The approach specifies objectives (classification accuracy, reward functions) and searches program space via gradient descent to find neural networks that meet those objectives. This aligns with the Software 2.0 framing.
You are viewing a single comment's thread from: