At this scale, it is impossible to even describe the rules of the game let alone the method to determine a winner.
If we pick a random start and end point within the network, many paths connect the two points together as an open ended solution. However if we include additional criteria, the constrained path sets become recognizably unique. For example, we can differentiate them with rules such as the paths must - include defined points, be shorter than a certain distance, avoid certain neurons, be restricted to certain diameters, include a certain number of turns, etc. Once criteria is assigned, these network paths can be interpreted and compared. This process of selection is analogous to decision making in the everyday world such as how we eat a meal. We can critique the choices of quantifiable neural networks with reasonable intelligibility.
The complexity of a neural network problem can also increase exponentially by including multiple rules and as the two points diverge to include additional paths. This complexity can even extend beyond the currently available paths to include future considerations where not all the paths can be definitively described yet. For example, most complex problems also have to take into account outside, uncontrollable forces so some rules must estimate probabilities for paths not yet visible. With multiple competing probability rules, we also have to rank their importance. This approximates the conditions of playing a game of go where an opponent’s moves must be contemplated and where strategy is difficult to critique while underway. Problems are less quantifiable and rely on qualitative interpretations of patterns. While this is more challenging, there are still clear limits to the variables at play since the game has rules and a clear method to determine a winner.
The deepest neural network problems are multifaceted global problems where we are trying to link together multiple qualitative neural networks into a fog of consequence. It requires a simultaneous understanding of multiple scales (forest, tree, leaf) while observing from multiple points of view (forward, beneath, within) and planning for systems of growth (embryo, child, adult). At this scale, it is impossible to even describe the rules of the game let alone the method to determine a winner because we are aiming for a network of improvement. Both the delineation of the solution AND the problem are neural networks and the solution changes according to the constant change in the definition of the problem. These neural networks are impossible without computational tools, if possible at all, as we can only interpret issues sequentially ourselves. Therefore, it is and always will be impossible for us to control the process without being completely dependent on computational tools to plan and reason. At this deepest level, the best we could hope for is to create a system with immeasurable power and good intentions.