Editorial overview: Computational neuroscience as a bridge between artificial intelligence, modeling and data
Pietro Verzelli,Tatjana Tchumatchenko,Jeanette Hellgren Kotaleski
DOI: https://doi.org/10.1016/j.conb.2023.102835
IF: 7.07
2024-01-06
Current Opinion in Neurobiology
Abstract:Computational neuroscience continues to be a broad and dynamic discipline that transforms itself as new experimental methods make the collection of new, often multi-modal data types possible. Initially, the focus was on models of single neurons, their ion channels, the statistics of their spikes, and their processing of inputs. New multiunit recording techniques that were possible first in vitro and then gradually in vivo allowed the field to expand into the question of how populations of neurons form circuits that represent sensory stimuli or motor output. Addressing these questions revealed many mathematically and conceptually challenging directions, such as low-dimensional manifolds, their synthesis, and representational meaning. Questions about neural synchrony, phase locking, neural attractors, and their relation to learning, memory, and sleep entered the stage. In the next decades, increased automatization of connectomics studies in organisms such as flies and mice allowed the correlation of the observed activity with the probing of the underlying connectivity. Therefore, neural activity could now, for the first time, be combined with the circuit structure that generated it, enabling detailed insights into the role of cell types in generating specific aspects of neural population activity. New machine-learning methods now enable tracking and segmenting behavior into smaller units while recording neural activity. These methods are complemented by advances in cell profiling and detailed molecular characterization of neural activity and structure, such as quantitative bridges between molecules and behavior, which can be unraveled for the first time. This is an exciting time to be a computational neuroscientist for multiple reasons. First, emerging datasets that cover multiple neural and behavioral levels are emerging and demand new computational approaches. Artificial neural networks can now enable online behavioral segmentation across species and behavioral tasks, but identifying the equivalent of a "grammar" that tells how to use a specific neural activity pattern in a specific region toward the construction of a behavioral trajectory in space is still a challenge. Also, at the level of cellular and molecular mechanisms, new proteomics and connectomics data sets allow unprecedented insights into the learning mechanisms at the single synapse or dendritic branch level, but translating these insights into learning rules in artificial neural networks or revealing their functional correlates in brain circuits remains a fundamental challenge. In this issue, computational neuroscientists, whose diverse expertise ranges from molecular interactions to behavioral characterization, share their perspectives on our current knowledge and future challenges and highlight recent breakthroughs. These reviews can be seen as covering five broad areas: 1) machine learning for neuroscience; 2) neurosciences for machine learning; 3) neural computation; 4) computational models; and 5) data analysis methods. Below, we try to map the different contributions to these categories, although several of the reviews fit in more than one.
neurosciences