April 28, 2015
MERL researcher and speech team leader, John Hershey, gave a talk at MIT entitled, “Deep Unfolding: Deriving Novel Deep Network Architectures from Model-based Inference Methods” on April 28, 2015.
Abstract: Model-based methods and deep neural networks have both been tremendously successful paradigms in machine learning. In model-based methods, problem domain knowledge can be built into the constraints of the model, typically at the expense of difficulties during inference. In contrast, deterministic deep neural networks are constructed in such a way that inference is straightforward, but their architectures are rather generic and it can be unclear how to incorporate problem domain knowledge. This work aims to obtain some of the advantages of both approaches. To do so, we start with a model-based approach and unfold the iterations of its inference method to form a layer-wise structure. This results in novel neural-network-like architectures that incorporate our model-based constraints, but can be trained discriminatively to perform fast and accurate inference. This framework allows us to view conventional sigmoid networks as a special case of unfolding Markov random field inference, and leads to other interesting generalizations. We show how it can be applied to other models, such as non-negative matrix factorization, to obtain a new kind of non-negative deep neural network that can be trained using a multiplicative back propagation-style update algorithm. In speech enhancement experiments we show that our approach is competitive with conventional neural networks, while using fewer parameters.