Canadian Institute for Advanced Research

About: Canadian Institute for Advanced Research is a facility organization based out in Toronto, Ontario, Canada. It is known for research contribution in the topics: Superconductivity & Gene. The organization has 839 authors who have published 3479 publications receiving 256281 citations.

Show, Attend and Tell: Neural Image Caption Generation with Visual Attention

Abstract: Inspired by recent work in machine translation and object detection, we introduce an attention based model that automatically learns to describe the content of images. We describe how we can train this model in a deterministic manner using standard backpropagation techniques and stochastically by maximizing a variational lower bound. We also show through visualization how the model is able to automatically learn to fix its gaze on salient objects while generating the corresponding words in the output sequence. We validate the use of attention with state-of-the-art performance on three benchmark datasets: Flickr9k, Flickr30k and MS COCO.

Improved Training of Wasserstein GANs

Abstract: Generative Adversarial Networks (GANs) are powerful generative models, but suffer from training instability. The recently proposed Wasserstein GAN (WGAN) makes progress toward stable training of GANs, but sometimes can still generate only low-quality samples or fail to converge. We find that these problems are often due to the use of weight clipping in WGAN to enforce a Lipschitz constraint on the critic, which can lead to undesired behavior. We propose an alternative to clipping weights: penalize the norm of gradient of the critic with respect to its input. Our proposed method performs better than standard WGAN and enables stable training of a wide variety of GAN architectures with almost no hyperparameter tuning, including 101-layer ResNets and language models over discrete data. We also achieve high quality generations on CIFAR-10 and LSUN bedrooms.

The Demography of massive dark objects in galaxy centers

Abstract: We construct dynamical models for a sample of 36 nearby galaxies with Hubble Space Telescope (HST) photometry and ground-based kinematics. The models assume that each galaxy is axisymmetric, with a two-integral distribution function, arbitrary inclination angle, a position-independent stellar mass-to-light ratio , and a central massive dark object (MDO) of arbitrary mass M•. They provide acceptable fits to 32 of the galaxies for some value of M• and ; the four galaxies that cannot be fitted have kinematically decoupled cores. The mass-to-light ratios inferred for the 32 well-fitted galaxies are consistent with the fundamental-plane correlation ∝ L0.2, where L is galaxy luminosity. In all but six galaxies the models require at the 95% confidence level an MDO of mass M• ~ 0.006Mbulge ≡ 0.006L. Five of the six galaxies consistent with M• = 0 are also consistent with this correlation. The other (NGC 7332) has a much stronger upper limit on M•. We predict the second-moment profiles that should be observed at HST resolution for the 32 galaxies that our models describe well. We consider various parameterizations for the probability distribution describing the correlation of the masses of these MDOs with other galaxy properties. One of the best models can be summarized thus: a fraction f 0.97 of early-type galaxies have MDOs, whose masses are well described by a Gaussian distribution in log (M•/Mbulge) of mean -2.28 and standard deviation ~0.51. There is also marginal evidence that M• is distributed differently for core and power law galaxies, with core galaxies having a somewhat steeper dependence on Mbulge.

Export Versus FDI with Heterogeneous Firms

Abstract: Multinational sales have grown at high rates over the last two decades, outpacing the remarkable expansion of trade in manufactures. Consequently, the trade literature has sought to incorporate the mode of foreign market access into the “new” trade theory. This literature recognizes that Ž rms can serve foreign buyers through a variety of channels: they can export their products to foreign customers, serve them through foreign subsidiaries, or license foreign Ž rms to produce their products. Our work focuses on the Ž rm’s choice between exports and “horizontal” foreign direct investment (FDI). Horizontal FDI refers to an investment in a foreign production facility that is designed to serve customers in the foreign market. Firms invest abroad when the gains from avoiding trade costs outweigh the costs of maintaining capacity in multiple markets. This is known as the proximity-concentration tradeoff. We introduce heterogeneous Ž rms into a simple multicountry, multisector model, in which Ž rms face a proximity-concentration trade-off. Every Ž rm decides whether to serve a foreign market, and whether to do so through exports or local subsidiary sales. These modes of market access have different relative costs: exporting involves lower Ž xed costs while FDI involves lower variable costs. Our model highlights the important role of within-sector Ž rm productivity differences in explaining the structure of international trade and investment. First, only the most productive Ž rms engage in foreign activities. This result mirrors other Ž ndings on Ž rm heterogeneity and trade; in particular, the results reported in Melitz (2003). Second, of those Ž rms that serve foreign markets, only the most productive engage in FDI. Third, FDI sales relative to exports are larger in sectors with more Ž rm heterogeneity. Using U.S. exports and afŽ liate sales data that cover 52 manufacturing sectors and 38 countries, we show that cross-sectoral differences in Ž rm heterogeneity predict the composition of trade and investment in the manner suggested by our model. We construct several measures of Ž rm heterogeneity, using different data sources, and show that our results are robust across all these measures. In addition, we conŽ rm the predictions of the proximityconcentration trade-off. That is, Ž rms tend to substitute FDI sales for exports when transport * Helpman: Department of Economics, Harvard University, Cambridge, MA 02138, Tel Aviv University, and CIAR (e-mail: ehelpman@harvard.edu); Melitz: Department of Economics, Harvard University, Cambridge, MA 02138, National Bureau of Economic Research, and Centre for Economic Policy Research (e-mail: mmelitz@ harvard.edu); Yeaple: Department of Economics, University of Pennsylvania, 3718 Locust Walk, Philadelphia, PA 19104, and National Bureau of Economic Research (e-mail: snyeapl2@ssc.upenn.edu). The statistical analysis of Ž rmlevel data on U.S. Multinational Corporations reported in this study was conducted at the International Investment Division, U.S. Bureau of Economic Analysis, under an arrangement that maintained legal conŽ dentiality requirements. Views expressed are those of the authors and do not necessarily re ect those of the Bureau of Economic Analysis. Elhanan Helpman thanks the NSF for Ž nancial support. We also thank Daron Acemoglu, Roberto Rigobon, Yona Rubinstein, and Dani Tsiddon for comments on an earlier draft, and Man-Keung Tang for excellent research assistance. 1 See Wilfred J. Ethier (1986), Ignatius Horstmann and James R. Markusen (1987), and Ethier and Markusen (1996) for models that incorporate the licensing alternative. We therefore exclude “vertical” motives for FDI that involve fragmentation of production across countries. See Helpman (1984, 1985), Markusen (2002, Ch. 9), and Gordon H. Hanson et al. (2002) for treatments of this form of FDI. 3 See, for example, Horstmann and Markusen (1992), S. Lael Brainard (1993), and Markusen and Anthony J. Venables (2000). 4 See also Andrew B. Bernard et al. (2003) for an alternative theoretical model and Yeaple (2003a) for a model based on worker-skill heterogeneity. James R. Tybout (2003) surveys the recent micro-level evidence on trade that has motivated these theoretical models. 5 This result is loosely connected to the documented empirical pattern that foreign-owned afŽ liates are more productive than domestically owned producers. See Mark E. Doms and J. Bradford Jensen (1998) for the United States and Sourafel Girma et al. (2002) for the United Kingdom.

International R&D Spillovers

Abstract: Investment in research and development (R&D) affects a country's total factor productivity. Recently new theories of economic growth have emphasized this link and have also identified a number of channels through which a country's R&D affects total factor productivity of its trade partners. Following these theoretical developments we estimate the effects of a country's R&D capital stock and the R&D capital stocks of its trade partners on the country's total factor productivity. We find large effects of both domestic and foreign R&D capital stocks on total factor productivity. The foreign R&D capital stocks have particularly large effects on the smaller countries in our sample (that consists of 22 countries). Moreover, we find that about one-quarter of the worldwide benefits of investment in R&D in the seven largest economies are appropriated by their trade partners.