In this study, we investigate an estimation of internal sources with poor information about their source profile, which is important in order to analyze the higher functions of the human brain. We used a 64-channel whole-head-type SQUID system to record MEG activities evoked by visually cognitive tasks. Three kinds of stimuli were presented: English words, nonsense words, and random dot patterns. Spatially distributed internal sources with latencies between 320 and 420 msec were estimated by using a sub-optimal least-squares subspace scanning technique. From three subjects data, the following results of distributed source estimations were shown: English words stimuli activated large area, the left temporal region, the parietal and right postero-temporal region, nonsense words activated the paretal region, random dot patterns activated the frontal region.