The self as agent-environment nexus: crossing disciplinary boundaries to help human selves and anticipate artificial selves
Lead Research Organisation:
University College London
Department Name: Institute of Neurology
Abstract
Both human and artificial agents learn from and potentially change their environments, but also align and adapt to them. Examples abound, from humans unconsciously tapping feet to the rhythm of music to a cleaning robot keeping a house dust-free. In humans, however, brain-world alignment is also closely linked with mental features like perception or self and mental health disorders lead to profound disruptions in how one considers 'the self'.
As AI agents acquire human and superhuman skill levels and Artificial General Intelligence looms closer, ethical considerations about how an agent will interact with its environment become paramount. What is it about an AI agent that will ensure it can co-exist peacefully in human society and transfer skills seamlessly from one setting to another?
Let us consider a real-world example of an adolescent who is watching television (on their laptop) via a streaming service. Suppose that individual has displayed anti-social behaviours at school or in their home and suppose that he has streamed several violent movies on his laptop over the past few days. Should the recommender system of the streaming service suggest a 'movie that you might like' that, again, contains strong scenes of violence? Or should the recommender system suggest, instead, a documentary on polar bears? This example illustrates three open questions and challenges in AI research. First, how would the recommender system infer the state of the young man? Second, how would a recommender system adapt its recommendations based upon this state, particularly given that the system itself has its own internal goals - i.e. to motivate people to consume more content? And third, is it ethical to probe the human at the human-computer interface in the first place?
The current proposal aims to converge both "free energy" and "dynamic temporo-spatial" approaches to mental features to develop a mathematical and computational framework for environmental alignment and adaptation of intelligent agents.
The primary, direct goal of this approach is to help human agents suffering from abnormal changes in their perception and self, as in psychiatric conditions like schizophrenia. This not only carries major ethical implications for understanding of ourselves, but will also inform the second, broader, aim of this project: paving the way for an understanding and mathematical theory of artificial selves.
As AI agents acquire human and superhuman skill levels and Artificial General Intelligence looms closer, ethical considerations about how an agent will interact with its environment become paramount. What is it about an AI agent that will ensure it can co-exist peacefully in human society and transfer skills seamlessly from one setting to another?
Let us consider a real-world example of an adolescent who is watching television (on their laptop) via a streaming service. Suppose that individual has displayed anti-social behaviours at school or in their home and suppose that he has streamed several violent movies on his laptop over the past few days. Should the recommender system of the streaming service suggest a 'movie that you might like' that, again, contains strong scenes of violence? Or should the recommender system suggest, instead, a documentary on polar bears? This example illustrates three open questions and challenges in AI research. First, how would the recommender system infer the state of the young man? Second, how would a recommender system adapt its recommendations based upon this state, particularly given that the system itself has its own internal goals - i.e. to motivate people to consume more content? And third, is it ethical to probe the human at the human-computer interface in the first place?
The current proposal aims to converge both "free energy" and "dynamic temporo-spatial" approaches to mental features to develop a mathematical and computational framework for environmental alignment and adaptation of intelligent agents.
The primary, direct goal of this approach is to help human agents suffering from abnormal changes in their perception and self, as in psychiatric conditions like schizophrenia. This not only carries major ethical implications for understanding of ourselves, but will also inform the second, broader, aim of this project: paving the way for an understanding and mathematical theory of artificial selves.
Planned Impact
Our proposal and study exert strong impact in three ways, that is, economic, clinical, and social impact.
Economic impact: Developing a novel mathematical-computational model of agent-environment interaction carries major implications. Such model will furnish an artificial agent that can respond dynamically to - and decide upon - events in an uncertain environmental context. Our model will thus provide the basis for widespread application and development of agent-environment interaction AI in both public and private settings that are characterized by uncertainty and context sensitivity. As our team members, especially in Canada (M. Fraser, J. Griffith, and Panangaden) show strong association and connection to various AI companies like Google and Facebook in Montreal/Toronto, there is a direct pathway for commercialization of our model. That could not only result in novel patents but also partnerships with some of the giants like Google and Facebook as well as the establishment of spin-off and start-up companies in both UK (London, Oxford) and Canada (Toronto, Montreal, Ottawa). To facilitate transfer from knowledge to application - as well as commercialization - we, due to the extensive connections of our team members (especially in Canada), will, from the very beginning, include the demands on AI as prioritized and in developed by companies lie Facebook and Google.
Clinical impact: Yet another economic impact will be in the field of mental health. Better diagnostic classification of psychiatric disorders like schizophrenia and depression will lead to earlier and more appropriate treatment. That, in turn, will significantly reduce economic burden of especially depression which is the most prevalent disorder (including both medical and psychiatric) according to the World Health Organization. That reduction will be first manifest in UK and Canada as we will implement first our mathematical-computational model of agent-environment interaction in specific psychiatric clinics (Royal Ottawa Mental Health Centre in Ottawa), Douglass/McGill in Montreal. If reducing economic burden, the model will be applied globally as for instance in Taiwan (Taipeh), Italy (Milano), China (Hangzhou), and Japan (Toyio) where the Canadian PI (G. Northoff) has extensive long-standing psychiatric affiliation and connections.
Social impact: Our model of agent-environment interaction and its application to psychiatric disorders carries major social impact. This is so as it will re-define what it is to be a human self and how it converges and is distinct from an artificial self. This will be channeled in our ethical debate. To disseminate our ethical knowledge and its impact, we will develop guidelines and criteria for the human self in its encounter with the artificial self, that is, what it means to be humans. This is especially supported by the fact that the Canadian PI, G. Northoff, has extensive expertise and background in neuroethics. He will work with the current team to discuss the ethical implications and, together also with other specialists in ethics, will develop the guidelines and criteria. It shall also be mentioned that G. Northoff is part of the Human Brain Project of the European Union in project No. 11 that, led by K. Evers and JP. Changeux, focuses on philosophical and ethical implications of neuroscience and AI. Hence, this will facilitate knowledge dissemination and enhance the world-wide impact of the ethical part of the project.
Economic impact: Developing a novel mathematical-computational model of agent-environment interaction carries major implications. Such model will furnish an artificial agent that can respond dynamically to - and decide upon - events in an uncertain environmental context. Our model will thus provide the basis for widespread application and development of agent-environment interaction AI in both public and private settings that are characterized by uncertainty and context sensitivity. As our team members, especially in Canada (M. Fraser, J. Griffith, and Panangaden) show strong association and connection to various AI companies like Google and Facebook in Montreal/Toronto, there is a direct pathway for commercialization of our model. That could not only result in novel patents but also partnerships with some of the giants like Google and Facebook as well as the establishment of spin-off and start-up companies in both UK (London, Oxford) and Canada (Toronto, Montreal, Ottawa). To facilitate transfer from knowledge to application - as well as commercialization - we, due to the extensive connections of our team members (especially in Canada), will, from the very beginning, include the demands on AI as prioritized and in developed by companies lie Facebook and Google.
Clinical impact: Yet another economic impact will be in the field of mental health. Better diagnostic classification of psychiatric disorders like schizophrenia and depression will lead to earlier and more appropriate treatment. That, in turn, will significantly reduce economic burden of especially depression which is the most prevalent disorder (including both medical and psychiatric) according to the World Health Organization. That reduction will be first manifest in UK and Canada as we will implement first our mathematical-computational model of agent-environment interaction in specific psychiatric clinics (Royal Ottawa Mental Health Centre in Ottawa), Douglass/McGill in Montreal. If reducing economic burden, the model will be applied globally as for instance in Taiwan (Taipeh), Italy (Milano), China (Hangzhou), and Japan (Toyio) where the Canadian PI (G. Northoff) has extensive long-standing psychiatric affiliation and connections.
Social impact: Our model of agent-environment interaction and its application to psychiatric disorders carries major social impact. This is so as it will re-define what it is to be a human self and how it converges and is distinct from an artificial self. This will be channeled in our ethical debate. To disseminate our ethical knowledge and its impact, we will develop guidelines and criteria for the human self in its encounter with the artificial self, that is, what it means to be humans. This is especially supported by the fact that the Canadian PI, G. Northoff, has extensive expertise and background in neuroethics. He will work with the current team to discuss the ethical implications and, together also with other specialists in ethics, will develop the guidelines and criteria. It shall also be mentioned that G. Northoff is part of the Human Brain Project of the European Union in project No. 11 that, led by K. Evers and JP. Changeux, focuses on philosophical and ethical implications of neuroscience and AI. Hence, this will facilitate knowledge dissemination and enhance the world-wide impact of the ethical part of the project.
Publications
Adams N
(2022)
Neurophysiological consequences of synapse loss in progressive supranuclear palsy
in Brain
Albarracin M
(2023)
Shared Protentions in Multi-Agent Active Inference
Anil Meera A
(2022)
Reclaiming saliency: Rhythmic precision-modulated action and perception.
in Frontiers in neurorobotics
Ansado J
(2021)
How brain imaging provides predictive biomarkers for therapeutic success in the context of virtual reality cognitive training.
in Neuroscience and biobehavioral reviews
Badcock P
(2023)
The separation distress hypothesis: The ultimate theory of depression?
in Neuropsychoanalysis
Barp A
(2022)
Geometry and Statistics
Benrimoh DA
(2020)
All grown up: Computational theories of psychosis, complexity, and progress.
in Journal of abnormal psychology
Bhat A
(2022)
Attenuated transcriptional response to pro-inflammatory cytokines in schizophrenia hiPSC-derived neural progenitor cells.
in Brain, behavior, and immunity
Bhat A
(2020)
Transcriptome-Wide Association Study Reveals Two Genes that Influence Mismatch Negativity
in SSRN Electronic Journal
Bhat A
(2021)
Immunoceptive inference: why are psychiatric disorders and immune responses intertwined?
in Biology & philosophy
Bouziane I
(2022)
Enhanced top-down sensorimotor processing in somatic anxiety.
in Translational psychiatry
Bowie C
(2022)
A 12-month projection to September 2022 of the COVID-19 epidemic in the UK using a dynamic causal model.
in Frontiers in public health
Bowie C
(2022)
Using a Dynamic Causal Model to validate previous predictions and offer a 12-month forecast of the long-term effects of the COVID-19 epidemic in the UK.
in Frontiers in public health
Brouillet D
(2023)
Relative fluency (unfelt vs felt) in active inference.
in Consciousness and cognition
Carhart-Harris RL
(2023)
Canalization and plasticity in psychopathology.
in Neuropharmacology
Castaldo F
(2023)
Multi-modal and multi-model interrogation of large-scale functional brain networks.
in NeuroImage
Ciaunica A
(2022)
I overthink-Therefore I am not: An active inference account of altered sense of self and agency in depersonalisation disorder.
in Consciousness and cognition
Collerton D
(2023)
Understanding visual hallucinations: A new synthesis
in Neuroscience & Biobehavioral Reviews
Constant A
(2022)
Integrating Evolutionary, Cultural, and Computational Psychiatry: A Multilevel Systemic Approach.
in Frontiers in psychiatry
Cooray GK
(2023)
Global dynamics of neural mass models.
in PLoS computational biology
Corcoran A
(2023)
Accelerating scientific progress through Bayesian adversarial collaboration
in Neuron
Da Costa L
(2023)
Reward Maximization Through Discrete Active Inference
in Neural Computation
Davila CE
(2022)
A Control-Theoretical System for Modulating Hippocampal Gamma Oscillations Using Stimulation of the Posterior Cingulate Cortex.
in IEEE transactions on neural systems and rehabilitation engineering : a publication of the IEEE Engineering in Medicine and Biology Society
De Ridder D
(2023)
A parahippocampal-sensory Bayesian vicious circle generates pain or tinnitus: a source-localized EEG study.
in Brain communications
Dezhina Z
(2023)
Establishing brain states in neuroimaging data.
in PLoS computational biology
Engleitner H
(2022)
GeoSPM: Geostatistical parametric mapping for medicine
in Patterns
Fagerholm E
(2023)
Selection entropy: The information hidden within neuronal patterns
in Physical Review Research
Fagerholm ED
(2023)
A primer on entropy in neuroscience.
in Neuroscience and biobehavioral reviews
Fermin ASR
(2022)
An insula hierarchical network architecture for active interoceptive inference.
in Royal Society open science
Fields C
(2023)
Control Flow in Active Inference Systems-Part II: Tensor Networks as General Models of Control Flow
in IEEE Transactions on Molecular, Biological and Multi-Scale Communications
Fields C
(2022)
A free energy principle for generic quantum systems.
in Progress in biophysics and molecular biology
Fields C
(2021)
A free energy principle for generic quantum systems
Fields C
(2023)
Control Flow in Active Inference Systems-Part I: Classical and Quantum Formulations of Active Inference
in IEEE Transactions on Molecular, Biological and Multi-Scale Communications
Fields C
(2022)
The free energy principle induces neuromorphic development
in Neuromorphic Computing and Engineering
Friston K
(2022)
Designing Ecosystems of Intelligence from First Principles
Friston K
(2023)
Computational psychiatry: from synapses to sentience.
in Molecular psychiatry
Friston K
(2023)
A Variational Synthesis of Evolutionary and Developmental Dynamics.
in Entropy (Basel, Switzerland)
Friston K
(2023)
A variational synthesis of evolutionary and developmental dynamics
Friston K
(2023)
Path integrals, particular kinds, and strange things
in Physics of Life Reviews
Friston K
(2022)
Maps and territories, smoke, and mirrors.
in The Behavioral and brain sciences
Friston K
(2023)
The free energy principle made simpler but not too simple
in Physics Reports
Friston KJ
(2020)
Bayesian Dysconnections.
in The American journal of psychiatry
Gandolfi D
(2022)
Emergence of associative learning in a neuromorphic inference network
in Journal of Neural Engineering
Golesorkhi M
(2021)
Temporal hierarchy of intrinsic neural timescales converges with spatial core-periphery organization.
in Communications biology
Golesorkhi M
(2021)
The brain and its time: intrinsic neural timescales are key for input processing.
in Communications biology
Hartwig M
(2022)
How Stress Can Change Our Deepest Preferences: Stress Habituation Explained Using the Free Energy Principle.
in Frontiers in psychology
HipĆ³lito I
(2021)
Markov blankets in the brain.
in Neuroscience and biobehavioral reviews
Horne CM
(2022)
The role of cognitive control in the positive symptoms of psychosis.
in NeuroImage. Clinical
Horne CM
(2021)
Cognitive control network connectivity differentially disrupted in treatment resistant schizophrenia.
in NeuroImage. Clinical
Description | Collaboration with Professor of Psychiatry |
Organisation | University of Kent |
Country | United Kingdom |
Sector | Academic/University |
PI Contribution | Through this award we have extended collaborations on patient recruitment. We have also been able to extend findings for test/retest reliability. |
Collaborator Contribution | Assistance in patient recruitment and NHS ethics. |
Impact | Collaboration with Professor of Psychiatry at KCL and University of Kent offering patient recruitment. |
Start Year | 2020 |