Unifying models of information processing across machine learning, artificial intelligence and neuroscience

 The structure and function of circuits in the brain have inspired many innovations in artificial intelligence (AI) and robotics. New mechanisms for how information is processed in the brain are constantly being discovered, both at the cellular level (how individual neurons, synapses and circuits process input signals received from our senses), and at the cognitive level (how decisions are made and actions planned). A better understanding of these processes, honed over millions of years of evolution, can help us make AI systems more robust and efficient. On the other hand, advances in AI provide neuroscientists with important clues on how the brain may process information and provide powerful new computational tools to add to their arsenal of research techniques. However, because researchers and software developers in these fields often have different objectives and use different terminology, workflows and approaches, there is still a big disconnect between these areas. This makes it extremely difficult for researchers to share and exchange their ideas, their latest findings, and their tools (such as models and software) with each other.

An example is the variety of approaches being taken to studying and building computer models of vision. Huge progress has been made in Machine Learning (ML) for image recognition and classification using deep convolutional neural networks. Many computational neuroscientists on the other hand investigate vision using spiking neuronal elements arranged in populations inspired by the visual processing pathway of the brain. Cognitive scientists try to understand object recognition and subsequent decision making from a more abstract, higher level. While all of these perspectives are important, they use very different software frameworks and terminology for building models and disseminating their work, limiting how progress in one domain can be readily interpreted and reused in another.

I aim to address this unnecessary disconnection between disciplines in this fellowship. I have extensive experience in computational neuroscience and the development of standards, tools, and infrastructure that enable building, sharing, and reuse of complex, biologically realistic models. I am the main developer of the well established NeuroML exchange format and associated software tools that are used widely by researchers and large scale brain initiatives around the world. To expand the scope of this into related domains, I recently initiated a new international collaboration to develop MDF (Model Description Format). MDF is designed to be a more general format for models across both AI and neuroscience - from complex deep learning models and artificial neural networks, all the way to biologically detailed neuronal models and models of cognition. I will build on my preliminary work in this area to expand the scope of MDF and create associated analysis methods to provide a powerful suite of tools for a wide range of researchers and application developers working with brain-inspired network models. This work will be guided by specific scientific use cases based on my previous research (cortical computation, in-silico emulation of worm behaviour), where widely varying approaches are used by different researchers to examine these complex systems.

An EPSRC Open Fellowship will provide the resources necessary to develop and expand these technologies while acquiring new scientific and professional skills necessary to lead in this area. The Plus Component is an absolutely crucial part of this, supporting me to actively engage with, and disseminate these approaches to researchers from a wide range of fields, as well as build a diverse community of users and developers around the technologies. Many of the barriers to communicating ideas across AI/ML/neuroscience are related to lack of the underlying