Using Epigenetically-Inspired Connectionist Models to Provide Transparency In The Modelling of Human Visceral Leismaniasis
Lead Research Organisation:
University of Hull
Department Name: Computer Science
Abstract
Biologically inspired connectionist models are made up of multiple interconnected units which are designed to mimic biological processes in nature which give rise to emergent phenomena. Typically, connectionist models are used as computational tools which are capable of learning by example, for instance predicting the next days activity on the stock exchange by learning from previous months data. Epigenetically inspired connectionist models (EICMs) are a particular type of biologically inspired connectionist model which allow for the activation and deactivation of their interconnected units whist they are solving a task. These models have been shown to break complex tasks down into smaller sub-tasks autonomously, with certain interconnected units being applied to certain sub-tasks, and other interconnected units being applied to other sub-tasks.
Biologically inspired connectionist models in general are difficult to interpret. Their decision making processes are an emergent property of their interconnected units, from which it is very difficult to provide an explanation as to why specific decisions have been made. Because of this, deriving confidence from the decisions they make is difficult. Having confidence in the decision making process is of importance especially when the tasks they are applied to are in domains which are considered "high risk" such as medical simulations and financial forecasting.
To address these issues, this work aims to develop a set of techniques which allow for EICMs to provide a rationale for their decision making process, essentially making its decisions transparent. This will be achieved by analysing the way the model breaks down complex tasks, which of its units are active at any given time and then correlating this with the behaviour of both the network and the task.
We apply the EICMs and the techniques developed in this project to improve the understanding of the often fatal disease human visceral leismaniasis (HVL). The immune response to HVL is a significant indicator of patient outcome and is the product of the interplay between multiple interacting cells, macrophages and specific cytokine responses. The project partner Simomics, a world leading disease modelling company, has a comprehensive data set which describes changes to the immune response in reference to HVL over varying timescales, and has provided it for use during this project.
The overall development of HVL and the immune response to it is not well understood. The techniques developed in this work which are able to provide a rationale for their decision making process, will be applied to learn the interplay and interactions between these processes. This will allow for model to provide an explanation of what processes are most important in the immune response over the duration of HVL infection.
By contributing to the field of biological modelling, which places a strong emphasis on transparency and confidence in results, other fields will be able to adopt the models developed in this work to provide transparency in other domains.
Biologically inspired connectionist models in general are difficult to interpret. Their decision making processes are an emergent property of their interconnected units, from which it is very difficult to provide an explanation as to why specific decisions have been made. Because of this, deriving confidence from the decisions they make is difficult. Having confidence in the decision making process is of importance especially when the tasks they are applied to are in domains which are considered "high risk" such as medical simulations and financial forecasting.
To address these issues, this work aims to develop a set of techniques which allow for EICMs to provide a rationale for their decision making process, essentially making its decisions transparent. This will be achieved by analysing the way the model breaks down complex tasks, which of its units are active at any given time and then correlating this with the behaviour of both the network and the task.
We apply the EICMs and the techniques developed in this project to improve the understanding of the often fatal disease human visceral leismaniasis (HVL). The immune response to HVL is a significant indicator of patient outcome and is the product of the interplay between multiple interacting cells, macrophages and specific cytokine responses. The project partner Simomics, a world leading disease modelling company, has a comprehensive data set which describes changes to the immune response in reference to HVL over varying timescales, and has provided it for use during this project.
The overall development of HVL and the immune response to it is not well understood. The techniques developed in this work which are able to provide a rationale for their decision making process, will be applied to learn the interplay and interactions between these processes. This will allow for model to provide an explanation of what processes are most important in the immune response over the duration of HVL infection.
By contributing to the field of biological modelling, which places a strong emphasis on transparency and confidence in results, other fields will be able to adopt the models developed in this work to provide transparency in other domains.
Planned Impact
The proposed research aims to develop connectionist models and computational tools which, inspired by epigenetic functionality in biology, allow for robust computational performance whilst allowing for transparent execution so that an explanation can be provided of its decision making processes.
Biological modelling impact
In general, when connectionist models are used in biological modelling, its rationale for decisions being made are hidden from view. In addition to understanding the immune response to human visceral leismaniasis, this work aims to provide a platform for biologists, immune modellers and other "high risk" fields to be able to use the tools and models developed in this work. By sharing the software developed in this project we aim to give other domains the ability to interact with this work with no prior training and no expertise in computer science.
Bio-inspired impact
The work in this project takes a different approach to the classical form of bio-inspired computational models, namely that architectures can change their topology during execution. By showing that this leads to the novel transparency where a rationale is provided for the decisions it is making, this could generate a lot of interest in the scientific community for such models. In addition, these ideas could be used in other areas of bio-inspired computation other than what is prescribed in this work, as the benefits might be transferable.
Economic impact
This project has potential to stimulate a wide range of economic areas, as understanding the rationale of connectionist architectures is imperative to instil confidence in the results they provide. Working with Simomics in this project, we will provide a free lightweight version of the software developed in this work. This is to be utilised for independent researchers and small businesses. A more comprehensive commercial piece of software will also be developed with support for more large scale businesses which have complex and specific needs.
Impact on society
The open source tools developed throughout this work will be designed to be used by both experts and non-experts in computer science. It will be made as simple as possible for users to work with these novel technologies, to re-visit previous work and to generate a better understanding how it has been interpreted by other models. In addition, the tool itself can also be used to highlight new areas of investigation which have been previously unknown. This, twinned with the simplicity of use could have a profound effect on a number of different groups, fields and industries.
Biological modelling impact
In general, when connectionist models are used in biological modelling, its rationale for decisions being made are hidden from view. In addition to understanding the immune response to human visceral leismaniasis, this work aims to provide a platform for biologists, immune modellers and other "high risk" fields to be able to use the tools and models developed in this work. By sharing the software developed in this project we aim to give other domains the ability to interact with this work with no prior training and no expertise in computer science.
Bio-inspired impact
The work in this project takes a different approach to the classical form of bio-inspired computational models, namely that architectures can change their topology during execution. By showing that this leads to the novel transparency where a rationale is provided for the decisions it is making, this could generate a lot of interest in the scientific community for such models. In addition, these ideas could be used in other areas of bio-inspired computation other than what is prescribed in this work, as the benefits might be transferable.
Economic impact
This project has potential to stimulate a wide range of economic areas, as understanding the rationale of connectionist architectures is imperative to instil confidence in the results they provide. Working with Simomics in this project, we will provide a free lightweight version of the software developed in this work. This is to be utilised for independent researchers and small businesses. A more comprehensive commercial piece of software will also be developed with support for more large scale businesses which have complex and specific needs.
Impact on society
The open source tools developed throughout this work will be designed to be used by both experts and non-experts in computer science. It will be made as simple as possible for users to work with these novel technologies, to re-visit previous work and to generate a better understanding how it has been interpreted by other models. In addition, the tool itself can also be used to highlight new areas of investigation which have been previously unknown. This, twinned with the simplicity of use could have a profound effect on a number of different groups, fields and industries.
People |
ORCID iD |
| Alexander Turner (Principal Investigator) |
Publications
Alexander Turner
(2021)
Stochasticity improves evolvability in artificial gene regulatory networks
| Description | One of the issues with typical artificial intelligence is that it is difficult to interpret it's decision making process. By using epigenetically inspired connectionist models, we have been able to adapt previously non-transparent machine learning architectures which are in wide use to present transparent properties. In particular, we were able to add these features to pre-existing deep neural networks. These architectures are more scalable and can be applied to larger and more complex data sets and used by a wider community. The results show that with minor trade offs in performance, we can greatly reduce the size of pre-existing neural networks, as well as reduce the amount of data processing done by the nodes within the network. This combination of features allows us to deduce rules about how the networks are functioning over time. This has far reaching effects in the broader communities as users will be able to generate more information about how their connectionist architectures are functioning than would be previously possible. In addition, significant effort has been made to ensure that these techniques developed to allow transparency are easy to use for non-experts in the field. The primary tool in achieving this is to incorporated the resources we have developed into open source software packages. |
| Exploitation Route | One of the issues with the previous work in epigenetically inspired connectionist models is that they worked well on smaller tasks / data sets. This work allows the epigenetically inspired connectionist models to be applied to much larger networks which can be trained much faster. More generally, it allows transparency to be part of more widely used connectionist architectures. We are currently in the process of bringing together this work to form a publication detailing these processes and development. It is planned to be submitted in April 2020. It is possible that future work could be based upon this to allow for more transparent decision making processes in a wider range of computational 'black box' tools. |
| Sectors | Digital/Communication/Information Technologies (including Software) |