RBMs weights during learning uniform distribution

I am implementing Bernoulli-Bernoulli RBM and trying to debug it. I want RBM to learn uniform distribution over two binary variables. So RBM has 2 visible and 4 hidden units. It seem that RBM will end up with zero weight matrix, zero bias vector for both visible and hidden units.

I use learning rate = 0.1, contrastive-divergence k = 3000 and full batch update. As this is a toy task I can compute marginal distribution over visible units. So after random initialization from normal distribution with mean = 0 and standart deviation = 0.01 I have: (b for visible bias, c for hidden bias)

W:

p(0,0) = 0,250 p(0,1) = 0,248 p(1,0) = 0,252 p(1,1) = 0,250

After 1000 iterations

W:

p(0,0) = 0,263 p(0,1) = 0,269 p(1,0) = 0,231 p(1,1) = 0,237

After 10000 iterations

W:

p(0,0) = 0,247 p(0,1) = 0,248 p(1,0) = 0,247 p(1,1) = 0,258

I did not use weight decay neither momentum. So my question is why does this happens? Is it bug in my implementation or such behaivor can be explained somehow?