Periodic Reporting for period 1 - GECKO (building GrEener and more sustainable soCieties by filling the Knowledge gap in social science and engineering to enable responsible artificial intelligence co-creatiOn)
Berichtszeitraum: 2021-01-01 bis 2022-12-31
There is a gap in studying responsible AI methods for sustainability, overarching social and information sciences. Current connected home technology (CHT) does not provide actionable outputs towards low energy outcome of commercial value, due to the absence of AI that deeply embeds users’ values and captures the complexity of lives at home. Responsible AI research is still in its infancy partly due to discipline barriers, rooted in orthogonal goals of AI and social science research, where, AI aims to build models to replace human reasoning, while social science aims to inform or guide human reasoning. GECKO aims to fill this gap by training the next generation of researchers and challenge traditional research approaches, acknowledging that socio-technical solutions must recognise complex interrelations between people and technology: addressing urgent sustainability needs, where a successful responsible AI decision support technology must embed social science understanding of people’s actions and how they interact with technology. The key research objectives of GECKO are to develop conceptual information science, responsible AI-driven methods to underpin design of technological solutions, investigate how in-depth understanding of everyday life can support the design, operation and energy-saving outcomes of emerging AI-driven low-carbon home technologies, how it can steer development of new human-centered information sciences and supporting development of novel responsible AI-driven and application-relevant information science methods.
The Information Sciences & Engineering Team has focused on developing a range of AI models that facilitate trust and explainability. Specifically, research has focused on exploring how to evaluate the explanations of deep learning models for time series prediction tasks, ensuring trust in algorithms through evaluation of generalisability and transferability of AI models, incorporating human-in-the-loop via active learning to address data uncertainty, development of adaptable deep learning structures to the environmental and living conditions, and detecting anomalies in the data streams. These were demonstrated for residential buildings and farms, to support governments globally to meet their NetZero Energy targets and taking into account renewables and emergence of electric vehicles.
The Social Science team has explored literature on CHTs, social practices, social justice and design methodologies relating to the development of human-centered AI. Data collection and fieldwork (inc. 2 living labs in eco-conscious residential buildings in Norway and a remote community in Aran Islands) has included: household interviews; a qual-quant study on energy impacts of new-build smart homes; a qualitative study of smart home designers’ user imaginaries; a desk-based mapping of errors and resistances in CHT applications; interviews with energy professionals on CHT and energy vulnerability; co-design workshops with smart energy professionals on domestic imaginaries; and piloting ‘walk with video’ methods on sustainable communities.
The Data Interpretation team has studied how algorithms can involve humans in the loop to provide explainability using intelligent personal assistants aiming to affect energy behaviour, creating novel design concepts for AI assistants. A living lab platform has been built (13 households) giving the users opportunity to monitor their energy consumption and get feedback. A scalable big data decentralised energy disaggregation scheme is being implemented, which alleviates the drawbacks of central data processing and promotes the adoption of such solutions. Optimisation methods under Demand Response schemes have been reviewed as a first step towards the determination of residential optimal consumption patterns.
The Social Science team has explored literature on CHTs, social practices, social justice and design methodologies relating to the development of human-centered AI. Data collection and fieldwork (inc. 2 living labs in eco-conscious residential buildings in Norway and a remote community in Aran Islands) has included: household interviews; a qual-quant study on energy impacts of new-build smart homes; a qualitative study of smart home designers’ user imaginaries; a desk-based mapping of errors and resistances in CHT applications; interviews with energy professionals on CHT and energy vulnerability; co-design workshops with smart energy professionals on domestic imaginaries; and piloting ‘walk with video’ methods on sustainable communities.
The Data Interpretation team has studied how algorithms can involve humans in the loop to provide explainability using intelligent personal assistants aiming to affect energy behaviour, creating novel design concepts for AI assistants. A living lab platform has been built (13 households) giving the users opportunity to monitor their energy consumption and get feedback. A scalable big data decentralised energy disaggregation scheme is being implemented, which alleviates the drawbacks of central data processing and promotes the adoption of such solutions. Optimisation methods under Demand Response schemes have been reviewed as a first step towards the determination of residential optimal consumption patterns.
The GECKO Team, as evidenced by academic publications and released public datasets, enabled by various training events, have made the following contributions so far:
The Information Science and Engineering Team:
- developed a novel framework for evaluating mathematical interpretability of deep 'black box' ML models and demonstrated how these can be used to determine the mechanism by which knowledge from one ML model can be effectively passed to another via knowledge distillation
- performed rigorous evaluation of what information can be mined at different smart meter sampling rates and how meaningful and trustworthy deep learning and domain-specific metrics are in CHT applications in ensuring ‘responsible’ and ‘ethical’, ‘bias-free’ outcomes
- demonstrated the challenges of transferability of non-intrusive load disaggregation to achieve net zero in agriculture by quantifying electricity consumption of various farming automation tools
- released a transformer-based deep learning software for load disaggregation, by processing entire sequences of data, understanding the significance of each part of the input sequence and assigning importance weights accordingly, to learn global dependencies in the sequence
-improved feature selection techniques for personal thermal comfort modelling
- tackle time series anomaly detection in smart homes
Social Science Team:
-Empirically, results have been achieved through in-depth qualitative methods such as interviews, workshops and observation with diverse communities including prosumers, smart home occupants, industry professionals, and other stakeholders. Major empirical advances have focussed on engaging diverse and often marginalised groups (e.g. island communities or energy vulnerable groups), and focussing on diverse CHTs across spatial scales rather than solely in the home.
-Methodologically, innovation has been demonstrated through detailed case study research on situated practices (e.g. in specific organisations or communities), with a focus on how practices are connected across scales (e.g. how domestic life connects with design imaginaries). Novel methodologies such as co-design and ‘walk with video’ methods have also been developed to engage and empower diverse groups.
-Conceptually, progress has been demonstrated through developments of social practice-based approaches, such as tracing connections between expert imaginaries and domestic practices, and combining practice theories with learning theory and design approaches. Understandings of ART AI and CHT have also been advanced through the development of more systemic and anticipatory approaches to social justice based on intersectionality.
-Expected results for the rest of the research centre on deepening these advances, and translating them into recommendations for improved design of CHTs, as well as more inclusive approaches to design that engage and empower diverse user groups.
Data Interpretation team:
-Demonstrated the challenges and needs involving smart users in the loop and optimising the AI/ML-models performance by reviewing literature and interviewing smart home users
-Demonstrated explainability for individual decisions by ML methods in different domains including energy demand forecasting in CHT, e-commerce, and personal thermal comfort prediction towards generating user-centered explanations.
-Implemented Graph-based neural networks that demonstrated high performance for load disaggregation providing trustworthy decisions.
-Demonstrated Deep Reinforcement Learning and Particle Swarm Optimization methods in residential Distributed Energy Resources scheduling and control.
The Information Science and Engineering Team:
- developed a novel framework for evaluating mathematical interpretability of deep 'black box' ML models and demonstrated how these can be used to determine the mechanism by which knowledge from one ML model can be effectively passed to another via knowledge distillation
- performed rigorous evaluation of what information can be mined at different smart meter sampling rates and how meaningful and trustworthy deep learning and domain-specific metrics are in CHT applications in ensuring ‘responsible’ and ‘ethical’, ‘bias-free’ outcomes
- demonstrated the challenges of transferability of non-intrusive load disaggregation to achieve net zero in agriculture by quantifying electricity consumption of various farming automation tools
- released a transformer-based deep learning software for load disaggregation, by processing entire sequences of data, understanding the significance of each part of the input sequence and assigning importance weights accordingly, to learn global dependencies in the sequence
-improved feature selection techniques for personal thermal comfort modelling
- tackle time series anomaly detection in smart homes
Social Science Team:
-Empirically, results have been achieved through in-depth qualitative methods such as interviews, workshops and observation with diverse communities including prosumers, smart home occupants, industry professionals, and other stakeholders. Major empirical advances have focussed on engaging diverse and often marginalised groups (e.g. island communities or energy vulnerable groups), and focussing on diverse CHTs across spatial scales rather than solely in the home.
-Methodologically, innovation has been demonstrated through detailed case study research on situated practices (e.g. in specific organisations or communities), with a focus on how practices are connected across scales (e.g. how domestic life connects with design imaginaries). Novel methodologies such as co-design and ‘walk with video’ methods have also been developed to engage and empower diverse groups.
-Conceptually, progress has been demonstrated through developments of social practice-based approaches, such as tracing connections between expert imaginaries and domestic practices, and combining practice theories with learning theory and design approaches. Understandings of ART AI and CHT have also been advanced through the development of more systemic and anticipatory approaches to social justice based on intersectionality.
-Expected results for the rest of the research centre on deepening these advances, and translating them into recommendations for improved design of CHTs, as well as more inclusive approaches to design that engage and empower diverse user groups.
Data Interpretation team:
-Demonstrated the challenges and needs involving smart users in the loop and optimising the AI/ML-models performance by reviewing literature and interviewing smart home users
-Demonstrated explainability for individual decisions by ML methods in different domains including energy demand forecasting in CHT, e-commerce, and personal thermal comfort prediction towards generating user-centered explanations.
-Implemented Graph-based neural networks that demonstrated high performance for load disaggregation providing trustworthy decisions.
-Demonstrated Deep Reinforcement Learning and Particle Swarm Optimization methods in residential Distributed Energy Resources scheduling and control.