First recall the variance definition.
Substitute random variable $X$ with $\hat \theta - \theta$, where $\hat \theta$ represents the estimate of the parameters given a certain architecture of the learning algorithm, and $\theta$ stands for the oracle optimal parameter setting.
Since $\theta$ is constant,
The second term is the definition of the mean squared error (MSE),
The third term is the definition of the squared bias,
Thus by moving the terms, finally we have,