Closed TMats closed 6 years ago
いわゆるprednet 実装 https://github.com/coxlab/prednet
The architecture is implemented as a custom layer1 in Keras. Code and model data is now compatible with Keras 2.0. Specifically, it has been tested on Keras 2.0.6 with Theano 0.9.0, Tensorflow 1.2.1, and Python 2.7. The provided weights were trained with the Theano backend. For previous versions of the code compatible with Keras 1.2.1, use fbcdc18. To convert old PredNet model files and weights for Keras 2.0 compatibility, see convert_model_to_keras2 in keras_utils.py.
predictive codingからインスピレーションを受けている
動画予測
モジュールが階層的に積み重なる
各モジュールはinput convolutional layer,recurrent representation layer,prediction layer,error representationの4つの部分から構成
対象の動く向きなど,潜在変数の学習を確認
これとmodel-base RLは相性が良さそう
predictive coding系の論文
https://arxiv.org/abs/1605.08104