Dr Alireza Tamaddoni-Nezhad

Unlike most computer vision approaches, which depend on hundreds or thousands of training images, humans can typically learn from a single visual example. Humans achieve this ability using background knowledge. Rule-based machine learning approaches such as Inductive Logic Programming (ILP) provide a framework for incorporating domain specific background knowledge. These approaches have the potential for human-like learning from small data or even one-shot learning, i.e. learning from a single positive example. By contrast, statistics based computer vision algorithms, including Deep Learning, have no general mechanisms for incorporating background knowledge. This paper presents an approach for one-shot rule learning called One-Shot Hypothesis Derivation (OSHD) based on using a logic program declarative bias. We apply this approach to two challenging human-like computer vision tasks: 1) Malayalam character recognition and 2) neurological diagnosis using retinal images. We compare our results with a state-of-the-art Deep Learning approach, called Siamese Network, developed for one-shot learning. The results suggest that our approach can generate human-understandable rules and outperforms the deep learning approach with a significantly higher average predictive accuracy.

Sustainable management of ecosystems and growth in agricultural productivity is at the heart of the United Nations' Sustainable Development Goals for 2030. New management regimes could revolutionize agricultural production, but require an evaluation of the risks and opportunities. Replacing existing conventional weed management with genetically modified, herbicide-tolerant (GMHT) crops, for example, might reduce herbicide applications and increase crop yields, but remains controversial owing to concerns about potential impacts on biodiversity. Until now, such new regimes have been assessed at the species or assemblage level, whereas higher-level ecological network effects remain largely unconsidered. Here, we conduct a large-scale network analysis of invertebrate communities across 502 UK farm sites to GMHT management in different crop types. We find that network-level properties were overwhelmingly shaped by crop type, whereas network structure and robustness were apparently unaltered by GMHT management. This suggests that taxon-specific effects reported previously did not escalate into higher-level systemic structural change in the wider agricultural ecosystem. Our study highlights current limitations of autecological assessments of effect in agriculture in which species interactions and potential compensatory effects are overlooked. We advocate adopting the more holistic system-level evaluations that we explore here, which complement existing assessments for meeting our future agricultural needs.

During the 1980s Michie defined Machine Learning in terms of two orthogonal axes of performance: predictive accuracy and comprehensibility of generated hypotheses. Since predictive accuracy was readily measurable and comprehensibility not so, later definitions in the 1990s, such as Mitchell’s, tended to use a one-dimensional approach to Machine Learning based solely on predictive accuracy, ultimately favouring statistical over symbolic Machine Learning approaches. In this paper we provide a definition of comprehensibility of hypotheses which can be estimated using human participant trials. We present two sets of experiments testing human comprehensibility of logic programs. In the first experiment we test human comprehensibility with and without predicate invention. Results indicate comprehensibility is affected not only by the complexity of the presented program but also by the existence of anonymous predicate symbols. In the second experiment we directly test whether any state-of-the-art ILP systems are ultra-strong learners in Michie’s sense, and select the Metagol system for use in humans trials. Results show participants were not able to learn the relational concept on their own from a set of examples but they were able to apply the relational definition provided by the ILP system correctly. This implies the existence of a class of relational concepts which are hard to acquire for humans, though easy to understand given an abstract explanation. We believe improved understanding of this class could have potential relevance to contexts involving human learning, teaching and verbal interaction.

We foresee a new global-scale, ecological approach to biomonitoring emerging within the next decade that can detect ecosystem change accurately, cheaply, and generically. Next-generation sequencing of DNA sampled from the Earth’s environments would provide data for the relative abundance of operational taxonomic units or ecological functions. Machine-learning methods would then be used to reconstruct the ecological networks of interactions implicit in the raw NGS data. Ultimately, we envision the development of autonomous samplers that would sample nucleic acids and upload NGS sequence data to the cloud for network reconstruction. Large numbers of these samplers, in a global array, would allow sensitive automated biomonitoring of the Earth’s major ecosystems at high spatial and temporal resolution, revolutionising our understanding of ecosystem change. Next-generation sequencing (NGS) can used to sample nucleic acids in the environment for the presence of species and ecological functions. Machine-learning software can search for ‘the ghosts of interactions past’ in the raw NGS data to reconstruct the networks of ecological interactions. NGS data and machine-learning in the cloud could be combined in the next generation of global biomonitoring. Autonomous NGS samplers would sequence and upload data for ecological network reconstruction, to detect ecosystem change accurately, cheaply and generically. Reconstruction of highly replicated networks of ecological interaction, using this next generation of biomonitoring, would provide general ecological information for ecosystem comparison and a revolution in the breadth of our understanding of the ecology of ecosystem change.

Species diversity, and the various interactions that occur between species, supports ecosystems functioning and benefit human societies. Monitoring the response of species interactions to human alterations of the environment is thus crucial for preserving ecosystems. Ecological networks are now the standard method for representing and simultaneously analyzing all the interactions between species. However, deciphering such networks requires considerable time and resources to observe and sample the organisms, to identify them at the species level and to characterize their interactions. Next-generation sequencing (NGS) techniques, combined with network learning and modelling, can help alleviate these constraints. They are essential for observing cryptic interactions involving microbial species, as well as short-term interactions such as those between predator and prey. Here, we present three case studies, in which species associations or interactions have been revealed with NGS. We then review several currently available statistical and machine-learning approaches that could be used for reconstructing networks of direct interactions between species, based on the NGS co-occurrence data. Future developments of these methods may allow us to discover and monitor species interactions cost-effectively, under various environmental conditions and within a replicated experimental design framework.

Despite early interest Predicate Invention has lately been under-explored within ILP. We develop a framework in which predicate invention and recursive generalisations are implemented using abduction with respect to a meta-interpreter. The approach is based on a previously unexplored case of Inverse Entailment for Grammatical Inference of Regular languages. Every abduced grammar H is represented by a conjunction of existentially quantified atomic formulae. Thus ¬H is a universally quantified clause representing a denial. The hypothesis space of solutions for ¬H can be ordered by θ-subsumption. We show that the representation can be mapped to a fragment of Higher-Order Datalog in which atomic formulae in H are projections of first-order definite clause grammar rules and the existentially quantified variables are projections of first-order predicate symbols. This allows predicate invention to be effected by the introduction of first-order variables. Previous work by Inoue and Furukawa used abduction and meta-level reasoning to invent predicates representing propositions. By contrast, the present paper uses abduction with a meta-interpretive framework to invent relations. We describe the implementations of Meta-interpretive Learning (MIL) using two different declarative representations: Prolog and Answer Set Programming (ASP). We compare these implementations against a state-of-the-art ILP system MC-TopLog using the dataset of learning Regular and Context-Free grammars as well learning a simplified natural language grammar and a grammatical description of a staircase. Experiments indicate that on randomly chosen grammars, the two implementations have significantly higher accuracies than MC-TopLog. In terms of running time, Metagol is overall fastest in these tasks. Experiments indicate that the Prolog implementation is competitive with the ASP one due to its ability to encode a strong procedural bias. We demonstrate that MIL can be applied to learning natural grammars. In this case experiments indicate that increasing the available background knowledge, reduces the running time. Additionally ASPM (ASP using a meta-interpreter) is shown to have a speed advantage over Metagol when background knowledge is sparse. We also demonstrate that by combining Metagol R (Metagol with a Regular grammar meta-interpreter) and Metagol CF (Context-Free meta-interpreter) we can formulate a system, Metagol RCF , which can change representation by firstly assuming the target to be Regular, and then failing this, switch to assuming it to be Context-Free. Metagol RCF runs up to 100 times faster than Metagol CF on grammars chosen randomly from Regular and non-Regular Context-Free grammars.

In a previous paper we described a machine learning approach which was used to automatically generate food-webs from national-scale agricultural data. The learned food-webs in the previous study consist of hundreds of ground facts representing trophic links between individual species. These species food-webs can be used to explain the structure and dynamics of particular eco-systems, however, they cannot be directly used as general predictive models. In this paper we describe the first steps towards this generalisation and present initial results on (i) learning general functional food-webs (i.e. trophic links between functional groups of species) and (ii) meta-interpretive learning (MIL) of general predictive rules (e.g. about the effect of agricultural management). Experimental results suggest that functional food-webs have at least the same levels of predictive accuracies as species food-webs despite being much more compact. In this paper we also present initial experiments where predicate invention and recursive rule learning in MIL are used to learn food-webs as well as predictive rules directly from data.

Interaction between species in microbial communities plays an important role in the functioning of all ecosystems, from cropland soils to human gut microbiota. Many statistical approaches have been proposed to infer these interactions from microbial abundance information. However, these statistical approaches have no general mechanisms for incorporating existing ecological knowledge in the inference process. We propose an Abductive/Inductive Logic Programming (A/ILP) framework to infer microbial interactions from microbial abundance data, by including logical descriptions of different types of interaction as background knowledge in the learning. This framework also includes a new mechanism for estimating the probability of each interaction based on the frequency and compression of hypotheses computed during the abduction process. This is then used to identify real interactions using a bootstrapping, re-sampling procedure. We evaluate our proposed framework on simulated data previously used to benchmark statistical interaction inference tools. Our approach has comparable accuracy to SparCC, which is one of the state-of-the-art statistical interaction inference algorithms, but with the the advantage of including ecological background knowledge. Our proposed framework opens up the opportunity of inferring ecological interaction information from diverse ecosystems that currently cannot be studied using other methods.

Apperly, I., Bundy, A., Cohn, A. , Colton, S., Cussens, J., d'Avila Garcez, A. S. , Hahn, U., Jamnik, M., Jay, C., Mareschal, D., Sammut, C., Schmid, U., Seed, A., Stahl, B., Steedman, M. & Tamaddoni-Nezhad, A.view all authorsEPJS_limit_names_shown_load( 'creators_name_29561_et_al', 'creators_name_29561_rest' ); (2022). Preface. CEUR Workshop Proceedings, 3227, ISSN 1613-0073

Statistical machine learning is widely used in image classification. However, most techniques (1) require many images to achieve high accuracy and (2) do not provide support for reasoning below the level of classification, and so are unable to support secondary reasoning, such as the existence and position of light sources and other objects outside the image. This paper describes an Inductive Logic Programming approach called Logical Vision which overcomes some of these limitations. LV uses Meta-Interpretive Learning (MIL) combined with low-level extraction of high-contrast points sampled from the image to learn recursive logic programs describing the image. In published work LV was demonstrated capable of high-accuracy prediction of classes such as regular polygon from small numbers of images where Support Vector Machines and Convolutional Neural Networks gave near random predictions in some cases. LV has so far only been applied to noise-free, artificially generated images. This paper extends LV by (a) addressing classification noise using a new noise-telerant version of the MIL system Metagol, (b) addressing attribute noise using primitive-level statistical estimators to identify sub-objects in real images, (c) using a wider class of background models representing classical 2D shapes such as circles and ellipses, (d) providing richer learnable background knowledge in the form of a simple but generic recursive theory of light reflection. In our experiments we consider noisy images in both natural science settings and in a RoboCup competition setting. The natural science settings involve identification of the position of the light source in telescopic and microscopic images, while the RoboCup setting involves identification of the position of the ball. Our results indicate that with real images the new noise-robust version of LV using a single example (i.e. one-shot LV) converges to an accuracy at least comparable to a thirty-shot statistical machine learner on both prediction of hidden light sources in the scientific settings and in the RoboCup setting. Moreover, we demonstrate that a general background recursive theory of light can itself be invented using LV and used to identify ambiguities in the convexity/concavity of objects such as craters in the scientific setting and partial obscuration of the ball in the RoboCup setting.

Since the late 1990s predicate invention has been under-explored within inductive logic programming due to difficulties in formulating efficient search mechanisms. However, a recent paper demonstrated that both predicate invention and the learning of recursion can be efficiently implemented for regular and context-free grammars, by way of metalogical substitutions with respect to a modified Prolog meta-interpreter which acts as the learning engine. New predicate symbols are introduced as constants representing existentially quantified higher-order variables. The approach demonstrates that predicate invention can be treated as a form of higher-order logical reasoning. In this paper we generalise the approach of meta-interpretive learning (MIL) to that of learning higher-order dyadic datalog programs. We show that with an infinite signature the higher-order dyadic datalog class $$H^2_2$$ H 2 2 has universal Turing expressivity though $$H^2_2$$ H 2 2 is decidable given a finite signature. Additionally we show that Knuth–Bendix ordering of the hypothesis space together with logarithmic clause bounding allows our MIL implementation Metagol $$_{D}$$ D to PAC-learn minimal cardinality $$H^2_2$$ H 2 2 definitions. This result is consistent with our experiments which indicate that Metagol $$_{D}$$ D efficiently learns compact $$H^2_2$$ H 2 2 definitions involving predicate invention for learning robotic strategies, the East–West train challenge and NELL. Additionally higher-order concepts were learned in the NELL language learning domain. The Metagol code and datasets described in this paper have been made publicly available on a website to allow reproduction of results in this paper.

Recent papers have demonstrated that both predicate invention and the learning of recursion can be efficiently implemented by way of abduction with respect to a meta-interpreter. This paper shows how Meta-Interpretive Learning (MIL) can be extended to implement a Bayesian posterior distribution over the hypothesis space by treating the meta-interpreter as a Stochastic Logic Program. The resulting \documentclass[12pt]{minimal}

Autonomous driving decision-making is a challenging task due to the inherent complexity and uncertainty in traffic. For example, adjacent vehicles may change their lane or overtake at any time to pass a slow vehicle or to help traffic flow. Anticipating the intention of surrounding vehicles, estimating their future states and integrating them into the decision-making process of an automated vehicle can enhance the reliability of autonomous driving in complex driving scenarios. This paper proposes a Prediction-based Deep Reinforcement Learning (PDRL) decision-making model that considers the manoeuvre intentions of surrounding vehicles in the decision-making process for highway driving. The model is trained using real traffic data and tested in various traffic conditions through a simulation platform. The results show that the proposed PDRL model improves the decision-making performance compared to a Deep Reinforcement Learning (DRL) model by decreasing collision numbers, resulting in safer driving.

Humanity is facing existential, societal challenges related to food security, ecosystem conservation, antimicrobial resistance, etc, and Artificial Intelligence (AI) is already playing an important role in tackling these new challenges. Most current AI approaches are limited when it comes to ‘knowledge transfer’ with humans, i.e. it is difficult to incorporate existing human knowledge and also the output knowledge is not human comprehensible. In this chapter we demonstrate how a combination of comprehensible machine learning, text-mining and domain knowledge could enhance human-machine collaboration for the purpose of automated scientific discovery where humans and computers jointly develop and evaluate scientific theories. As a case study, we describe a combination of logic-based machine learning (which included human-encoded ecological background knowledge) and text-mining from scientific publications (to verify machine-learned hypotheses) for the purpose of automated discovery of ecological interaction networks (food-webs) to detect change in agricultural ecosystems using the Farm Scale Evaluations (FSEs) of genetically modified herbicide-tolerant (GMHT) crops dataset. The results included novel food-web hypotheses, some confirmed by subsequent experimental studies (e.g. DNA analysis) and published in scientific journals. These machine-leaned food-webs were also used as the basis of a recent study revealing resilience of agro-ecosystems to changes in farming management using GMHT crops.

Classical biomonitoring techniques have focused primarily on measures linked to various biodiversity metrics and indicator species. Next-generation biomonitoring (NGB) describes a suite of tools and approaches that allow the examination of a broader spectrum of organizational levels-from genes to entire ecosystems. Here, we frame 10 key questions that we envisage will drive the field of NGB over the next decade. While not exhaustive, this list covers most of the key challenges facing NGB, and provides the basis of the next steps for research and implementation in this field. These questions have been grouped into current- and outlook-related categories, corresponding to the organization of this paper.

Plant diseases are one of the main causes of crop loss in agriculture. Machine Learning, in particular statistical and neural nets (NNs) approaches, have been used to help farmers identify plant diseases. However, since new diseases continue to appear in agriculture due to climate change and other factors, we need more data-efficient approaches to identify and classify new diseases as early as possible. Even though statistical machine learning approaches and neural nets have demonstrated state-of-the-art results on many classification tasks, they usually require a large amount of training data. This may not be available for emergent plant diseases. So, data-efficient approaches are essential for an early and precise diagnosis of new plant diseases and necessary to prevent the disease’s spread. This study explores a data-efficient Inductive Logic Programming (ILP) approach for plant disease classification. We compare some ILP algorithms (including our new implementation, PyGol) with several statistical and neural-net based machine learning algorithms on the task of tomato plant disease classification with varying sizes of training data set (6, 10, 50 and 100 training images per disease class). The results suggest that ILP outperforms other learning algorithms and this is more evident when fewer training data are available.