Recommender systems aim to find an accurate and efficient mapping from historic data of user-preferred items to a new item that is to be liked by a user. Towards this goal, energy-based sequence generative adversarial nets (EB-SeqGANs) are adopted for recommendation by learning a generative model for the time series of user-preferred items. By recasting the energy function as the feature function, the proposed EB-SeqGANs is interpreted as an instance of maximum-entropy imitation learning.

Estimating the log-likelihood gradient with respect to the parameters of a Restricted Boltzmann Machine (RBM) typically requires sampling using Markov Chain Monte Carlo (MCMC) techniques. To save computation time, the Markov chains are only run for a small number of steps, which leads to a biased estimate. This bias can cause RBM training algorithms such as Contrastive Divergence (CD) learning to deteriorate. We adopt the idea behind Population Monte Carlo (PMC) methods to devise a new RBM training algorithm termed Population-Contrastive-Divergence (pop-CD). Compared to CD, it leads to a consistent estimate and may have a significantly lower bias. Its computational overhead is negligible compared to CD. However, the variance of the gradient estimate increases. We experimentally show that pop-CD can significantly outperform CD. In many cases, we observed a smaller bias and achieved higher log-likelihood values. However, when the RBM distribution has many hidden neurons, the consistent estimate of pop-CD may still have a considerable bias and the variance of the gradient estimate requires a smaller learning rate. Thus, despite its superior theoretical properties, it is not advisable to use pop-CD in its current form on large problems.

Andrej Karpathy's post "The Unreasonable Effectiveness of Recurrent Neural Networks" made splashes last year. The basic premise is that you can create a recurrent neural network to learn language features character-by-character. But is the resultant model any different from a Markov chain built for the same purpose? I implemented a character-by-character Markov chain in R to find out. First, let's play a variation of the Imitation Game with generated text from Karpathy's tinyshakespeare dataset.

As members of the deep learning R&D team at SVDS, we are interested in comparing Recurrent Neural Network (RNN) and other approaches to speech recognition. Until a few years ago, the state-of-the-art for speech recognition was a phonetic-based approach including separate components for pronunciation, acoustic, and language models. Typically, this consists of n-gram language models combined with Hidden Markov models (HMM). We wanted to start with this as a baseline model, and then explore ways to combine it with newer approaches such as Baidu's Deep Speech. While summaries exist explaining these baseline phonetic models, there do not appear to be any easily-digestible blog posts or papers that compare the tradeoffs of the different freely available tools.

In this article, we extend the conventional framework of convolutional-Restricted-Boltzmann-Machine to learn highly abstract features among abitrary number of time related input maps by constructing a layer of multiplicative units, which capture the relations among inputs. In many cases, more than two maps are strongly related, so it is wise to make multiplicative unit learn relations among more input maps, in other words, to find the optimal relational-order of each unit. In order to enable our machine to learn relational order, we developed a reinforcement-learning method whose optimality is proven to train the network.

In this article, a semi-supervised classification algorithm implementation will be described using Markov Chains and Random Walks. We have the following 2D circles dataset (with 1000 points) with only 2 points labeled (as shown in the figure, colored red and blue respectively, for all others the labels are unknown, indicated by the color black). Now the task is to predict the labels of the other (unlabeled) points. From each of the unlabeled points (Markov states) a random walk with Markov transition matrix (computed from the row-stochastic kernelized distance matrix) will be started that will end in one labeled state, which will be an absorbing state in the Markov Chain. This problem was discussed as an application of Markov Chain in a lecture from the edX course ColumbiaX: CSMM.102x

An ensemble of models (EM), where each model is constructed on a diverse subset of feature variables, is proposed to rank rare class items ahead of majority class items in a highly unbalanced two class problem. The proposed ensemble relies on an algorithm to group the feature variables into subsets where the variables in a subset work better together in a model and the variables in different subsets work better in separate models. The strength of the EM depends on the algorithm's ability to identify strong and diverse subsets of feature variables. A second phase of ensembling is achieved by aggregating several EMs each optimized on a diverse evaluation metric. The resulting ensemble is called ensemble of models and metrics (EMM). Here, the diverse/complementary evaluation metrics ensure increased diversity among EMs to aggregate. The ensembles are applied to the protein homology data, downloaded from the 2004 KDD cup competition website, to rank proteins in such a way that the rare homologous proteins are found ahead of the majority non-homologous proteins. The ensembles are constructed using feature variables which are various scores from sequence alignments of amino acids in a candidate protein and three dimensional descriptions of a native protein representing functional and structural similarity of proteins. While prediction performances of the EMs are better than the contemporary state-of-the-art ensembles and competitive to the winning procedures of the $2004$ KDD cup competition, the performances of the EMM are found on the top of all. In this application, we have two diverse EMs constructed on two complementary evaluation metrics average precision and rank last, where the former is robust against ranking close homologs and the latter is robust against ranking distant homologs. The advantage of using EMM is that it is robust against both close and distant homologs.

Anima Anandkumar pioneered the research of finding global optimal in non-convex problems, a big pain point in machine learning. Our protagonist for this week's Techie Tuesdays, Anima is an academician who represents the best of both worlds--industry and academia. She has contributed significantly to major AI and ML projects at Amazon. This will be a treat for all machine learning enthusiasts. In my two hours of conversation with Anima Anandkumar, Principal Scientist at Amazon Web Services, I've had the most potent dose of technical knowledge ever injected. Not that I didn't expect it while talking to an ex-faculty of UC Irvine (soon to be an endowed professor at Caltech), known for her research on non-convex problems (in deep learning). Our Techie Tuesdays protagonist of the week, Anima has worked towards establishing a strong collaboration between academia and industry. She follows an unconventional style of teaching, the one she would have loved as a student.

Observable operator models (OOMs) and related models are one of the most important and powerful tools for modeling and analyzing stochastic systems. They exactly describe dynamics of finite-rank systems and can be efficiently and consistently estimated through spectral learning under the assumption of identically distributed data. In this paper, we investigate the properties of spectral learning without this assumption due to the requirements of analyzing large-time scale systems, and show that the equilibrium dynamics of a system can be extracted from nonequilibrium observation data by imposing an equilibrium constraint. In addition, we propose a binless extension of spectral learning for continuous data. In comparison with the other continuous-valued spectral algorithms, the binless algorithm can achieve consistent estimation of equilibrium dynamics with only linear complexity.

We propose a Bayesian nonparametric mixture model for prediction- and information extraction tasks with an efficient inference scheme. It models categorical-valued time series that exhibit dynamics from multiple underlying patterns (e.g. user behavior traces). We simplify the idea of capturing these patterns by hierarchical hidden Markov models (HHMMs) - and extend the existing approaches by the additional representation of structural information. Our empirical results are based on both synthetic- and real world data. They indicate that the results are easily interpretable, and that the model excels at segmentation and prediction performance: it successfully identifies the generating patterns and can be used for effective prediction of future observations.