Speaker: Xue Gu Title: Predictive Coding Model Abstract: Neurons in cortical layers 2 and 3 have endstopping property: neurons can respond vigorously to a specific line segment, and this response is reduced for eleminated when the same stimulus goes beyond the neuron's receptive field. Rao & Ballard has shown that this property may result from the predictive coding of natural image. A hierarchical model of predictive coding is used to simulate this property. Neurons that are trained can successfully reconstruct natural images. How does it work with 3-D motion environment? Optical flow will be generated from a virtual environment for convenience. The data then will be fed into the model. Speaker: Shan He Title: Speech Rocognition: Acoustic Modeling with Graphemes Abstract: In speech recognition, acoustic models play an important role -- capturing the acoustic properties of speech and providing the probability of the observed acoustic signal given a hypothesized word sequence. Traditionally, phonemes are used as trainind units in acoustic modeling. In our experiment, we try to build a new acoustic model with graphemes, which is the sub-word unit usually used in Text-to-Speech synthesis. In the presentation, we will describe the experiment procedure and evaluate the result.