Collaborating Authors

Deterministic Single-Pass Algorithm for LDA

Neural Information Processing Systems

We develop a deterministic single-pass algorithm for latent Dirichlet allocation (LDA) in order to process received documents one at a time and then discard them in an excess text stream. Our algorithm does not need to store old statistics for all data. The proposed algorithm is much faster than a batch algorithm and is comparable to the batch algorithm in terms of perplexity in experiments. Papers published at the Neural Information Processing Systems Conference.

Algorithms for Generating Ordered Solutions for Explicit AND/OR Structures

Journal of Artificial Intelligence Research

The proposed algorithms use a best first search technique and report the solutions using an implicit representation ordered by cost. In this paper, we present two versions of the search algorithm -- (a) an initial version of the best first search algorithm, ASG, which may present one solution more than once while generating the ordered solutions, and (b) another version, LASG, which avoids the construction of the duplicate solutions. The actual solutions can be reconstructed quickly from the implicit compact representation used. We have applied the methods on a few test domains, some of them are synthetic while the others are based on well known problems including the search space of the 5-peg Tower of Hanoi problem, the matrix-chain multiplication problem and the problem of finding secondary structure of RNA. Experimental results show the efficacy of the proposed algorithms over the existing approach. Our proposed algorithms have potential use in various domains ranging from knowledge based frameworks to service composition, where the AND/OR structure is widely used for representing problems.

The Language of Search

Journal of Artificial Intelligence Research

This paper is concerned with a class of algorithms that perform exhaustive search on propositional knowledge bases. We show that each of these algorithms defines and generates a propositional language. Specifically, we show that the trace of a search can be interpreted as a combinational circuit, and a search algorithm then defines a propositional language consisting of circuits that are generated across all possible executions of the algorithm. In particular, we show that several versions of exhaustive DPLL search correspond to such well-known languages as FBDD, OBDD, and a precisely-defined subset of d-DNNF. By thus mapping search algorithms to propositional languages, we provide a uniform and practical framework in which successful search techniques can be harnessed for compilation of knowledge into various languages of interest, and a new methodology whereby the power and limitations of search algorithms can be understood by looking up the tractability and succinctness of the corresponding propositional languages.

Can an AI Recommend the Best Algorithm for Me?


So, when I need artificial intelligence to automate my business, why can't I get recommendations for which machine learning algorithm best suits my individual needs? Every business is unique, and there are hundreds of algorithms available, each one with individual strengths and weaknesses. Just like I don't look at every individual book when choosing which one to read, I don't have the time, resources, or knowledge to try out each and every algorithm. I want artificial intelligence to recommend a short list of algorithms for me to try on my data.

Parallelized Stochastic Gradient Descent

Neural Information Processing Systems

With the increase in available data parallel machine learning has become an increasingly pressing problem. In this paper we present the first parallel stochastic gradient descent algorithm including a detailed analysis and experimental evidence. Unlike prior work on parallel optimization algorithms our variant comes with parallel acceleration guarantees and it poses no overly tight latency constraints, which might only be available in the multicore setting. Our analysis introduces a novel proof technique --- contractive mappings to quantify the speed of convergence of parameter distributions to their asymptotic limits. As a side effect this answers the question of how quickly stochastic gradient descent algorithms reach the asymptotically normal regime.