Conference 2019 Abstracts

The abstracts will be added to this page until June on a continuous basis.

Explainable AI and regulatory framework

European regulations require that products marketed in the territory must be safe and reliable. In this context, it is necessary for AI developers to be able to design systems that meet sets of criteria to ensure that products comply with existing and soon-to-come regulatory requirements. From the point of view of the lawyer and the evaluator, it is also essential that system inspection be made possible, in particular for the purpose of liability determination or conformity assessment.

However, the AI stakeholders face two main issues: first, the scarcity of reference frameworks (optimal characteristics, test methods, performance thresholds, etc.), such as normative and regulatory standards, is a limitation to the development and deployment of intelligent systems. On the other hand, system analysis requires methods for querying and testing AI algorithms that may be provided by explainability solutions (whether by design or with software overlays).

In the spirit of the European New Approach Directives, technical characteristics of explainability solutions should not be constrained by regulations, but guidance may be provided through standards. As an independant evaluating and certifying body, LNE contributes to the development of standards and test methods for the qualification of AI systems. LNE wishes to explore the design of a reference framework for explainability, including the type of information to be extracted for compliance assessment, the conditions of the application of explainability solutions, but also the assessment of the performance of these solutions.

LNE expects to lead a proposal gathering partners offering different implementation methods for explainable AI, for at least one application which may be subject to specific regulations (such as medical applications).

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

Explainable AI with Knowledge Graphs and Semantic Explanations

Semantic technologies, such as knowledge graphs, ontologies and reasoning have been developed as a bridge between human and machine conceptualizations of a domain of interest.

They may provide human-centric and semantic interpretation and be used to produce semantic explanations, i.e. explanations based on semantic concepts coming from knowledge graphs and ontologies.

I will show several ideas and proposals on how knowledge graphs, ontologies and schemas, both from arbitrary domain and from the domain of machine learning in particular, may be combined with machine learning for: i) providing semantic explanations, ii) facilitating generation of other types of explanations (textual or visual explanations).

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

The role of Semantic web in Biodiversity data management

The biodiversity research discipline studies the totality and variability of organisms, their morphology and genetics, life history and habitats, and geographical ranges (Srivastava, 2011) . Biodiversity is usually used to refer to biological diversity at three levels: genetics, species, and ecology. The current rapid loss of biodiversity is cause for alarm. Biodiversity has a great and direct impact on human life and the survival of mankind. It affects our access to clean water, food, fuel, medical drug compounds, and animal feed. It is thus crucial to raise awareness for the importance of biodiversity and to develop methods to better understand and preserve it. Globally, a number of initiatives, such as the Convention on Biological Diversity (CBD) , the Intergovernmental Platform on Biodiversity (IPBES) , have been undertaken. These initiatives require scientific research into primary biological, chemical, and physical processes to develop predictive models and inform strategic decisions. To achieve these goals, reliable and constant biodiversity data should be made available in a consistent way. Biodiversity Informatics is the application of informatics techniques to biodiversity data for improved management, presentation, discovery, exploration and analysis. It is clear that improved discovery and accessibility of biodiversity data helps addressing both scientific and social issues. Furthermore, it is essential for informed decisions for sustainable development of biotic resources.

To deal with these heterogeneous datasets, a set of data management techniques should be applied. The semantic web enhances data exchange, discovery, and integration by providing common formats to achieve a formalized conceptual scheme. Therefore, in this presentation, we are going to draw links and relationships between Semantic web and its roles in managing biodiversity data. In this talk, we first demonstrate the importance of semantics in biodiversity data management, we then survey current approaches proposed so far, focusing on our own development in this direction. After that we are going to demonstrate the links between machine learning and data integration. Finally, we also aim to describe the future trends and research issues still need to be faced.

Keynote talk

Novel Computational Approaches for Environmental Sustainability (June 12)

Explainable artificial intelligence for physical and technical systems

Physical and technical systems are peculiar for machine learning.

The solution of a variety of engineering problems necessitates the identification of an accurate system model serving as a basis of its supervisory control. While physical laws provide a way of generalizable modeling, the main danger originates in the incompleteness of factors taken into account.

Observation-based identification learns a good phenomenological model in terms of numerical approximation. However, it is tough to associate such a model with a priori knowledge. This background is frequently formulated as complex engineering models, partly describing the system only qualitatively. This discrepancy is a significant cause of prohibiting explainability.

The proposed approach originates in our experience in the performability analysis of critical infrastructures. Here a qualitative model summarizes the interactions in the system under evaluation, as extracted of operation logs and outcomes of benchmark experiments. (This model-building process can be well-supported by Inductive Logic Programming for Answer Set Programming, a special kind of machine learning). These models are well-interpretable, as they enrich an a priori engineering model with the newly extracted knowledge, thus delivering a representation directly consumable for the domain expert.

Evolving technologies, like physically constrained neural networks, integrate this model into the learning model to assure, that the phenomenological model complies with the engineering one. Moreover, the qualitative model may serve as a checking automaton during runtime in critical applications.

Short talk or poster (to be defined)

Explainable Machine Learning-based Artificial Intelligence (June 11)

Online explainability

The growing impact that artificial intelligence is having on our everyday lives, combined with the opaqueness of deep learning technology, has brought the problem of reliability and explainability of predictive systems to the top of the AI community agenda. However, most research on explainable learning focuses on post-hoc explanation, where one aims at explaining the reasons for the predictions made by a learned model. The case of Alice and Bob, two chatbots that Facebook shut down after discovering that they ended up developing their own “secret language” to communicate, is a clear example of the limitations of this approach to explainability. We argue that online explainability, in which the user is involved in the learning loop and interactively provides feedback to guide the learning process to the desired direction, is crucial in order to develop truly reliable and trustworthy AI.

Short talk or poster (to be defined)

Explainable Machine Learning-based Artificial Intelligence (June 11)

On Combing Deep-Learning and Classical Control Theoretic Approaches

End to end deep learning of control policies have gathered much attention in recent times. Their attractiveness stems from the fact that in principal they can work with very minimal assumptions on the problem structure (robot dynamics, environment model etc.). But at the same time, control policies learned through end-to-end approaches have a certain opaqueness, i.e if it does not work, it is difficult to pinpoint the exact reason.

On the other hand, conventional control theoretical motion planning and control comes with certain guarantees and explainability. For example, we can answer questions like, whether robot state converges from a certain set of initial conditions under the given feedback control policy. However, classical approaches require more information and assumptions about the problem structure.

In our research group, we are developing ways to optimally integrate deep-learning based approaches with classical control theoretic algorithms for safety critical applications like autonomous driving and human-robot collaborative manufacturing. The key focus is on understanding on what parameters of control theoretic algorithms needs to be learned in order to make them reliable in real world setting. The explainability of our approach stem from the fact that learned parameters have a physical meaning and thus the performance shortcomings can be clearly explained and analyzed.

Short talk or poster (to be defined)

Explainable Machine Learning-based Artificial Intelligence (June 11)

Cloud-based Federated Learning for Environmental Data

Analysis of environmental data is the key to most computational approaches for sustainability. However, such data feature complex spatial pattern at different scales, due to the combination of several spatial phenomena or various influencing factors. Moreover, they are highly dispersed geographically. For this reason, there is the need to develop approaches that can easily scale with the high volume of data coming from different sources and the computational demand of their analysis.

Federated learning has been proposed as a solution to these issues. In this approach, data are analyzed at their source (e.g. on IoT devices). This allows (1) high scalability, (2) low network overhead, (3) specialized machine learning models, and (4) protection of privacy. To account for global information, an aggregate model still co-exists with local models. However, data-collecting devices are usually battery-powered and have limited storage and computational resources. Since our goal is to provide accurate environmental data analysis, respecting the hardware constraints of measurement devices, we need to explore solutions that enhance the capabilities of measuring devices by exploiting Cloud/Edge resources.

To this end, we plan to (1) apply dynamic data reduction and compression techniques, to ensure efficient use of the existing network and storage infrastructure, keeping only the most relevant data at data sources, and (2) exploit dynamic offloading of federated learning processes to the Cloud/Edge, in order to increase battery lifetime of measurement devices.

Short talk

Novel Computational Approaches for Environmental Sustainability (June 12)

Exploring Internal Representations and Extracting Rules from Deep Neural Networks

Artificial deep neural networks are a powerful tool, able to extract information from large datasets and, using this acquired knowledge, make accurate predictions on previously unseen data. As a result, they are being applied in a wide variety of domains ranging from genomics to autonomous driving, from speech recognition to gaming. Many areas, where neural network-based solutions can be applied, require a validation, or at least some explanation, of how the system makes its decisions. This is especially true in the medical domain where such decisions can contribute to the survival or death of a patient. Unfortunately, the very large number of parameters required by deep neural networks is extremely challenging to cope with for explanation methods, and these networks remain for the most part black boxes. This demonstrates the real need for accurate explanation methods able to scale with this large quantity of parameters and to provide useful information to a potential user. Our research aims at providing tools and methods to improve the interpretability of deep neural networks.

In this context, we developed a method allowing a user to interrogate a trained neural network and reproduce internal representations, at various depths within the network. This allows for the discovery of biases that might have been overlooked in the training dataset and enable the user to verify and potentially discover new features that have been captured from the data by the network.

Another tool, based on rule extraction, is a method that emphasizes the regions of an image that are relevant to a certain class, through a local approximation of a neural net. This method is of particular interest when the detection of a certain feature or characteristic is particularly complex, and where artificial neural nets exceed human performance. This is especially the case in some medical diagnosis tasks.

To understand how features extracted by the network are combined to produce specific predictions, a third approach aims at extracting logical rules that reflect the behavior of the network’s fully connected layers. Such approach consists in (1) using a trained network to extract features from a set of images, (2) training a Random Forest to create a set of rules, based on those features, that behave in the same manner than the network, and (3) ranking those rules according to their contribution to the prediction. An analyst can then select the top-N rules allowing for an interpretation.

Short talk

Explainable Machine Learning-based Artificial Intelligence (June 11)

From shallow to deep learning for inverse imaging problems: Some recent approaches

In this talk we discuss the idea of data-driven regularisers for inverse imaging problems. We are in particular interested in the combination of model-based and purely data-driven image processing approaches. In this context we will make a journey from “shallow” learning for computing optimal parameters for variational regularisation models by bilevel optimization to the investigation of different approaches that use deep neural networks for solving inverse imaging problems. Alongside all approaches that are being discussed, their numerical solution and available solution guarantees will be stated.

Keynote talk

Explainable Machine Learning-based Artificial Intelligence (June 11)

Towards an explainable and convivial AI based tools: Illustration on medicine applications

Since 2010, the numerical Artificial Intelligence (AI) based on Machine Learning (ML) has produced impressive results, mainly in the fields of the pattern recognition and the natural language processing, succeeding to the previous dominance of the symbolic AI, centered on the logical reasoning. The integration of ML methods into industrial processes gives hope for new growth drivers. These impressive results could be considered in a first approach as the end of the mathematical models as the statistical analysis is able to reproduce phenomena. In true, Machine Learning is based on inductive models theorized by Francis Bacon in 1620. The use of inductive models requires to explain the prediction obtained on data, which is currently not often the case for industrial Machine Learning applications.

Consequently, the operational benefit of using Machine Learning methods is reco\-gnized but is hampered by the lack of understanding of their mechanisms, at the origin of operational, legal and ethical operational problems. This affects highly the operational acceptability of AI tools. This is largely dependent on the ability of engineers, decision-makers and users to understand the meaning and the properties of the results produced by these tools. In addition, the increasing delegation of decision-making offered by AI tools competes with tried and tested business rules, sometimes constituting certified expert systems. Machine Learning could be thus consider now as a colossus with a feet of clay. It is important to note that this difficult problem will not be solved only by mathematicians and by computer scientists. Indeed, it requires a large scientific collaboration for example with philosophers of science to investigate the properties of the inductive model, cognitive psychologists to evaluate the quality of an explanation and anthropologists to study the relation and the communication between humans and these AI tools.

The first part of the talk presents the challenges and the benefits coming from Artificial Intelligence for Industry and Services, in particular for the medicine. Medicine is changing in depth its paradigm moving from a reactive to a proactive discipline for reducing the costs while improving the healthcare quality. It is useful to remember that, before the success of Machine Learning, some automatic healthcare tools have been developed. For example, the MyCin healthcare program , developed in the seventies at Standford University, was developed to identify bacteria causing severe infections, such as bacteremia and meningitis and recommend antibiotics, with the dosage adjusted for patient's body weight. It was based on the Good Old-Fashioned AI (expert system). It is relevant to see that MyCin was never actually used in practice not for any weakness in its performance but largely for ethical and legal issues related to the use of computers in medicine. It was also already difficult to explain the logic of its operation and even more to detect contradictions.

The second part of the talk summarizes our research activities conducted with Frank Varenne, philosopher of science, and Judith Nicogossian, anthropobiolgist. Its main objectives is to provide and evaluate explanations of ML methods tools considered as a black box. The first step of this project, presented in this talk, is to show that the validation of this black box differs epistemologically from the one set up in the framework of mathematical and causal modeling of physical phenomena. The form of the explanation has to be evaluated and chosen to minimize the cognitive bias of the user. This also raises an ethical problem about the possible drift of producing more persuasive and transparent explanations. The evaluation must therefore take into account the management of the compromise between the need for transparency and the need for intelligibility. Another important point concerns the conviviality of the AI based tool that is to say the user's capability to work with independent efficiency. A philosophical and anthropological approach is required to define the conviviality of an AI tool which will be translated in terms of rules guiding its conception. Last but not least, an anthropological standpoint will be summarized in particular in the definition of the nature and the properties of the "phygital" communication, between IA and users.

Finally, the last part of the talk proposes some future research directions needed in our opinion to be included the CHIST-ERA program.

Keynote talk

Explainable Machine Learning-based Artificial Intelligence (June 11)

Explanation of Smart 5G Network Intrusion Detection using Attack Trees

Successful Intrusion Detection systems heavily rely on machine learning to detect anomaly. However, particularly in 5G networks, detected attacks contain complex information representing technical details about the network components (e.g., virtual BBU (vBBU), virtual RRH (vRRH), controllers, NFV orchestrator, involved EPC functions, etc.), its heterogeneous structure, security policies, and involved actors and their capabilities. This heterogeneous 5G infrastructure makes it hard for users to interpret machine generated attack data. Explanation is needed to clarify the attacks to users. This can happen using visualization techniques, for example, interactive tree graphs for improved user interaction allowing zooming in and out of details of attacks. In addition, explanation is needed to highlight which parts of attacks target which parts of the 5G network infrastructure and what parts of the security policies are violated. The challenge is to link up the Intrusion Detection Intelligence, analyse it, explain it, and feed back incident response decisions to users as well as to different levels of the 5G network infrastructure to enforce security policies in response to detected attacks. An important backbone for this process is to have models in which these heterogeneous scenarios can be encoded adequately yet concisely. A possibility is to use logical representation. However, the logics need to be powerful enough to represent entities, structures, and policies and yet rigorous and sufficiently supported with analysis and verification capabilities. Candidates are higher order temporal logics extended with attack trees and other security notions. A demonstrator platform will be provided using a cloud-native 5G set-up and Software Defined Network controllers.

Short talk or poster (to be defined)

Explainable Machine Learning-based Artificial Intelligence (June 11)

Explaining Visual Classification using Attributes

The performance of deep Convolutional Neural Networks (CNN) has been reaching or even exceeding the human level on large number of tasks. Some examples are image classification, Mastering Go game, speech understanding etc. However, their lack of decomposability into intuitive and understandable components make them hard to interpret, i.e. no information is provided about what makes them arrive at their prediction.

We propose a technique to interpret CNN classification task and justify the classification result with visual explanation and visual search. The model consists of two sub networks: a deep recurrent neural network for generating textual justification and a deep convolutional network for image analysis. This multimodal approach generates the textual justification about the classification decision. To verify the textual justification, we use the visual search to extract the similar content from the training set.

We evaluate our strategy on a novel CUB dataset with the ground-truth attributes. We make use of these attributes to further strengthen the justification by providing the attributes of images.

Short talk

Explainable Machine Learning-based Artificial Intelligence (June 11)

Explaining Visual Classification using Attributes

The performance of deep Convolutional Neural Networks (CNN) has been reaching or even exceeding the human level on large number of tasks. Some examples are image classification, Mastering Go game, speech understanding etc. However, their lack of decomposability into intuitive and understandable components make them hard to interpret, i.e. no information is provided about what makes them arrive at their prediction.

We propose a technique to interpret CNN classification task and justify the classification result with visual explanation and visual search. The model consists of two sub networks: a deep recurrent neural network for generating textual justification and a deep convolutional network for image analysis. This multimodal approach generates the textual justification about the classification decision. To verify the textual justification, we use the visual search to extract the similar content from the training set.

We evaluate our strategy on a novel CUB dataset with the ground-truth attributes. We make use of these attributes to further strengthen the justification by providing the attributes of images.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

Trust through explainability: Technical and legal perspective

Explainability of an AI system is needed to build user's trust. However, explainability is not a feature that could be added to existing AI black-box system. We claim that AI systems have to be build explainable by design. To achieve this goal, they should be designed as hybrid systems, where the machine learning component is integrated with a knowledge-based component. We demonstrate how we achieved it in the area of context-aware systems, where we proposed a knowledge-driven human-computer interaction process of context mediation. Furthermore, trust and explainability cannot be addressed only on the technical level. In our interdisciplinary work on the intersection of AI and law, we consider the legal notion of liability. We claim that the analysis of legal liability is needed for building trust to AI systems. We analyze how it can be applied to AI systems, as it is plays crucial role in certain application areas. Moreover, we emphasize that explainability of AI system should be in fact a requirement from the legal point of view.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

Explainable Machine Learning based on Instances

Example-based explanation methods select particular instances of the dataset to explain the behavior of machine learning models or to explain the underlying data distribution. That is, once the model is build, it is intended to be explained based on instances whose information has been used to buid it. For instance, a training instance is called influential when its deletion from the training data significantly changes the parameters or predictions of the model.

Implicitly, some machine learning methods work example-based. Support Vector Machines look for those instances (vectors) that define the frontier (hyperplane) between two different classes. Given a new non-labeled instance, Knn methods locates the k closest labeled instances in the training set to predict class for the new instance. Thus, it is possible to explain this machine learning approaches using the relevant instances. In fact, it has been shown that example-based explanations performe significantly better than feature-based explanation in order to help the user to understands the reasons behind a predictions, to provide the user with relevant information, to increase the confidenciability of the users, etc.

Here we discuss the need increase the resources to build new Explainable Machine Learning methods based on Instances, where the focus of the development of the method is on the interpretability based on examples.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

Explainable Machine Learning based on Instances

Example-based explanation methods select particular instances of the dataset to explain the behavior of machine learning models or to explain the underlying data distribution. That is, once the model is build, it is intended to be explained based on instances whose information has been used to buid it. For instance, a training instance is called influential when its deletion from the training data significantly changes the parameters or predictions of the model.

Implicitly, some machine learning methods work example-based. Support Vector Machines look for those instances (vectors) that define the frontier (hyperplane) between two different classes. Given a new non-labeled instance, Knn methods locates the k closest labeled instances in the training set to predict class for the new instance. Thus, it is possible to explain this machine learning approaches using the relevant instances. In fact, it has been shown that example-based explanations performe significantly better than feature-based explanation in order to help the user to understands the reasons behind a predictions, to provide the user with relevant information, to increase the confidenciability of the users, etc.

Here we discuss the need increase the resources to build new Explainable Machine Learning methods based on Instances, where the focus of the development of the method is on the interpretability based on examples.

Short talk

Explainable Machine Learning-based Artificial Intelligence (June 11)

Experts-based Recommendation System for Explainable Machine Learning methods in Data Science projects

In this poster, we discuss an interdisciplinary, open educational resource to provide help for Data Science researchers and practitioners looking for Explainable Machine Learning methods.

A significant set of success cases studies of Explainable Machine Learning will be collected and organized. Related publications, corresponding code, etc will be included. A free resource for researchers and practitioners to find and follow the latest state-of-the-art Explainable Machine Learning methods will be created.

In addition, a recommendation system based on experts opinions will be developed using the variety of information previously collected. Given an input information regarding the problem under consideration (as complete as possible) the system will look for the most similar explainable solutions and provides a guide for the resarcher.

In our opinion, to develop this project, an online community of data scientists, machine learners and experts on a number of application domains should be involved as part of a CHIST-ERA project.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

Explaining personalisation for a happier life: Recommender systems for wellbeing and leisure

This talk introduces the fundamentals of recommender systems as a data-driven AI tool for driving personalised user experiences. We then move onto the 'not-so-conventional' personalisation domains beyond e-commenrce, namely health, wellbeing, leisure and tourism in cities, highlighting state-of-the-art and ongoing challenges. A discussion on how explainability can be used to motivate personalised and more convincing recommendations for the end user.

Short talk or poster (to be defined)

Explainable Machine Learning-based Artificial Intelligence (June 11)

From explaining models to explaining decisions and systems

Explainability has been investigated in several ways in the field of machine learning: there are more interpretable models (e.g., decision trees) and more accurate models (e.g., deep networks), and one can try to explain the behavior of even complex models in a more understandable way.

However, when embedded in large AI systems, explainability is much less well studied. Even if we can explain the behavior of a predictive model, we may fail in explaining the actions which a system takes or recommends. Still, understanding actions forms a large part of what humans expect from explanations by AI, e.g., when researchers in collaborative projects need to decide on a next action, when patients want to understand the possible treatments, or when data subjects want to understand the effects of privacy agreements.

I will suggest a number of ideas for research towards AI-based explainability of actions (or more generally policies) of systems exploiting artificial and human intelligence.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

Needs of explainable AI in global healthcare challenges

The focus of this talk is today’s challenges of Artificial Intelligence in Medicine (AIM) and the need of explainability to support the global strategies recently defined by international healthcare authorities.

From a machine learning perspective, the support of multidisciplinary medical teams in such global healthcare problems imply the integration of: (1) a myriad of clinical data sources; and (2) knowledge from multiple levels of the healthcare administration.

We claim that the trust on AIM is the baseline of successful decision support systems in real clinical settings. Indeed, learned AIM models can be trustworthy when they have the validation of a clinical team. However, due to the complexity and the variety of clinicians involved in these scenarios, we believe that a formal research on explainable AIM is required to build trust mechanisms from a technical point of view.

In particular, following the WHO’s recommendations, the EU is implementing the European ONE-health action plan, drawing their attention to global antimicrobial resistance. We show our experience in developing a clinical decision support system for antimicrobial stewardship medical teams and its evaluation in 9 hospitals. We identify current needs, technical requirements to scale AIM systems and the need of explainability.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

Bias and Discrimination in AI – Towards more transparent and explainable attribute-sensitive decisions

With the widespread and pervasive use of AI for automated decision-making systems, AI bias is becoming more apparent and problematic. One of its negative consequences is discrimination: the unfair, unequal, or simply different treatment of individuals based on certain characteristics. However, the relationship between bias and discrimination is still unclear. In this talk, I will discuss current research we are conducting under the frame of an EPSRC-funded project about bias and discrimination in AI from an interdisciplinary perspective that embeds technical, legal, social and ethical dimensions. I will show that finding solutions to bias and discrimination in AI requires robust cross-disciplinary collaborations that will advance on the task of making AI more transparent and explainable to help assess whether AI systems discriminate against users and how to mitigate that.

Keynote talk

Explainable Machine Learning-based Artificial Intelligence (June 11)

Visualisation and crowdsourcing tools for quantitative studies

Data-driven environmental sustainability should consider not just getting more data, but smarter representation of data. Today's research data handling and publication processes are still largely decoupled. Static datasets are published somewhere, referenced from papers, sometimes updated and then out of sync with the publication. Involving non-researchers in such a data representation is not trivial due to a lack of attractiveness, but would be important to benefit from a larger pool of domain knowledge, and would contribute to make science more attractive overall.

In our work at the Service Prototyping Lab at Zurich University of Applied Sciences, we have gathered some experience with building collaborative dashboards including annotations of interesting data positions, hyperlinking updated data from publications, and conditional formulation of findings to co-evolve with continuously updated data. This short talk aims to inspire raising the barrier for research data handling for future projects to help that the current problem - not enough data about the environment - does not contribute to the next problem - too much data.

Short talk

Novel Computational Approaches for Environmental Sustainability (June 12)

Smart Spaces Lab infrastructure

We highlight the infrastructure available in our Smart Spaces lab, the projects which have been run with it, as well as other related resources available in our research group (http://ie.cs.mdx.ac.uk/)

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

The Need to Empirically Evaluate Explanation Quality

Organisations face growing legal and social responsibilities to be able to explain decisions they have made using autonomous systems. Though there is much focus on how these decisions impact the public, there is also a need for these decisions to be clear and interpretable internally for employees. In many sectors, this means provisioning textual explanations around decisions made with technical or expertise-driven information in such a way that non-expert users can understand, thus supporting problem-solving in real-time. As an example, our current work with a telecommunications organisation is centred on empowering desk-based agents to better understand autonomous decision-making using specialist field-engineer notes. In this domain we have implemented various low-level (word-matching between problem and solution, confidence metrics) high-level (summarisation of similarities/differences) and co-created (hazard identification) textual explanation methods.

Increasingly we face difficulties in empirically evaluating the quality of these explanations; a problem which becomes even more challenging as the complexity of the provisioned explanation grows. Though we can easily examine whether an explanation contains the necessary content, it is more difficult to determine whether this content is placed in a suitable context to answer the user’s need for an explanation (e.g. its subjective quality). In this talk we will discuss our current work on eXplainable AI (XAI) and position it with the state-of-the-art by examining the output of several national and international workshops on the subject. In particular, we will highlight an important gap in current XAI research; the ability to empirically evaluate the quality of an explanation. We will present our findings in this domain and why we believe that empirical evaluation of explanation quality is key for the growth of XAI methods in future.

Short talk

Explainable Machine Learning-based Artificial Intelligence (June 11)

Explainable AI in process of complexity reduction of DL models and in boosting theirs performance

The performance of AI systems based on deep learning models is exceeding the human level on an increasing number of tasks like image classification and segmentation, object detection, sentiment analysis, speech recognition, language translation or game playing. Deep learning models don’t need feature extractors, they are applied in a black box manner, no information is provided about what exactly makes them arrive at their predictions.

Our experience in reducing DL and ML models (because of the size and adaptation them to real time systems) help to understand (sensitivity analysis - based on a gradient, heat map computation, conditionality reduction) how specific layer is sensitive, which filters have low contribution in activation or are correlated together, which region or subset of weights can be removed from a layer, or how input data variations have influence on prediction accuracy and categories distinction for further model shrinking. It enables to acquire knowledge about model ability to generalize each category. Recently we can also observe changes in context DL models. LSTM and GRU cells in recurrent networks are exchange with less complex different types of attention mechanism. There is no answer if we can exchange them in most of the task and can achieve the same accuracies like in RNNs. Additionally, there is no explicit rule what combination of attention mechanism to use for specific task.

Most of the state-of-the-art rules extraction techniques are often designed for specific network architectures and specific domains, and therefore not easily adaptable to new applications. With the emergence of Deep Learning, rules extraction and interpret-ability methods that can work for networks with thousands of neurons are in high demand. We have already implemented some memetic approaches and RL methods for reducing complexity of the DL networks by acquiring the knowledge before this process. Our further goal is to use GA and reinforcement leaning combined with other ML methods to acquire knowledge of the AI model mechanism to do not only reduction of a model, but also improvement in accuracy verification of the system in a specific task. The specific goal to explain will be defined as a combination of fitness or rewards form.

Presented approaches will help to better understand how the prediction is made and which parts and regions take main part in this process. It will give more control on a prediction process. Further it will help to shrink the model by removing some less important parts of huge learning models in the specific tasks and can improve the model by changing its architecture and adapt it to specific task and input data. Another goal is to find using genetic and reinforcement learning techniques new methods or rules how DL models deal and make decision in specific tasks. In our presentation existing already implemented approaches will be described and new methods which we want to include to extract more hidden information from DL models in wide range of specific tasks.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

Explainable AI in process of complexity reduction of DL models and in boosting theirs performance

The performance of AI systems based on deep learning models is exceeding the human level on an increasing number of tasks like image classification and segmentation, object detection, sentiment analysis, speech recognition, language translation or game playing. Deep learning models don’t need feature extractors, they are applied in a black box manner, no information is provided about what exactly makes them arrive at their predictions.

Our experience in reducing DL and ML models (because of the size and adaptation them to real time systems) help to understand (sensitivity analysis - based on a gradient, heat map computation, conditionality reduction) how specific layer is sensitive, which filters have low contribution in activation or are correlated together, which region or subset of weights can be removed from a layer, or how input data variations have influence on prediction accuracy and categories distinction for further model shrinking. It enables to acquire knowledge about model ability to generalize each category. Recently we can also observe changes in context DL models. LSTM and GRU cells in recurrent networks are exchange with less complex different types of attention mechanism. There is no answer if we can exchange them in most of the task and can achieve the same accuracies like in RNNs. Additionally, there is no explicit rule what combination of attention mechanism to use for specific task.

Most of the state-of-the-art rules extraction techniques are often designed for specific network architectures and specific domains, and therefore not easily adaptable to new applications. With the emergence of Deep Learning, rules extraction and interpret-ability methods that can work for networks with thousands of neurons are in high demand. We have already implemented some memetic approaches and RL methods for reducing complexity of the DL networks by acquiring the knowledge before this process. Our further goal is to use GA and reinforcement leaning combined with other ML methods to acquire knowledge of the AI model mechanism to do not only reduction of a model, but also improvement in accuracy verification of the system in a specific task. The specific goal to explain will be defined as a combination of fitness or rewards form.

Presented approaches will help to better understand how the prediction is made and which parts and regions take main part in this process. It will give more control on a prediction process. Further it will help to shrink the model by removing some less important parts of huge learning models in the specific tasks and can improve the model by changing its architecture and adapt it to specific task and input data. Another goal is to find using genetic and reinforcement learning techniques new methods or rules how DL models deal and make decision in specific tasks. In our presentation existing already implemented approaches will be described and new methods which we want to include to extract more hidden information from DL models in wide range of specific tasks.

Short talk

Explainable Machine Learning-based Artificial Intelligence (June 11)

Causal-AI: Explainability of AI Models through Cause and Effect Reasoning

Interpretability of artificial intelligence (AI) models is one of the most discussed topics in contemporary AI research (Holm, 2019). Leading architects of AI, like Turing Award winner Judea Pearl are very critical with the current machine learning (ML) concentration on (purely data-driven) deep learning and its non-transparent structures (Ford, 2018). "These and other critical views regarding different aspects of the machine learning toolbox, however, are not a matter of speculation or personal taste, but a product of mathematical analyses concerning the intrinsic limitations of data-centric systems that are not guided by explicit models of reality" (AAAI-WHY 2019). In order to achieve a human-like AI, it is necessary to tell the AI how humans come up with decisions, how they plan and how they imagine things. Humans do that through causal reasoning (Pearl & Mackenzie, 2018). Therefore, in this talk (and project proposal), we will focus on aspects for integrating causal inference wit h machine learning, stimulated, among others, by Pearl's New Science of Cause and Effect, in order to come up with know-how that is complementary to the current deep learning expertise.

Specifically, based on the Software Competence Center Hagenberg's (SCCH) experience of carrying out AI-related research projects together with industry partners, the following research topics are relevant from an industrial point of view:

  • Learning causal models from industrial data sets with applications for, e.g., imputation of missing data based on causal inference
  • Extraction and generation of causal models from knowledge graphs and large heterogeneous and unstructured data sets, e.g. for identifying cause-effect relationships of system failures from system logs and development artifacts (code, architecture/requirements/test specifications)
  • Research on potential integration of several causal models to create comprehensive domain knowledge models
Short talk or poster (to be defined)

Explainable Machine Learning-based Artificial Intelligence (June 11)

Machine Intelligence for Smart Water System

The freshwater resources have been constantly depleting worldwide and it is forecasted that several countries will face acute water shortage in coming few years. The alarming situation, also highlighted in a recent United Nations study, “Global Water Crisis: The Facts”, calls for an efficient management of water distribution systems in cities and towns. It is really surprising to see that the issue of water management is continuously ignored across, and most of the cry is about water scarcity.

With an increasing interest in Artificial Intelligence based scientific applications, we suggest a “Smart Water System”, which fully exploits state-of-art Artificial Intelligence and Machine Learning technologies. The system would provide an Information Management System for collecting, storing, and monitoring of water system related data including water sources (e.g. reservoir) data, pipe network data, customer service data and business data (e.g. billing). The data would be processed by an “Artificial Intelligence Engine” to evaluate and optimize specific business operations of a water supply system.

Research empowered by machine intelligence in the field of smart water distribution systems would lead to development of data-driven computational prototypes for addressing the following research challenges:

  • On-line Simulation of Water Distribution Network
  • Water Demand Prediction
  • Operational Optimization
  • Leakage/Pipe-Burst Detection and Localization
  • Pressure and Flow Sensors Placement Optimization
Short talk

Novel Computational Approaches for Environmental Sustainability (June 12)

Operning the Black Box? - The European Legal Framework

Explainable AI (XAI) is not only relevant from the perspective of developers who want to understand how their system or model is working in order to debug or improve it. XAI is also a LEGAL ISSUE: For those affected by an algorithmic decision, it is important to comprehend why the system arrived at this decision in order to understand the decision, develop trust in the technology and - if the algorithmic decision making process is illegal - initiate appropriate remedies against it. Last but not least, XAI enables experts (and regulators) to audit decisions and verify whether legal regulatory standards have been complied with. All these arguments strike in favor for OPENING THE BLACK BOX. On the other hand, there are a number of legal arguments against full transparency of AI systems, esp. the interest to protect trade secrets, national security, and privacy.

Against this background, I will try to explore the European legal framework for XAI in my short talk.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

History-Aware Explainability in Self-adaptation using Temporal Models

On the one hand, there has been a growing interest towards the application of AI-based for self-adaptation under uncertainty. On the other hand, self-explanation is one of the self-* properties that has been neglected. This is paradoxical as self-explanation is inevitably needed when using such techniques. We argue that a self-adaptive system (SAS) needs an infrastructure and capabilities to look at its own history to explain and reason why the system has reached its current state. Such an infrastructure and capabilities need to be built based on the right conceptual models in such a way that the system's history can be stored, queried to be used in the context of the decision-making algorithms. We framed explanation capabilities in four architectural incremental levels, from forensic self-explanation to automated history-aware (HA) systems. Incremental capabilities imply that capabilities at level n should be available for capabilities at level n+1. The poster shows results for the first two levels using temporal graph-based models in the domain of Bayesian learning. Future work is also outlined.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)

History-Aware Explainability in Self-adaptation using Temporal Models

On the one hand, there has been a growing interest towards the application of AI-based for self-adaptation under uncertainty. On the other hand, self-explanation is one of the self-* properties that has been neglected. This is paradoxical as self-explanation is inevitably needed when using such techniques. We argue that a self-adaptive system (SAS) needs an infrastructure and capabilities to look at its own history to explain and reason why the system has reached its current state. Such an infrastructure and capabilities need to be built based on the right conceptual models in such a way that the system's history can be stored, queried to be used in the context of the decision-making algorithms. We framed explanation capabilities in four architectural incremental levels, from forensic self-explanation to automated history-aware (HA) systems. Incremental capabilities imply that capabilities at level n should be available for capabilities at level n+1. We will discuss early but reassuring results for the first two levels using temporal graph-based models in the domain of Bayesian learning. Future work is also outlined looking for collaborations.

Short talk or poster (to be defined)

Explainable Machine Learning-based Artificial Intelligence (June 11)

Quality of data for computer vision algorithm

Focus on cognitive multimedia processing, open challenges and standard:

  1. Collecting "good" data for AI using AI
  2. Qualifying AI based computer vision in real life scenario
Short talk or poster (to be defined)

Explainable Machine Learning-based Artificial Intelligence (June 11)

Using a combination of numerical process models and data to make decisions in the environment

Environmental science has developed to the stage where there are simulators (numerical models), often involving the solutions of PDEs, that give good representations of the real world. In addition our data collection capabilities have increased in recent years so we can now collect very large amounts of data on the natural world. How can we best combine both data and models, both of which are uncertain, to make good decisions about environmental policy? In this paper I will look at ways that machine learning and statistical modelling can be used to bring uncertain data and models together to make better decisions. Examples will include the calibration of climate and other environmental models, modelling air pollution and climate prediction for future emission scenarios not covered by the RCPs.

Keynote talk

Novel Computational Approaches for Environmental Sustainability (June 12)

Integrated modelling approaches to advance in the assessment of the impacts of plant protection products

Plant protection is a vital part of current agricultural and horticultural practices assuring yield and quality. Application of agrochemicals for plant protection requires dedicated practices such as spraying and seed treatments.

Sustainable plant protection required minimizing environmental risks associated with drift of agrochemicals during field operations. Mitigation measures to reduce the risk to the environment include buffer zones and drift reduction technologies. The acceptance and range of measures vary widely with limited harmonization. To assist development of more uniform measures, implement effective practices and assess new application technologies, computational modelling provides comprehensive and objective insight into the drift process affected by operational, environmental and field factors.

Challenges to overcome to achieve more reliable and effective modelling frameworks of drift include improved models of the interactions of particles/droplets with canopy and soil structures affected by environmental conditions, the temporal and spatial scales affecting dispersion of particles, droplets and vapor, and integrating into the models the properties and dynamics of application technology and operations, and impact on plants, humans, animals and ecosystems. Computational Fluid Dynamics provides a means to implement such framework. Still, the multiscale nature of the drift process requires to build dedicated, more efficient and user friendly simulation platforms that solve and integrate the models into predictive tools to support drift risk assessment.

Short talk

Novel Computational Approaches for Environmental Sustainability (June 12)

Scalable Constraint-based Optimisation

Declarative methods for combinatorial optimisation (such as modeling as a CSP) can form the basis of highly scalable solvers. These may be used in several application contexts, some of which may combine with machine-learning techniques. Domain applications include natural resource management.

Short talk or poster (to be defined)

Novel Computational Approaches for Environmental Sustainability (June 12)

Design Space Exploration of Deep Learning Models for embedded vision and health devices

Deep Neural Networks (DNNs) suffer from energy-hungry implementation due to their computation and memory intensive processing patterns. Health wearable devices and embedded systems have small and low power processors which are much slower compared to desktop and server processors. To overcome this problem, in our research project we propose an automatic framework that designs a highly optimized and energy efficient set of deep learning architectures consist of Convolutional Neural Network (CNNs) and Recurrent Neural Networks (RNNs) for health and vision applications. The proposed framework provides efficient methods to explore and prune the design space to find improved set of neural network architectures. Application of these methods cover different health applications, like cardiac arrhythmia monitoring, which is regarded as a key point of the distributed healthcare devices. Other applications can be activity monitoring by wearable sensor devices for elderly people, and fall detection system.

Short talk or poster (to be defined)

Explainable Machine Learning-based Artificial Intelligence (June 11)

Graph-representation and learning framework for Smart Cities big data analytics

Given the increased dynamism and complexity of modern world, researchers struggle to cope to exploit as much as possible meaningful knowledge from the rapidly growing abundance of available data. The necessity for efficient representation of various interdependences among huge amounts of heterogeneous data, is emerging to be one of the most challenging tasks.

We propose development of a graph-based representation and accordingly learning framework for more efficient capturing of spatial and temporal dependences among datastreams. In fact, graphs enable richer and more effective representation of data and their relations giving more meaning to the available datasets, enabling in turn better efficiency in entire machine learning process and its application (including feature extraction and selection, classification, prediction etc.).

In particular, we consider Smart Grids, especially energy consumption as well as renewables generation forecasting. Improved accuracy of predictions in this field directly leads to optimization of Demand Side Management for all customers as well as assessment of renewables and EV insertion in specific areas leading to remarkable improvements in sustainability of cities. Integration of various heterogeneous energy consumption related data (i.e., electricity, gas and water consumption) and efficient capturing of their interdependences would be the very next step towards sustainability boosting urban development recommendation systems.

Short talk

Novel Computational Approaches for Environmental Sustainability (June 12)

Machine Learning Enabled Transparent Manufacturing

Benefiting from the advancement of information and communication technology, more and more data related to product manufacturing process can be collected. Owing to long manufacturing cycle, dynamic manufacturing process and diverse sensors, the collected data featured by high volume, high velocity and high variety generate a typical big data pool. Therefore, a rapidly increasing number of big data analytics research on production application emerges. Accordingly various knowledge patterns underlid the data are recognized to support production decision-making activities. On one hand, the data analytics results accelerate the decision-making efficiently. On the other hand, the investigation on physics of manufacturing process are slacked up. Actually they do not collide with each other, but complement each other. Hence, a machine learning enabled transparent manufacturing framework is proposed.

Because the energy consumption of manufacturing equipments such as CNC machines is directly related to the workpiece, cutters and cutting parameters, it is selected as an indirect indicator to monitor the manufacturing process due to the cost-effective power sensors. By applying machine learning algorithms, the energy consumption patterns which corresponding to multifarious combinations of machines, cutters and cutting parameters can be recognized from the monitoring power data. Since cutters wear out gradually along with the consecutive manufacturing process, the energy consumption of producing the identical components with the same cutting parameters are different. Therefore the energy consumption for individual combination obtained from the collected data fluctuate within an interval. In contrast, the energy consumption of CNC machines can also be calculated based on the physics of cutting process which are expressed as the integral of the product of cutting force and cutting speed over cutting time. The accuracy of cutting force estimation heavily relies on the cutting force coefficients which are related to the tool wear condition. Generally the cutting force coefficients are calculated based on very limited number of physical experiments, a single value for each coefficient are applied for cutting force estimation. Apparently it cannot reflect the real cutting process. Therefore, their conjunction can explain the manufacturing process in more accurate way and make the manufacturing process transparent.

Since the energy consumption models based on collecting data and cutting physics have their own pros and cons, a hybrid models based decision making engine are expect to be developed to support production activities efficiently within transparent manufacturing scenario.

Poster

Explainable Machine Learning-based Artificial Intelligence (June 11)