Unveiling the Neuromatch course on NeuroAI
Intelligence and generalization in artificial and natural systems
Neuromatch is unveiling its new course on NeuroAI. This is an advanced course for students who have taken the computational neuroscience (CN) and deep learning (DL) NMA courses, or equivalent. Applications are open from March 1st until March 17th (for TAs) and March 24th (for students). The course will take place July 15th–July 26th, 2024.
As with the rest of the Neuromatch courses, it’s focused on practical computational concepts given in an interactive tutorial format through the Colab platform. Students in small virtual pods go through the course accompanied by a TA, exploring the state-of-the-art in neuroAI. Neuromatch is a 501(c)(3) non-profit dedicated to making science education more accessible and equitable. So far, 10,000 students from over 101 countries and territories have gone through the CN, DL and computational climate science courses.
What is neuroAI?
Our fearless course director Xaq Pitkow had a mammoth task: to define a burgeoning field that resists a short definition. Through many iterations of the curriculum, we’ve come to a story that I think is broad, inclusive and narratively satisfying.
NeuroAI is the study of the common principles behind natural and artificial intelligent systems. There are two main strains of practice in neuroAI:
The neuro → AI route: learning principles from brains that can be applied to create more robust and efficient artificially intelligent systems. This subfield is often known as neuro-inspired AI.
The AI → neuro route: taking inspiration from advances in artificial intelligence to understand natural intelligence. This subfield has also been referred to as neuroconnectionism.
These threads of research have ancient roots. Homer, in the Ilyad, describes the story of Hephaestus, god of metallurgy, who built a set of artificially intelligent agents (robots) made of gold to assist him in his workshop, in part to relieve him of physical limitations brought by his limp. Understanding biological intelligence–knowing oneself–is a theme that runs from classic Greece to ancient Vedic and Taoist traditions, with the first medical investigations of the brain documented in an ancient Egyptian papyrus dated 1600 BC.
The fact that these roots are so deep hints that both neuro and AI are tapping into deeply human desires to understand and build. One of the founding principles of the course is to instill a radical curiosity in students toward the goals of other disciplines. That means immersing oneself in the language and social context of both these disciplines and committing to read the primary literature from both sides. This is the key to bridging the gap between AI and neuroscience.
Generalization and inductive biases
We could have told many stories about natural and artificial intelligence to build a coherent narrative: phylogenetic refinement vs. the artificial evolution of AI systems; prediction and prospective learning; optimization of behaviour under resource constraints. Each of these contains a facet of truth about neuroAI, and one could create an entire course around each of these ideas.
I really like where we ended up with the overarching narrative: intelligent systems generalize. Intelligent systems must be able to adapt to changing circumstances–including circumstances previously unseen–rapidly and effectively. Because of the no-free lunch theorem, generalizing to arbitrary circumstances is impossible; it is only by having strong inductive biases that we’re able to generalize beyond our immediate experiences.
Starting from that premise, the course covers a wide range of topics from basic neuroscience all the way to theoretical deep learning, from synapses to symmetry. It’s a wide scope, and recognizing that a lot of the questions in our field are unsettled, we build each day in a T-shaped structure: we go broad in the intro lectures, and go deep in the weeds in the tutorials. We always tie back the days to the question of generalization. Our goal was to give students from all backgrounds the ability to read other literatures through a dense map of concepts tying the two fields together, and to continue on their learning journey far beyond the end of the course.
Course structure
There are 9 days of content, spread over two weeks:
Overview of neuroAI
Comparing tasks
Comparing networks
Micro-architecture
Macro-architecture
Cognitive structures
Micro-learning
Macro-learning
Mysteries
Students get to practice these concepts deeper in a set of projects led by the amazing Eva Dyer.
Sign up now
Applications are open from March 1st until March 17th (for TAs) and March 24th (for students). Course tuition is adjusted according to local cost-of-living to be affordable for all. It’s a great occasion to be part of the inaugural class for what we hope will be a foundation for the next generation of neuroAI scientists.