Periodic Reporting for period 1 - ASTOUND (Improving social competences of virtual agents through artificial consciousness based on the Attention Schema Theory)
Période du rapport: 2022-12-01 au 2023-11-30
INTRODUCTION:
Even if state of the art AI can solve some problems as competently as a human, it may not have a broader context to judge the value of the solution, especially when handling new and unexpected situations. Humans leverage their awareness, but what about machines?
Given current technology, it should be possible to build a machine that contains a model of artificial consciousness. That machine would attribute a kind of artificial consciousness to itself and to the people it interacts with. Moreover, it would use that attribution to make predictions about human behavior and to put input data in context. The ASTOUND project will provide an Integrative Approach For Awareness Engineering to establish artificial consciousness in machines. To achieve this goal the project will focus on the following activities:
- Develop an AI architecture for Artificial Consciousness based on the Attention Schema Theory (AST). It is a novel approach to human and social cognition that reconciles several cognitive neuroscience theories of consciousness, stating that subjective awareness results from the construction of an internal model of the “state of attention”.
- Implement the developed architecture into a contextually-aware conversational agent. This is to verify the hypothesis that an artificial consciousness based on AST will unambiguously improve selection of appropriate language registry, adaptation to the interlocutor, and long term coherence.
- Propose new evaluation mechanisms that can be used for evaluating the capabilities of the system, but also to measure consciousness in humans and machines.
OBJECTIVES
The final goal is to pose the basis for effective collaboration between humans and machines, promoting machines from the status of tools to those of (empathic) partners. ASTOUND, more specifically, aims at verifying the hypothesis that providing a virtual agent with artificial consciousness, in the sense of an attention schema, will unambiguously improve its performance in a task of natural language understanding. The innovative consciousness architecture based on the AST will be incorporated into a cross-domain, self-adaptive conversational agent. It will be designed to learn and adapt to new contexts and topics based on user interaction.
Embodying awareness into a conversational agent is expected to deliver a series of consciousness-related features that will bring extra values in terms of user experience and system performance.
Even if state of the art AI can solve some problems as competently as a human, it may not have a broader context to judge the value of the solution, especially when handling new and unexpected situations. Humans leverage their awareness, but what about machines?
Given current technology, it should be possible to build a machine that contains a model of artificial consciousness. That machine would attribute a kind of artificial consciousness to itself and to the people it interacts with. Moreover, it would use that attribution to make predictions about human behavior and to put input data in context. The ASTOUND project will provide an Integrative Approach For Awareness Engineering to establish artificial consciousness in machines. To achieve this goal the project will focus on the following activities:
- Develop an AI architecture for Artificial Consciousness based on the Attention Schema Theory (AST). It is a novel approach to human and social cognition that reconciles several cognitive neuroscience theories of consciousness, stating that subjective awareness results from the construction of an internal model of the “state of attention”.
- Implement the developed architecture into a contextually-aware conversational agent. This is to verify the hypothesis that an artificial consciousness based on AST will unambiguously improve selection of appropriate language registry, adaptation to the interlocutor, and long term coherence.
- Propose new evaluation mechanisms that can be used for evaluating the capabilities of the system, but also to measure consciousness in humans and machines.
OBJECTIVES
The final goal is to pose the basis for effective collaboration between humans and machines, promoting machines from the status of tools to those of (empathic) partners. ASTOUND, more specifically, aims at verifying the hypothesis that providing a virtual agent with artificial consciousness, in the sense of an attention schema, will unambiguously improve its performance in a task of natural language understanding. The innovative consciousness architecture based on the AST will be incorporated into a cross-domain, self-adaptive conversational agent. It will be designed to learn and adapt to new contexts and topics based on user interaction.
Embodying awareness into a conversational agent is expected to deliver a series of consciousness-related features that will bring extra values in terms of user experience and system performance.
SCIENTIFIC THEORY
ASTOUND proposes to build a plausible human-brain architecture for consciousness capable of demonstrating an undeniable added value in terms of human-machine interaction. This will be obtained by combining state of the art deep neural networks models with an Attention Schema.
The Attention Schema Theory (AST) , first conceived by Dr. Michael Graziano (Princeton Neuroscience Institute, part of ASTOUND’s team), explains the brain basis of subjective awareness in a mechanistic and scientifically testable manner. According to the AST, the brain is an information-processing device with the capacity to focus its processing resources more on some signals than on others.
On the other hand, the team at Universidad Politécnica de Madrid (Spain) has recently proposed state-of-the-art modules for conversational systems, as well as automatic dialogue evaluation metrics such a PoE and Fined-Eval which are SotA models for self-assessment metrics. In addition, new Deep Learning algorithms for automatic classification of human movements using inertial sensors and multimodal processing for trustworthiness evaluation will be used to enrich and better understand the human-chatbot interaction.
In addition, Ecole Normale Supérieure in Paris (ENS) will work on detecting biases in language to establish mechanisms for evaluating and removing undesired behaviors in conversational system inherited from the training data used. Finally, the team from University Medical Center Hamburg-Eppendorf (UKE) will propose innovative mechanisms to evaluate the effects on humans when chatting with a conscious chatbot versus one without consciousness.
in the first year of the project, IndeepAI (IAI) has worked, in collaboration with Dr. Graziano, in proposing an implementation for AST that could be applied to Transformer-based models. They also performed several experiments to test if the proposed mechanism is able to provide such capabilities as we are looking for. Current results are promising according to the AST, but still need additional improvements and experiments to move into more complex problems and also to show its application on conversational systems. In addition, Universidad Poilitécnica de Madrid (UPM) has created a novel dialogue dataset in the domain of art-works that is being used for training the proposed chatbot. Finally, University Medical Center Hamburg (UKE) has been working on evaluating the current capabilities of LLMs in Theory of Mind tests, together with a large number of human users, to better understand how to evaluate consciousness in both humans and chatbots; this activity is planned in order to propose a new type of test that we termed Turing-Graziano test for evaluating consciousness.
ASTOUND proposes to build a plausible human-brain architecture for consciousness capable of demonstrating an undeniable added value in terms of human-machine interaction. This will be obtained by combining state of the art deep neural networks models with an Attention Schema.
The Attention Schema Theory (AST) , first conceived by Dr. Michael Graziano (Princeton Neuroscience Institute, part of ASTOUND’s team), explains the brain basis of subjective awareness in a mechanistic and scientifically testable manner. According to the AST, the brain is an information-processing device with the capacity to focus its processing resources more on some signals than on others.
On the other hand, the team at Universidad Politécnica de Madrid (Spain) has recently proposed state-of-the-art modules for conversational systems, as well as automatic dialogue evaluation metrics such a PoE and Fined-Eval which are SotA models for self-assessment metrics. In addition, new Deep Learning algorithms for automatic classification of human movements using inertial sensors and multimodal processing for trustworthiness evaluation will be used to enrich and better understand the human-chatbot interaction.
In addition, Ecole Normale Supérieure in Paris (ENS) will work on detecting biases in language to establish mechanisms for evaluating and removing undesired behaviors in conversational system inherited from the training data used. Finally, the team from University Medical Center Hamburg-Eppendorf (UKE) will propose innovative mechanisms to evaluate the effects on humans when chatting with a conscious chatbot versus one without consciousness.
in the first year of the project, IndeepAI (IAI) has worked, in collaboration with Dr. Graziano, in proposing an implementation for AST that could be applied to Transformer-based models. They also performed several experiments to test if the proposed mechanism is able to provide such capabilities as we are looking for. Current results are promising according to the AST, but still need additional improvements and experiments to move into more complex problems and also to show its application on conversational systems. In addition, Universidad Poilitécnica de Madrid (UPM) has created a novel dialogue dataset in the domain of art-works that is being used for training the proposed chatbot. Finally, University Medical Center Hamburg (UKE) has been working on evaluating the current capabilities of LLMs in Theory of Mind tests, together with a large number of human users, to better understand how to evaluate consciousness in both humans and chatbots; this activity is planned in order to propose a new type of test that we termed Turing-Graziano test for evaluating consciousness.
POTENTIAL IMPACT
The first direct contribution of ASTOUND will be an engineering method for the realization of “conscious” conversational agents (Chatbots) able to deliver more human-like and natural communication. These agents will also be able to learn from their own mistakes, through self-assessment. We can expect to drive the European Chatbot market segment in multiple verticals, such as Healthcare, Education, BFSI (Banking, Financial Services, and Insurance), Retail and Ecommerce, Telecom, and Manufacturing.
Society has the potential to benefit from conscious AI, enabled by this project, at multiple levels.
- In the Healthcare Sector, providing human-like understanding and empathy to conversational agents could support mental health applications like personal AI therapists.
- Conscious conversational agents could potentially impact Education, allowing more effective e-learning tools for student assistance thanks to an extended coherence of which a conscious agent would be capable.
On December 2022, we released a SotA model for automatic evaluation of open-domain dialogue systems at turn-level (Panel of Experts). However, the advent of LLMs such as ChatGPT has created a new paradigme and opportunities which we are now trying to improve by proposing more complex models and taking advantage of the emerging capabilities of these LLMs by focusing on new problems such as memorability, better ethical and moral value alignment, higher commonsense capabilities and integrated multimodal capabilities.
The first direct contribution of ASTOUND will be an engineering method for the realization of “conscious” conversational agents (Chatbots) able to deliver more human-like and natural communication. These agents will also be able to learn from their own mistakes, through self-assessment. We can expect to drive the European Chatbot market segment in multiple verticals, such as Healthcare, Education, BFSI (Banking, Financial Services, and Insurance), Retail and Ecommerce, Telecom, and Manufacturing.
Society has the potential to benefit from conscious AI, enabled by this project, at multiple levels.
- In the Healthcare Sector, providing human-like understanding and empathy to conversational agents could support mental health applications like personal AI therapists.
- Conscious conversational agents could potentially impact Education, allowing more effective e-learning tools for student assistance thanks to an extended coherence of which a conscious agent would be capable.
On December 2022, we released a SotA model for automatic evaluation of open-domain dialogue systems at turn-level (Panel of Experts). However, the advent of LLMs such as ChatGPT has created a new paradigme and opportunities which we are now trying to improve by proposing more complex models and taking advantage of the emerging capabilities of these LLMs by focusing on new problems such as memorability, better ethical and moral value alignment, higher commonsense capabilities and integrated multimodal capabilities.