Could it be that we've already solved the enigma of "how the brain works" and all that's left is to complete the details?

Central to this somewhat provocative question lies the differentiation between the computational and implementation levels of analysis. The impressive capabilities of our brains stem from computational principles that find parallels in the field of artificial neural networks (aNNs), with units swapping signals based on distinct connectivity structures and communication strengths.

Echoing Feynman's principle of 'what I cannot create, I do not understand,' does this high-level consideration on computations in aNNs provide enough groundwork for constructing or simulating a brain? Suppose for instance we had comprehensive knowledge of the connectivity architecture (connectome) and communication weights (synaptic weights) among cells. What other biological intricacies ought to be considered to faithfully construct or simulate a brain using a large-scale aNN? Striking cell morphologies, intricate synaptic and cellular mechanisms, a multitude of diverse peptides and molecules swirling around – are these all indispensable components that the simulation must encompass?

Our approach is to start simple, examining how much progress can be achieved by using conventional CNN/RNN networks and constraining their architectures and connection strengths based on biological insights from (1) the inputs received, e.g., sensory information, (2) cellular activities, (3) functional connections, and (4) the behavior of the animal.

We focus on the mouse visual system where we investigate how it processes 'naturalistic' stimuli such as visual textures. To address the points listed above, we train mice in visual discrimination tasks and adopt all-optical techniques to both record from and perturb large groups of neurons at the level of individual cells.

Artificial neural networks constrained using this biological data can then be utilized as predictive models, for instance, by predicting outcomes – also behavioral – in response to optogenetic perturbations or new stimuli. When deviations from these predictions arise, they introduce novel constraints that can be employed to enhance and fine-tune the models further.

If you like this approach, click on this link