We address the problem of online model adaptation when learning representations from non-stationary data streams. For now, we focus on single hidden-layer sparse linear autoencoders (i.e. sparse dictionary learning), although in the future, the proposed approach can be extended naturally to general multi-layer autoencoders and supervised models. We propose a simple but effective online model-selection, based on alternating-minimization scheme, which involves “birth” (addition of new elements) and “death” (removal, via l1/l2 group sparsity) of hidden units representing dictionary elements, in response to changing inputs; we draw inspiration from the adult neurogenesis phenomenon in the dentate gyrus of the hippocampus, known to be associated with better adaptation to new environments. Empirical evaluation on both real-life and synthetic data demonstrates that the proposed approach can considerably outperform the state-of-art non-adaptive online sparse coding of Mairal et al. (2009) in the presence of non-stationary data, especially when dictionaries are sparse.