Manzil Zaheer

Gaussian LDA for Topic Models with Word Embeddings

Continuous space word embeddings learned from large, unstructured corpora have been shown to be effective at capturing semantic regularities in language. In this paper we replace LDA’s parameterization of “topics” as categorical distributions over opaque word types with multivariate Gaussian distributions on the embedding space. This encourages the model to group words that are a priori known to be semantically related into topics. To perform inference, we introduce a fast collapsed Gibbs sampling algorithm based on Cholesky decompositions of covariance matrices of the posterior predictive distributions. We further derive a scalable algorithm that draws samples from stale posterior predictive distributions and corrects them with a Metropolis–Hastings step. Using vectors learned from a domain-general corpus (English Wikipedia), we report results on two document collections (20-newsgroups and NIPS). Qualitatively, Gaussian LDA infers different (but still very sensible) topics relative to standard LDA. Quantitatively, our technique outperforms existing models at dealing with OOV words in held-out documents.

Co-Learning Bayesian Model Fusion: Efficient Performance Modeling of Analog and Mixed-Signal Circuits Using Side Information

Efficient performance modeling of todays analog and mixed-signal (AMS) circuits is an important yet challenging task. In this paper, we propose a novel performance modeling algorithm that is referred to as Co-Learning Bayesian Model Fusion (CL-BMF). The key idea of CL-BMF is to take advantage of the additional information collected from simulation and/or measurement to reduce the performance modeling cost. Different from the traditional performance modeling approaches which focus on the prior information of model coefficients (i.e. the coefficient side information) only, CL-BMF takes advantage of another new form of prior knowledge: the performance side information. In particular, CL-BMF combines the coefficient side information, the performance side information and a small number of training samples through Bayesian inference based on a graphical model. Two circuit examples designed in a commercial 32nm SOI CMOS process demonstrate that CL-BMF achieves up to 5× speed-up over other state-of-the-art performance modeling techniques without surrendering any accuracy.

Question Answering on Knowledge Bases and Text using Universal Schema and Memory Networks.

Existing question answering methods infer answers either from a knowledge base or from raw text. While knowledge base (KB) methods are good at answering compositional questions, their performance is often affected by the incompleteness of the KB. Au contraire, web text contains millions of facts that are absent in the KB, however in an unstructured form. {\it Universal schema} can support reasoning on the union of both structured KBs and unstructured text by aligning them in a common embedded space. In this paper we extend universal schema to natural language question answering, employing \emph{memory networks} to attend to the large body of facts in the combination of text and KB. Our models can be trained in an end-to-end fashion on question-answer pairs. Evaluation results on \spades fill-in-the-blank question answering dataset show that exploiting universal schema for question answering is better than using either a KB or text alone. This model also outperforms the current state-of-the-art by 8.5

Multiple-Population Moment Estimation: Exploiting Interpopulation Correlation for Efficient Moment Estimation in Analog/Mixed-Signal Validation

F_1

mTunes: efficient post-silicon tuning of mixed-signal/RF integrated circuits based on Markov decision process

points.\footnote{Code and data available in \url{this https URL}}

MPME-DP: multi-population moment estimation via dirichlet process for efficient validation of analog/mixed-signal circuits

Moment estimation is an important problem during circuit validation, in both presilicon and postsilicon stages. From the estimated moments, the probability of failure and parametric yield can be estimated at each circuit configuration and corner, and these metrics are used for design optimization and making product qualification decisions. The problem is especially difficult if only a very small sample size is allowed for measurement or simulation, as is the case for complex analog/mixed-signal circuits. In this paper, we propose an efficient moment estimation method, called multiple-population moment estimation (MPME), that significantly improves estimation accuracy under small sample size. The key idea is to leverage the data collected under different corners/configurations to improve the accuracy of moment estimation at each individual corner/configuration. Mathematically, we employ the hierarchical Bayesian framework to exploit the underlying correlation in the data. We apply the proposed method to several datasets including postsilicon measurements of a commercial high-speed I/O link, and demonstrate an average error reduction of up to 2×, which can be equivalently translated to significant reduction of validation time and cost.

Deep Sets

Uncertainty prevails in IC manufacturing and circuit operation. In particular, process variability has a huge impact on circuit performance, especially for mixed-signal/RF circuits, leading to unacceptable yields. Additionally, environmental uncertainties, such as temperature fluctuation and channel variation, further deteriorate performances in field. To combat variability, circuits are often made reconfigurable by adding tunable knobs to recover circuit performance in the post-manufacturing stage. However, as the number of knobs increases, knob tuning becomes challenging due to the huge search space. In fact, knob-tuning policies can have an observable impact on final performance and power consumption. In this paper, we propose mTunes, a method based on the Markov decision process for dynamically choosing the “right” knob tuning sub-routine from a pre-defined set achieving a balance between performance and power constraints. The proposed method has been applied to a reconfigurable RF front-end design, showing 60% improvement in yield compared to static tuning policies.

A Generic Approach for Escaping Saddle points

Moment estimation is one of the most important tasks to appropriately characterize the performance variability of todays nanoscale integrated circuits. In this paper, we propose an efficient algorithm of multi-population moment estimation via Dirichlet Process (MPME-DP) for validation of analog and mixed-signal circuits with extremely small sample size. The key idea is to partition all populations (e.g., different environmental conditions, setup configurations, etc.) into groups. The populations within the same group are similar and their common knowledge can be extracted to improve the accuracy of moment estimation. As will be demonstrated by the silicon measurement data of a high-speed I/O link, MPME-DP reduces the moment estimation error by up to 65% compared to other conventional estimators.