https://www.eet-china.com/mp/a123145.html
https://www.math.sinica.edu.tw/mathmedia/journals/4386
基於物理的機器學習
Here’s a table summarizing the analogy between the loss function in optimization and scalar potential in physics:
Aspect | Loss Function (Optimization) | Scalar Potential (Physics) |
---|---|---|
Definition | A scalar function that measures the "error" or "cost" for a given set of parameters . | A scalar field that represents potential energy per unit charge or mass at a point in space. |
Domain | Parameter space (e.g., weights of a neural network). | Physical space (e.g., 3D spatial coordinates). |
Gradient () | The gradient points in the direction of steepest increase in the loss. | The gradient points in the direction of steepest increase in potential. |
Negative Gradient () | gives the direction of steepest decrease in the loss. | gives the direction of steepest decrease in potential energy. |
Physical Interpretation | Minimizing adjusts parameters to improve model predictions (reduce error). | Systems naturally move to minimize potential energy, leading to equilibrium states. |
Movement | Parameters are updated iteratively in the direction to minimize loss. | Particles "move downhill" in the direction to minimize potential energy. |
Landscape | Loss function forms a "loss landscape" with hills, valleys, and minima. | Potential forms a "potential landscape" with similar features. |
Goal | Find the global minimum of the loss . | Find the state of lowest potential energy . |
Stochasticity | Stochastic gradient descent uses noisy estimates of , introducing randomness to avoid local minima. | Deterministic; no stochasticity in the potential field (but randomness can exist in other contexts, like thermal motion). |
Applications | Optimization in machine learning (e.g., neural network training). | Describing forces in physical systems (e.g., gravitational or electric fields). |
沒有留言:
張貼留言