I give a try to read the arXiv paper of Calandriello et al. 2017 but failed to understand the link between the actual paper formula in section 3, Sequential RLS Sampling 馃憞

and the code implementation of compute_tau:
def compute_tau(centers_dict: CentersDictionary,
X: np.ndarray,
similarity_func: callable,
lam_new: float,
force_cpu=False):
.
.
.
diag_norm = np.asarray(similarity_func.diag(X))
# (m x n) kernel matrix between samples in dictionary and dataset X
K_DU = xp.asarray(similarity_func(centers_dict.X, X))
# The estimator proposed in Calandriello et al. 2017 is
# diag(XX' - XX'S(SX'XS + lam*I)^(-1)SXX')/lam
# Here for efficiency we collect an S inside the inverse and compute
# diag(XX' - XX'(X'X + lam*S^(-2))^(-1)XX')/lam
# note that in the second term, we take care of dropping the rows/columns of X associated
# with 0 entries in S
U_DD, S_DD, _ = np.linalg.svd(xp.asnumpy(similarity_func(centers_dict.X, centers_dict.X)
+ lam_new * np.diag(centers_dict.probs)))
U_DD, S_root_inv_DD = __stable_invert_root(U_DD, S_DD)
E = xp.asarray(S_root_inv_DD * U_DD.T)
# compute (X'X + lam*S^(-2))^(-1/2)XX'
X_precond = E.dot(K_DU)
# the diagonal entries of XX'(X'X + lam*S^(-2))^(-1)XX' are just the squared
# ell-2 norm of the columns of (X'X + lam*S^(-2))^(-1/2)XX'
tau = (diag_norm - xp.asnumpy(xp.square(X_precond, out=X_precond).sum(axis=0))) / lam_new 馃憟
Like,
- Is
X'X reflect the kernel matrix $\mathbf{K}_t=\boldsymbol{\Phi}_t^{\top}\boldsymbol{\Phi}_t$? And what about $XX'$?
- I couldn't understand what is
X_precond here and why svd decomposition needed.
- In section 3 there was the definition of a dictionary, "we redefine a dictionary as a collection $\mathcal{I}={(i,\widetilde{p_i},q_i)}$, where $i$ is the index of the point $x_i$ stored in the dictionary, $\widetilde{p_i}$ tracks the probability used to sample it, and $q_i$ is the number of copies (multiplicity) of i." - here I couldn't understand the $q_i$
- Overall, I feel, I didn't understand the
EXPAND and SHRINK for Algorithm 1 intuitively. It will be a great help if you comment something on this.
It would be greatly appreciated if you could assist me in resolving this matter.
I give a try to read the arXiv paper of Calandriello et al. 2017 but failed to understand the link between the actual paper formula in section 3, Sequential RLS Sampling 馃憞
and the code implementation of
compute_tau:Like,
X'Xreflect the kernel matrixX_precondhere and whysvddecomposition needed.EXPANDandSHRINKfor Algorithm 1 intuitively. It will be a great help if you comment something on this.It would be greatly appreciated if you could assist me in resolving this matter.