This is a reimplement for linear Gaussian simulation in the paper " Predicting What You Already Know Helps: Provable Self-Supervised Learning " https://arxiv.org/pdf/2008.01064.pdf.
I run the experiment as the same set up in the paper for 1000 times.Surprisingly, using psi_* (the transformed feature to low dimension) is better tham using X_2 directly in 735 cases.
Any comments are welcome to the Issues.
