1: \begin{abstract}

2: % Teleportation accelerates convergence rate, can be used to improve generalization, and is compatible with various optimization algorithms. 

3: 

4: In overparametrized models, different values of the parameters may result in the same loss value.

5: Parameter space symmetries are transformations that change the model parameters but leave the loss invariant. 

6: Teleportation applies such transformations to accelerate optimization. However, the exact mechanism behind this algorithm's success is not well understood.

7: % In this paper, we show that 

8: % symmetry teleportation, intercalated with gradient steps, results in faster overall convergence.

9: In this paper, we show that teleportation not only speeds up optimization in the short-term, but gives overall faster time to convergence.

10: Additionally, we show that teleporting to minima with different curvatures improves generalization and provide insights on the connection between the curvature of the minima and generalization ability.

11: Finally, we show that integrating teleportation into a wide range of optimization algorithms and

12: optimization-based meta-learning improves convergence.

13: \end{abstract}

14: