Weakly-supervised word learning is improved by an active online algorithm

When infants learn words, they do not generally occur in isola- tion but as parts of continuous utterances and with several pos- sible related meanings. Details of the efficient learning techniques of infants remain unknown. In this paper we intro- duce a dynamic concept matrix (DCM) algorithm that learns acoustic models for a set of keywords from given pairs of con- tinuous utterances and related keyword labels. DCM is an in- cremental, active online algorithm. Specifically, each training utterance is first recognized with the current word models, and the recognition result is used to guide training further. In low- noise conditions DCM shows significant improvement in con- vergence rate and final recognition scores to an earlier passive CM model on TIDIGITS and CAREGIVER UK Y2 datasets. The results suggest that in ambiguous learning situations it may be beneficial for the learner to observe the learning situa- tion, make hard decisions if some known words/objects were recognized and update the models based on the decisions.