We propose a novel high dimensional nonparametric model named ATLAS which naturally generlizes the sparse additive model. Given a covariate of interest $X_j$, the ATLAS model assumes the mean function can be locally approximated by a sparse additive function whose sparsity pattern may vary from the global perspective. We propose to infer the marginal influence function $f_j^*(z) = \mathbb{E}[f(X_1,\ldots, X_d) \mid X_j = z]$ using a new kernel-sieve approach that combines the local kernel regression with the B-spline basis approximation. We prove the rate of convergence for estimating $f_j^*$ under the supremum norm. We also propose two types of confidence bands for $f_j^*$ and illustrate their statistical-comptuational tradeoffs. Thorough numerical results on both synthetic data and real-world genomic data are provided to demonstrate the efficacy of the theory.