Co-operation Between AI and Neuroscience Research, and Future Problems

Based on an article published in 'Frontiers in Computational Neuroscience' May 6 2020

Development of Artificial Intelligence (AI) has heavily utilised research done in neuroscience. Neural Networks are a prime example of biological neuroscience research models of neurons being transferred into AI research and development. AI has also aided neuroscience researches in understanding biological brains. However, the diverging goals and applications for neuroscience research and AI research are a threat to the continued synergy between the fields.

The Divergence

The authors of the article have defined a three axis model of the goals of researchers in the AI and neuroscience fields in order to show the difference. They define research as answering some of the following questions; "what is it?", "how does it work?" and "what does it do?". These can be simplified into understanding the form, mechanism or function respectively.

Using this model, the authors summarised that most neuroscience research is focused on the form or what makes up the brain. Secondary is research into the mechanism or how the brain works. The obvious reasons for this stem from the fact that neuroscience can study biological brains which are a physical object the we can learn about, hence form. Neuroscience also relates to medicine, and would be interested in how the brain works, and what areas of the brain perform what functions. This is so that they can develop treatments for neurological illnesses, hence they are interested in the form and the mechanisms in the brain.

AI research is very different, particularly in the fact that the models are artificial, and development is in the creation of these models. AI research is also focused at solving particular jobs such as image recognition which is based on function, making this its primary focus. The secondary focus is the mechanisms as this reflects how the AI systems work and learn, and this is the information that developers use to create the AI models and artificial neural networks.

Neuroscience	Artificial Intelligence
1. Form	1. Function
2. Mechanism	2. Mechanism

Importance of Neuroscience for AI

Following a formal understanding of the visual cortex in human brains, a precursor neural network to modern day CNNs was able to be created. AI researchers very often explore neuroscience inspired approaches to AI design and modelling. However, history has shown that such a model still needs to provide benchmarked improvements over current models for continued development from AI researchers.

Neuroscience research is in an exciting era as the technology used to study the brain is become more and more advanced, allowing measurements of more neurons and a larger time frame. Higher resolution and multidimensional data is also able to be analysed more effectively than ever before using new data science and greater computational power. Temporal dynamics seem to play a major role in biological brains and it is now able to be fully observed. With this, neuroscience is on the cusp of a 'complete connectome', which is a full map of the neuron connections in a human brain. This would remove the roadblock in AI research that was the question; "how is the human brain wired?".

Tasks for a computer can be divided into two categories, those that a human is better at, and those that a computer is better at. AI has always been focused on tasks a human is better at, and has been looking at ways to improve the computer at said task. Neuroscience research is often learning about these tasks that a human can beat a computer at, however, there is no perfect understanding of the human brain in this field which limits the ways in which AI can improve from neuroscience. Understanding of the brain in these tasks, when there is any, can sometimes come from macro understandings of function, such as understandings of memory. This is disconnected from the physical neural circuitry. As said before, AI research could benefit from better functional and mechanical understanding, but the lack of mechanical understanding of these areas is very limiting. This is combined with the functional understanding being unproven and split into multiple plausible theories for the same task. There is a divide in near-future neuroscience research, as the tools look into an understanding of form and mechanics, but do not look at function. This divide is a major challenge and creates the difficulties in the continued co-operation between neuroscience and AI.

Conclusion

For future "cross-pollination" between neuroscience and AI research, it would be very effective for a shift in communication. Neuroscience can extrapolate their understanding of form and mechanics of the brain towards the implementation of the mechanics into artificial neural networks. This can be extrapolated to the function of the system which would represent the goals of an AI model. Doing this is likely to significantly aid in the future development of AI particularly in tasks were humans out perform computers, and neuroscience is an effective source of new approaches.

References

Subject paper: "Crossing the Cleft: Communication Challenges Between Neuroscience and Artificial Intelligence",
https://doi.org/10.3389/fncom.2020.00039