. Finally, we introduce an efficient algorithm for reasoning based on k-entailment.

Collective inference is a popular approach for solving tasks as knowledge graph completion within the statistical relational learning field. There are many existing solutions for this task, however, each of them is subjected to some limitation, either by restriction to only some learning settings, lacking interpretability of the model or theoretical test error bounds. We propose an approach based on cautious inference process which uses first-order rules and provides PAC-style bounds.

Neural-symbolic learning cycle is a machine learning paradigm for learning predictive models by a combination of both artificial neural networks and symbolic rules. There is a vast number of proposed methods for each part of the cycle, yet a proper experimental evaluation, e.g. asymptotic behavior, of the cycle is still missing. One key component for such an evaluation is to have a dataset with expert knowledge to verify whether the cycle produces meaningful results. The aim of this paper is to fill this gap by introducing novel artificial datasets which, in turn, can be used for a proper experimental evaluation of the cycle.

We present a method to prune hypothesis spaces in the con- text of inductive logic programming. The main strategy of our method consists in removing hypotheses that are equivalent to already consid- ered hypotheses. The distinguishing feature of our method is that we use learned domain theories to check for equivalence, in contrast to existing approaches which only prune isomorphic hypotheses. Specifically, we use such learned domain theories to saturate hypotheses and then check if these saturations are isomorphic. While conceptually simple, we exper- imentally show that the resulting pruning strategy can be surprisingly effective in reducing both computation time and memory consumption when searching for long clauses, compared to approaches that only con- sider isomorphism.

Lifted Relational Neural Networks (LRNNs) describe relational domains using weighted first-order rules which act as templates for constructing feed-forward neural networks. While previous work has shown that using LRNNs can lead to state-of-the-art results in various ILP tasks, these results depended on hand-crafted rules. In this paper, we extend the framework of LRNNs with structure learning, thus enabling a fully automated learning process. Similarly to many ILP methods, our structure learning algorithm proceeds in an iterative fashion by top-down searching through the hypothesis space of all possible Horn clauses, considering the predicates that occur in the training examples as well as invented soft concepts entailed by the best weighted rules found so far. In the experiments, we demonstrate the ability to automatically induce useful hierarchical soft concepts leading to deep LRNNs with a competitive predictive power.

Reducing accidental complexity in planning problems is a well-established method for increasing efficiency of classical planning. Removal of superfluous facts and actions, and problem transformation by recursive macro actions are representatives of such methods working directly on input planning problems. Despite of its general applicability and thorough theoretical analysis, there is only a sparse amount of experimental results. In this paper, we adopt selected reduction methods from literature and amend them with a generalization-based reduction scheme and auxiliary reductions. We show that all presented reductions are polynomial in time to the size of an input problem. All reductions applied in a recursive manner produce only safe (solution preserving) abstractions of the problem, and they can implicitly represent exponentially long plans in a compact form. Experimentally, we validate efficiency of the presented reductions on the IPC benchmark set and show average 24% reduction over all problems. Additionally, we experimentally analyze the trade-off between increase of coverage and decrease of the plan quality