Although neuroscience has inspired many elements of artificial neuronal networks, the mammalian visual system is still markedly different from current state-of-the-art deep neural networks in terms of its circuit architecture, robustness, and ability to learn. Two groups at the Bernstein center (Bethge, Sinz) are part of a multi-university consortium funded by the MICrONs program within the Obama BRAINinitiative, that is setting out to narrow the gap between current state-of-the-art deep learning and algorithms of the mammalian visual system by exploring circuit level functional and anatomical patterns of populations of cortical neurons.
The approach is guided by data collected from cortical networks across three hierarchically organized visual areas in a behaving mouse at an unprecedented scale and detail. Using two- and three-photon imaging the team is collecting the distributed patterns of neural activity in response to artificial and natural stimuli from nearly all ~100,000 neurons in 1mm3 of cortex in vivo. The functional imaging is followed by dense electron-microscopy reconstruction of the connectome in the same tissue. Using this historic dataset, the team aims at a full characterization of the functional properties, morphology, and connectivity of all neurons in all six layers across three interconnected areas of the mammalian cortex, and the computations they implement.