Why that fixed it

The earlier machine connected each taste straight to the price, with one weight per taste. Correcting its cabernet weight for one bottle damaged the guess for the other cabernet. The same number was being pulled toward two incompatible prices.

The middle row gives each wine type its own signal and therefore its own output weight.

When a plummy cabernet arrives, its middle unit reports 1. The other three report 0. The correction rule multiplies every change by that unit value, so only the plummy cabernet weight moves. The other three changes are multiplied by zero.

A correction for one wine can no longer damage another. The interference is not merely smaller. It is gone.

Nothing about a weight has changed. What changed is what the weights are attached to. A weight attached to a taste says what that taste is worth on its own. A weight attached to a wine-type signal says what that whole combination is worth.

The machine can now hold the right kind of statement.

There is a general shape here:

A layer in the middle turns things that cannot be told apart into things that can.

The four bottles arrive as sets of tastes that overlap. By the time they leave the middle row, they have become four separate signals. The hard part was never attaching a price. It was telling the bottles apart.

So the problem is solved.

Except we have not yet shown that the machine can build that useful middle row for itself.