Course overview
Modern deep learning methods provide some of the best tools to model behavior and brain function today. Excitingly, AI systems have become the first artificial models capable of matching human performance in sophisticated cognitive tasks, such as visual recognition, language processing, and strategic planning. This unique capability makes them a key test bed for neuroscience research: by studying how these AI systems solve complex problems, we can generate and test hypotheses about the computational principles that biological brains might use. Moreover, thanks to amazing progress in neuroscientific experimental recording techniques over the last decade, we now have access to vast amounts of complex data, which can be used in computational modeling, across multiple modalities – from neural activity of thousands of neurons, to anatomical details of neuronal circuits, to whole brain neural recordings during complex behavior of humans and animals. These exciting developments—in both AI methodology and neuroscientific recordings—have inspired an emerging area of research at the intersection of neuroscience and AI.
The course gives a hands-on introduction to modern AI methods, including deep learning, and how it can be used for analyzing and modeling brain activity and behavior. Experts in the field will teach the basics of AI, and how to use them as models of the brain, cognition, and behavior.
Course Directors
Georg August University Göttingen, Germany
Johns Hopkins University, USA
Max Planck Institute for Software Systems, Germany
HHMI Janelia Research Campus, USA
Course Faculty
Jacob Yates – Berkley, USA
Martin Schrimpf – EPFL, Switzerland
Kevin Miller – Google DeepMind, UK
Patrick Mineault – Amaranth Foundation, USA
Carsen Stringer – HHMI Janelia Research Campus, USA
Chris Summerfield – Oxford University, UK
John Krakauer – Champalimaud, Portugal
Memming Park – Champalimaud, Portugal
Andreas Tolias – Stanford, USA
Jonathan Pillow – Princeton University, USA
Shailee Jain – UC San Fransisco, USA
Rui Ponte Costa, University of Oxford, UK
Instructors
Janne Lappalainen – University Tübingen, Germany
Farah (Fengtong) Du – HHMI Janelia Research Campus, USA
Mathis Pink – MPI Software Systems, Germany
Manasi Malik – Johns Hopkins University, USA
Course Content
Programme
The first two weeks will consist of full day hands-on lectures and tutorials from introductory to more advanced topics. The third week is reserved for more extensive group research projects.
First week
- Introduction to NeuroAI and primer on deep learning for computer vision and language
- Neural network models of vision and calcium imaging
- Models for neuronal dynamics
- Data-driven models of visual cortex during active vision
- Understanding choice behavior using reinforcement learning models
Second week
- Engineering a less intelligent AI (the fly)
- System models of vision and language in the primate brain
- Skill, practice, plasticity, and the cognitive-motor interface
- Foundation models in neuroscience
Third week
The third week has lectures on more advanced topics. The remainder of the days are focused on project work.
Confirmed talks in the third week are:
- Learning and generalization in human and neural networks
- Engineering a less artificial intelligence
Students will design projects applying machine learning methods to neural recording and behavioral data sets.
Datasets
We encourage course faculty to bring datasets that can be used for projects. Here are some examples of datasets that will be available for project work
- Calcium imaging data from mouse primary visual cortex in response to natural images and videos from the SENSORIUM competitions 2022 and 2023
- Large dataset of human choices on a 4-armed drifting bandit task.
- Data from foveal V1 of free-viewing marmosets along with continuously monitored gaze. Data from Marmoset retina with the same eye traces.
- Publicly available data from the Brain-Score benchmark (vision and language). The vision data include non-human and human primate electrophysiology and fMRI recordings spanning the visual ventral stream and associated object recognition behaviors. The language data are human neuroimaging and reading time measurements. Meta-data from thousands of model alignment scores on all of the Brain-Score benchmarks.
- A dataset with over 29,000 neurons responding to up to 65,000 natural image presentations in mouse V1. The neurons were expressing jGCaMP8s and we recorded their activity using two-photon calcium imaging at a rate of 30Hz.
- Neuropixels speech dataset comprising spiking activity from >500 single neurons in awake human participants undergoing brain surgery. Neurons were recorded while participants listened to simple English sentences. Associated metadata for each neuron such as its anatomical position along the superior temporal gyrus and cortical depth will also be provided.
Techniques
Through hands on tutorials and project work, students will have the opportunities to learn about computational approaches in NeuroAI such as:
- Deep learning models for vision and language (CNNs, transformers)
- Reinforcement learning algorithms
- State space models for neuronal time series
- Machine learning for biophysical models of neuronal circuits
Champalimaud Centre for the Unknown, Portugal
The Champalimaud Foundation is a private, non-profit organization, established in 2005 and dedicated to research excellence in biomedical science. Completed in 2010, the Champalimaud Centre for the Unknown is a state-of-the-art centre that houses the Champalimaud Clinical Centre and the Champalimaud Research, with its three parallel programs – the Champalimaud Neuroscience Programme, the Physiology and Cancer Programme, and the Experimental Clinical Research Programme.
Initially focused on a system and circuit approach to brain function and behavior, the Centre expanded to incorporate molecular and cell biological expertise. The Centre comprises 26 research groups (circa 400 researchers) leading independent curiosity-based research.
Facilities
The Centre provides Facilities dedicated for Training, some in their entirety, for use by the CAJAL Advanced Neuroscience Training Programme. These include the Teaching Laboratory, a fully equipped open lab space for 20-30 students that can be dynamically reconfigured to support a full range of neuroscience courses. It also overlooks, via floor to ceiling windows, a tropical garden and the river. The experimental spaces include: Imaging Lab: A dark-room containing a full size optical table is used for advanced imaging setups (two-photon microscopy, SPIM, etc.) and custom (course-designed) optical systems.
Registration
Fee : 3 500 € (includes tuition fee, accommodation and meals)
Applications are closed.
The CAJAL programme offers 4 stipends per course (waived registration fee, not including travel expenses). Please apply through the course online application form. In order to identify candidates in real need of a stipend, any grant applicant is encouraged to first request funds from their lab, institution or government.
Kindly note that if you benefited from a Cajal stipend in the past, you are no longer eligible to receive this kind of funding. However other types of funding (such as partial travel grants from sponsors) might be made available after the participants selection pro- cess, depending on the course.
Sponsors
