Deep neural networks can model images with rich latent representations, but they cannot naturally conceptualize structures of object categories in a human-perceptible way. This paper addresses the problem of learning object structures in an image modeling process without supervision. We propose an autoencoding formulation to discover landmarks as explicit structural representations. The encoding module outputs landmark coordinates, whose validity is ensured by constraints that reflect the necessary properties for landmarks. The decoding module takes the landmarks as a part of the learnable input representations in an end-to-end differentiable framework. Our discovered landmarks are semantically meaningful and more predictive of manually annotated landmarks than those discovered by previous methods. The coordinates of our landmarks are also complementary features to pretrained deep-neural-network representations in recognizing visual attributes. In addition, the proposed method naturally creates an unsupervised, perceptible interface to manipulate object shapes and decode images with controllable structures.

The Internet distance map service (IDMaps) [P. Francis, S. Jamin, C. Jin, D. Raz, Y. Shavitt, and L. Zhang, 2001] provides distance estimates between any pair of hosts connected to the Internet. The IDMaps system comprises two component types: tracers that measure distance between IP address prefixes, and servers that collect measurement results and answer distance queries. The distance information protocol (DIP) is used for tracers to report measured distance data to servers. The dynamics on the Internet topology, the distributed nature of autonomous tracers and servers, and the vast size of the data set require that DIP provide highly adaptive and scalable data dissemination from tracers to servers. DIP is a soft-state announce/listen protocol and scales independently from the total amount of measurement data by all tracers. DIP achieves its scalability through combination of staged timers, positive feedback, and feedback suppression techniques, which enable DIP to disseminate only the most useful measurement data to servers in a dynamic way. Simulations verified DIP's scalability and adaptability under various network conditions.

There is an increasing need to quickly and efficiently learn network distances, in terms of metrics such as latency or bandwidth, between Internet hosts. For example, Internet content providers often place data and server mirrors throughout the Internet to improve access latency for clients, and it is necessary to direct clients to the nearest mirrors based on some distance metric in order to realize the benefit of the mirrors. We suggest a scalable Internet-wide architecture, called IDMaps, which measures and disseminates distance information on the global Internet. Higher level services can collect such distance information to build a virtual distance map of the Internet and estimate the distance between any pair of IP addresses. We present our solutions to the measurement server placement and distance map construction problems in IDMaps. We show that IDMaps can indeed provide useful distance estimations to applications such as nearest mirror selection.

The IDMaps project aims to provide a distance map of the Internet from which relative distances between hosts on the Internet can be gauged. Many distributed systems and applications can benefit from such a distance map service, for example, a common method to improve user-perceived performance of the Internet is to place data and server mirrors closer to clients. When a client tries to access a mirrored server, which mirror should it access? With IDMaps, the closest mirror can be determined based on distance estimates between the client and the mirrors. In this paper we investigate both graph theoretic methods and ad hoc heuristics for instrumenting the Internet to obtain distance maps. We evaluate the efficacy of the resulting distance maps by comparing the determinations of the closest replica using known topologies against those obtained using the distance maps.

There is an increasing need for Internet hosts to be able to quickly and efficiently learn the distance, in terms of metrics such as latency or bandwidth, between Internet hosts. For example, to select the nearest of multiple equal content Web servers. This paper explores technical issues related to the creation of a public infrastructure service to provide such information. In so doing, we suggest an architecture, called IDMaps, whereby Internet distance information is distributed over the Internet, using IP multicast groups, in the form of a virtual distance map. Systems listening to the groups can estimate the distance between any pair of IP addresses by running a spanning tree algorithm over the received distance map. We also presents the results of experiments that give preliminary evidence supporting the architecture. This work thus lays the initial foundation for future work in this new area.