The thresholded t-map produced by the General Linear Model (GLM) gives an effective summary of activation patterns in functional brain images and is widely used for feature selection in fMRI related classification tasks. As part of a project to build content-based retrieval systems for fMRI images, we have investigated ways to make GLM more adaptive and more robust in dealing with fMRI data from widely differing experiments. In this paper we report on exploration of the Finite Impulse Response model, combined with multiple linear regression, to identify the “locally best Hemodynamic Response Function (HRF) for each voxel” and to simultaneously estimate activation levels corresponding to several stimulus conditions. The goal is to develop a procedure for processing datasets of varying natures. Our experiments show that Finite Impulse Response (FIR) models with a smoothing factor produce better retrieval performance than does the canonical double gamma HRF in terms of retrieval accuracy.