TechTalks from event: ICML 2011

Neural Networks and Statistical Methods

  • Minimum Probability Flow Learning Authors: Jascha Sohl-Dickstein; Peter Battaglino; Michael DeWeese
    Fitting probabilistic models to data is often difficult, due to the general intractability of the partition function and its derivatives. Here we propose a new parameter estimation technique that does not require computing an intractable normalization factor or sampling from the equilibrium distribution of the model. This is achieved by establishing dynamics that would transform the observed data distribution into the model distribution, and then setting as the objective the minimization of the KL divergence between the data distribution and the distribution produced by running the dynamics for an infinitesimal time. Score matching, minimum velocity learning, and certain forms of contrastive divergence are shown to be special cases of this learning technique. We demonstrate parameter estimation in Ising models, deep belief networks and an independent component analysis model of natural scenes. In the Ising model case, current state of the art techniques are outperformed by at least an order of magnitude in learning time, with lower error in recovered coupling parameters.
  • The Importance of Encoding Versus Training with Sparse Coding and Vector Quantization Authors: Adam Coates; Andrew Ng
    While vector quantization (VQ) has been applied widely to generate features for visual recognition problems, much recent work has focused on more powerful methods. In particular, sparse coding has emerged as a strong alternative to traditional VQ approaches and has been shown to achieve consistently higher performance on benchmark datasets. Both approaches can be split into a training phase, where the system learns a dictionary of basis functions from unlabeled data, and an encoding phase, where the dictionary is used to extract features from new inputs. In this work, we investigate the reasons for the success of sparse coding over VQ by decoupling these phases, allowing us to separate out the contributions of the training and encoding in a controlled way. Through extensive experiments on CIFAR, NORB and Caltech 101 datasets, we compare sparse coding and several other training and encoding schemes, including a form of VQ paired with a soft threshold activation function. Our results show not only that we can use fast VQ algorithms for training without penalty, but that we can just as well use randomly chosen exemplars from the training set. Rather than spend resources on training, we find it is more important to choose a good encoder---which can often be as simple as a feed forward non-linearity. Among our results, we demonstrate state-of-the-art performance on both CIFAR and NORB.
  • Learning Recurrent Neural Networks with Hessian-Free Optimization Authors: James Martens; Ilya Sutskever
    In this work we resolve the long-outstanding problem of how to effectively train recurrent neural networks (RNNs) on complex and difficult sequence modeling problems which may contain long-term data dependencies. Utilizing recent advances in the Hessian-free optimization approach citep{hf}, together with a novel damping scheme, we successfully train RNNs on two sets of challenging problems. First, a collection of pathological synthetic datasets which are known to be impossible for standard optimization approaches (due to their extremely long-term dependencies), and second, on three natural and highly complex real-world sequence datasets where we find that our method significantly outperforms the previous state-of-the-art method for training neural sequence models: the Long Short-term Memory approach of citet{lstm}. Additionally, we offer a new interpretation of the generalized Gauss-Newton matrix of citet{schraudolph} which is used within the HF approach of Martens.
  • On Random Weights and Unsupervised Feature Learning Authors: Andrew Saxe; Pang Wei Koh; Zhenghao Chen; Maneesh Bhand; Bipin Suresh; Andrew Ng
    Recently two anomalous results in the literature have shown that certain feature learning architectures can yield useful features for object recognition tasks even with untrained, random weights. In this paper we pose the question: why do random weights sometimes do so well? Our answer is that certain convolutional pooling architectures can be inherently frequency selective and translation invariant, even with random weights. Based on this we demonstrate the viability of extremely fast architecture search by using random weights to evaluate candidate architectures, thereby sidestepping the time-consuming learning process. We then show that a surprising fraction of the performance of certain state-of-the-art methods can be attributed to the architecture alone.