Neuroscience meets AI

While the discipline of NeuroAI is still in the process of being formed, many opinion and review pieces covering sections of this upcoming field already exist. The recent review "Toward Next-Generation Artificial Intelligence: Catalyzing the NeuroAI Revolution" (https://www.nature.com/articles/s41467-023-37180-x) is a good starting point to learn about potential overlaps of Neuroscience and AI. Focusing more on potential benefits on the AI / ML side, "Deep learning needs a prefrontal cortex" (https://baicsworkshop.github.io/pdf/BAICS_10.pdf) and “Building artificial neural circuits for domain-general cognition: a primer on brain-inspired systems-level architecture” (https://arxiv.org/abs/2303.13651) provide good starting points. For highlighting the impact on neuroscience, “Deep learning framework for neuroscience” (https://www.nature.com/articles/s41593-019-0520-2) and “The neuroconnectionist research programme" (https://arxiv.org/abs/2209.03718) are great places to start. In addition, Patrick Mineault is writing on NeuroAI In his blog (https://xcorr.net). None of these get to fully define the field but can be good starting points to learn about current questions, assumptions and key references in the field.

Any research question that is part of artificial intelligence or (computational) neuroscience can be a potential research question within NeuroAI — as such, providing an exhaustive list is clearly not possible. What I do want to provide here instead is a (very subjective) list of research directions which could benefit to a special degree from being pursued on the interaction of neuroscience in AI. I provide references discussing each of these research directions.

• The interaction between structure and function in neural network (see https://www.cell.com/trends/cognitive-sciences/fulltext/S1364-6613(20)30026-7 and https://www.nature.com/articles/s42256-023-00748-9)

• The role of neural representations for computations in distributed computing systems (https://arxiv.org/abs/2310.13018 and https://www.annualreviews.org/content/journals/10.1146/annurev-neuro-092619-094115)

• Efficient local and distributed learning (https://arxiv.org/abs/2305.11252 and https://www.cell.com/trends/cognitive-sciences/fulltext/S1364-6613(19)30061-0)

• Planning and exploration in open-ended environments (https://www.cell.com/neuron/fulltext/S0896-6273(21)01035-7 and https://royalsocietypublishing.org/doi/full/10.1098/rsos.230539)

• Interpretability of large-scale neural networks (https://ieeexplore.ieee.org/abstract/document/9521221)

• Hardware-algorithm co-optimisation, including neuromorphic computing (https://ieeexplore.ieee.org/document/10144567/ and https://www.nature.com/articles/s41586-020-2782-y) 

It is an incredibly exciting time to work on the intersection of biological and artificial intelligent systems. The experimental and modelling tools in neuroscience are incredibly sophisticated and at the same time we see new artificially cognitive systems being developed every day, challenging researchers to include them in our overall understanding of intelligent systems and distributed computing. While being part of such a rapidly evolving field is incredibly exciting, starting in it as a new researcher can be very intimidating. So where should students start? While everybody learns a bit differently, I think it can be helpful to think of your pathway in NeuroAI in three stages: (1) the prerequisites, (2) your knowledge pool, and (3) research project.

(1) The prerequisites: There are some basic tools you will need if you want to work in NeuroAI that are probably best studied by yourself before joining a research project, which are fundamental principles coding in Python and some sort of neural network library (probably best to start with Pytorch). Luckily there are various resources on this online and Google Colaboratory will provide students with enough free computing resources needed for learning. You should probably also be familiar with some basic things in neuroscience. These courses from the University of Washington and Imperial College London might be good starting points but there is nothing specific a student will need to know — it is more about knowing roughly what neuroscience research is about. While it is more specifically focused on computational neuroscience and theoretical neuroscience, you might find my online lecture material helpful: https://www.jachterberg.com/cbu_methods 

(2) Your knowledge pool: Once you know the basics, you gradually fill your knowledge pool by reading papers or watching more lectures covering neuroscience and / or AI. This is a free exploration stage where you can just use YouTube, Google Scholar, and perhaps Twitter / X to identify content on research you find interesting. It is now worth noting that both AI and Neuroscience are massive fields, so it clearly is not practical to strive to know everything. I think this phase is mostly good to dig a bit deeper until you found an overall topic that catches your interest (best in the form of a recent research paper). Another topic you can fill your knowledge pool with is your knowledge of how write good code — coding is your main tool in this discipline, so gradually picking up good practices will really be helpful (good resources for this could be https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005510   and  https://goodresearch.dev/index.html ). 

(3) Research projects: I have always learned best when working on a specific project and so my recommendation tends to be that after having learned basic prerequisites and finding some topics you are interested in, try to identify a little project you can work on by yourself, or even better, try to join a research group where you can work with. Many professors will be excited to hear from students of all backgrounds that are motivated to work with them and their labs, as long as they know the basics in a field. Having that said, professors also get tons of emails, so if you do not get a response, it can often be a good idea to identify a senior graduate student or postdoc in a lab you are interested in. When contacting them, give a brief overview of what research interests and relevant skills are, and attach a general academic CV – that should be sufficient for an initial contact. If you have worked on a relevant project already and maybe have some code available on GitHub, then that can be a great proof of proficiency and motivation. An alternative to joining a classical academic lab, you can also find people to work and collaborate with in open research communities, like OpenBioML (https://openbioml.org), MedARC, or EleutherAI. If you are a trainee in NeuroAI, you mind also find this more general list of resources helpful: https://github.com/8erberg/NeuroAI_Trainee_Resources 

Generally speaking, if you have any questions regarding this process or want to discuss research ideas, please drop me an email. I generally respond to all email from students interested in neuroscience and / or AI. You can find my current email address at the bottom of the webpage or on the ‘Home’ page.

This page is not specifically a guide on how to find a good PhD place and plenty of detailed articles have been written about this (for example Tim Sandhu’s blog post on this but there are many others). My general advice is to first and foremost find a supervisor who will support you on your path to becoming a researcher. Having the right supervisor is usually much more important for your development than the specific university you are at. If you are interested in identifying researchers working on NeuroAI that might be suitable PhD supervisors, I recommend you have a look at the authors of the papers in the reading list above. Let me add that applying for PhD can be difficult, especially at prestigious universities, but if you like research, it is likely worth trying. 

A last note in this section is aimed at the Gates Cambridge Scholarship that supported my PhD studies, as I frequently get asked about it from prospective students. I do think the Gates community is great and you should absolutely apply for it if you are considering Cambridge for your graduate studies. At the same time, it is an extremely competitive scholarship, and a successful application is to a large degree a mix of matching the specific criteria Gates list on their webpage and also a good chunk of luck. Cambridge is a great place to pursue a PhD, but you should definitely consider alternative scholarships alongside Gates. 

As with the topics above, I am happy to be contacted with question about PhD applications, as long as they are within the realm of neuroscience and / or AI. The application process can differ quite a lot between subjects, so I will likely not be of much help for subjects outside my own area of expertise.