Interpolation consistency training

For two sampled unlabeled minibatches, obtain fixed teacher predictions f_theta_prime(u_j) and f_theta_prime(u_k). Draw lambda from a distribution on [0,1], with Beta(a,a), a>0, used in the reported method. Form u_mix=lambda*u_j+(1−lambda)*u_k and q_mix=lambda*f_theta_prime(u_j)+(1−lambda)*f_theta_prime(u_k). Minimize labeled cross-entropy plus a ramped coefficient times a discrepancy between f_theta(u_mix) and q_mix; the experiments use mean squared error, while Equation1 allows a general discrepancy. Update the student from gradients with respect to theta, holding the teacher targets fixed. Algorithm1 explicitly computes these gradients, updates theta_prime←rho*theta_prime+(1−rho)*theta using the pre-step student, and then takes the student optimization step; this order differs from averaging the already-updated student. The displayed algorithm returns the student parameters. Although Algorithm1 and Figure2 display ordinary labeled cross-entropy, Section3.3 reports applying supervised MixUp to labeled examples in the experiments. This protocol choice is not silently substituted into the displayed core algorithm.