In e-commerce, product content, especially product images have a significant influence on a customer's journey from product discovery to evaluation and finally, purchase decision. Since many e-commerce retailers sell items from other third-party marketplace sellers besides their own, the content published by both internal and external content creators needs to be monitored and enriched, wherever possible. Despite guidelines and warnings, product listings that contain offensive and non-compliant images continue to enter catalogs. Offensive and non-compliant content can include a wide range of objects, logos, and banners conveying violent, sexually explicit, racist, or promotional messages. Such images can severely damage the customer experience, lead to legal issues, and erode the company brand. In this paper, we present a computer vision driven offensive and non-compliant image detection system for extremely large image datasets. This paper delves into the unique challenges of applying deep learning to real-world product image data from retail world. We demonstrate how we resolve a number of technical challenges such as lack of training data, severe class imbalance, fine-grained class definitions etc. using a number of practical yet unique technical strategies. Our system combines state-of-the-art image classification and object detection techniques with budgeted crowd-sourcing to develop a solution customized for a massive, diverse, and constantly evolving product catalog.

The problem of optimal actuation for channel and source coding was recently formulated and solved in a number of relevant scenarios. In this class of models, actions are taken at encoders or decoders, either to acquire side information in an efficient way or to control or probe effectively the channel state. In this paper, the problem of embedding information on the actions is studied for both the source and the channel coding setups. In both cases, a decoder is present that observes only a function of the actions taken by an encoder or a decoder of an action-dependent point-to-point link. For the source coding model, this decoder wishes to reconstruct a lossy version of the source being transmitted over the point-to-point link, while for the channel coding problem, the decoder wishes to retrieve a portion of the message conveyed over the link. For the problem of source coding with actions taken at the decoder, a single letter characterization of the set of all achievable tuples of rate, distortions at the two decoders, and action cost is derived, under the assumption that the mentioned decoder observes a function of the actions noncausally, strictly causally or causally. A special case of the problem in which the actions are taken by the encoder is also solved. A single-letter characterization of the achievable capacity-cost region is then obtained for the channel coding setup with actions. Examples are provided that shed light into the effect of information embedding on the actions for the action-dependent source and channel coding problems.

The model of a side information vending machine (VM) accounts for scenarios in which the measurement of side information sequences can be controlled via the selection of cost-constrained actions. In this paper, the three-node cascade source coding problem is studied under the assumption that a side information VM is available at the intermediate and/or end node of the cascade. A single-letter characterization of the achievable tradeoff among the transmission rates, distortions in the reconstructions at the intermediate and end node, and cost for acquiring the side information is derived for a number of relevant special cases. It is shown that a joint design of the description, source, and control signals used to guide the selection of the actions at downstream nodes is generally necessary for an efficient use of the available communication links. In particular, for all the considered models, layered coding strategies prove to be optimal, whereby the base layer fulfills two network objectives: 1) determining the actions of downstream nodes and 2) simultaneously providing a coarse description of the source. Design of the optimal coding strategy is shown via examples to depend on both the network topology and action costs. Examples also illustrate the involved performance tradeoffs across the network.

The problem of optimal actuation for channel and source coding was recently formulated and solved in a number of relevant scenarios. In this class of models, actions are taken either to acquire side information in an efficient way for source coding, or to control or probe effectively the channel state for channel coding. In this paper, the problem of embedding information on the actions is introduced by considering the presence of an additional decoder that observes only a function of the actions. For the source coding model, this decoder wishes to reconstruct a lossy version of the source being transmitted over the point-to-point link, while, for the channel coding problem, the decoder wishes to retrieve a portion of the message conveyed over the link. In both cases, single-letter performance characterizations are provided for various special cases, along with specific examples.

A bidirectional link between two nodes, Node 1 and Node 2, is studied in which Node 2 is able to acquire information from the environment, e.g., via access to a remote data base or via sensing. Information acquisition is expensive in terms of system resources, e.g., time, bandwidth and energy, and thus should be done efficiently. As a result of the forward communication from Node 1 to Node 2, the latter wishes to obtain information about the the data available at Node 1 and of the data obtained from the environment. The forward link is, however, also used by Node 1 to query Node 2 with the aim of retrieving information from the environment on the backward link. The problem is formulated using the concept of side information “vending machine”, and the optimal trade-off between communication rates, distortions of the estimates produced at the two nodes and costs for information acquisition at Node 2 is derived.

Source coding with a side information “vending machine” is a recently proposed framework in which the statistical relationship between the side information and the source, instead of being given and fixed as in the classical Wyner–Ziv problem, can be controlled by the decoder. This control action is selected by the decoder based on the message encoded by the source node. Unlike conventional settings, the message can thus carry not only information about the source to be reproduced at the decoder, but also control information aimed at improving the quality of the side information. In this paper, the analysis of the tradeoffs between rate, distortion, and cost associated with the control actions is extended from the previously studied point-to-point setup to two basic multiterminal models. First, a distributed source coding model is studied, in which two encoders communicate over rate-limited links to a decoder, whose side information can be controlled. The control actions are selected by the decoder based on the messages encoded by both source nodes. For this setup, inner bounds are derived on the rate-distortion-cost region for both cases in which the side information is available causally and noncausally at the decoder. These bounds are shown to be tight under specific assumptions, including the scenario in which the sequence observed by one of the nodes is a function of the source observed by the other and the side information is available causally at the decoder. Then, a cascade scenario in which three nodes are connected in a cascade and the last node has controllable side information is also investigated. For this model, the rate-distortion-cost region is derived for general distortion requirements and under the assumption of causal availability of side information at the last node.

In lossy source coding with side information at the decoder (i.e., the Wyner–Ziv problem), the estimate of the source obtained at the decoder cannot be generally reproduced at the encoder, due to its dependence on the side information. In some applications, this may be undesirable, and a common reconstruction (CR) requirement, whereby one imposes that the encoder and decoder be able to agree on the decoder's estimate, may be instead in order. The rate-distortion function under the CR constraint has been derived recently for a point-to-point (Wyner–Ziv) problem. In this paper, this result is extended to three multiterminal settings with three nodes, namely the Heegard–Berger (HB) problem, its variant with cooperating decoders, and the cascade source coding problem. The HB problem consists of an encoder broadcasting to two decoders with respective side information. The cascade source coding problem is characterized by a two-hop system with side information available at the intermediate and final nodes. For the HB problem with the CR constraint, the rate-distortion function is derived under the assumption that the side information sequences are (stochastically) degraded. The rate-distortion function is also calculated explicitly for three examples, namely Gaussian source and side information with quadratic distortion metric, and binary source and side information with erasure and Hamming distortion metrics. The rate-distortion function is then characterized for the HB problem with cooperating decoders and (physically) degraded side information. For the cascade problem with the CR constraint, the rate-distortion region is obtained under the assumption that side information at the final node is physically degraded with respect to that at the intermediate node. For the latter two cases, it is worth emphasizing that the corresponding problem without the CR constraint is still open. Outer and inner bounds on the rate-distortion region are also obtained for the cascade problem under the assumption that the side information at the intermediate node is physically degraded with respect to that at the final node. For the three examples mentioned above, the bounds are shown to coincide. Finally, for the HB problem, the rate-distortion function is obtained under the more general requirement of constrained reconstruction, whereby the decoder's estimate must be recovered at the encoder only within some distortion.

Source coding with a side information "vending machine" is a recently proposed framework in which the statistical relationship between the side information available at the decoder and the source sequence can be controlled by the decoder based on the message received from the encoder. In this paper, the characterization of the optimal rate-distortion performance as a function of the cost associated with the control actions is extended from the previously studied point-to-point set-up to two multiterminal models. First, a distributed source coding model is studied, in which two encoders communicate over rate-limited links to a decoder, whose side information can be controlled based on the control actions selected by one of the encoders. The rate-distortion-cost region is characterized under the assumption of lossless reconstruction of the source encoded by the node that does not control the side information. Then, a three-node cascade scenario is investigated, in which the last node has controllable side information. The rate-distortion-cost region is derived for general distortion requirements and under the assumption of "causal" availability of side information at the last node.

In lossy source coding with side information at the decoder (i.e., the Wyner-Ziv problem), the estimate of the source obtained at the decoder cannot be generally reproduced at the encoder, due to its dependence on the side information. In some applications this may be undesirable, and a Common Reconstruction (CR) requirement, whereby one imposes that encoder and decoder be able to agree on the decoder's estimate, may be instead in order. The rate-distortion function under the CR constraint has been recently derived for the point-to-point (Wyner-Ziv) problem. In this paper, this result is extended to the Heegard-Berger (HB) problem and to its variant with cooperating decoders. Specifically, for the HB problem, the ratedistortion function is derived under the assumption that the side information sequences at the two decoders are stochastically degraded. The rate-distortion function is also calculated explicitly for the special case of binary source and erased side information with Hamming distortion metric. The rate-distortion function is then characterized also for the HB problem with cooperating decoders and physically degraded side information.

The problem of multiple descriptions for a Gaussian source is considered, in which all the decoders have access to correlated Gaussian side-information. Two variations of this problem are studied. First, the rate-distortion tradeoff is characterized under the assumption of a common reconstruction constraint, in which the estimate produced at each of the decoders is also exactly recoverable at the encoder. Secondly, a generalization of this setup is studied in which possibly different distortions are tolerable between the estimates at the encoder and the respective estimates at each of the decoders.

Action-dependent channels model scenarios in which transmission takes place in two successive phases. In the first phase, the encoder selects an “action” sequence, with the twofold aim of conveying information to the receiver and of affecting in a desired way the state of the channel to be used in the second phase. In the second phase, communication takes place in the presence the mentioned action-dependent state. In this work, two extensions of the original action-dependent channel are studied. In the first, the decoder is interested in estimating not only the message, but also the state sequence within an average per-letter distortion. Under the constraint of common reconstruction (i.e., the decoder's estimate of the state must be recoverable also at the encoder) and assuming non-causal state knowledge at the encoder in the second phase, we obtain a single-letter characterization of the achievable rate-distortion-cost trade-off. In the second extension, we study an action-dependent degraded broadcast channel. Under the assumption that the encoder knows the state sequence causally in the second phase, the capacity-cost region is identified. Various examples, including Gaussian channels and a model with a “probing” encoder, are also provided to show the advantage of a proper joint design of the two communication phases.

The model of a side information “vending machine” accounts for scenarios in which acquiring side information is costly and thus should be done efficiently. In this paper, the three-node cascade source coding problem is studied under the assumption that a side information vending machine is available either at the intermediate or at the end node. In both cases, a single-letter characterization of the available trade-offs among the rate, the distortions in the reconstructions at the intermediate and at the end node, and the cost in acquiring the side information are derived under given conditions.

1An encoder wishes to minimize the bit rate necessary to guarantee that a decoder is able to calculate a symbol-wise function of a sequence available only at the encoder and a sequence that can be measured only at the decoder. This classical problem, first studied by Yamamoto, is addressed here by including two new aspects: (i) The decoder obtains noisy measurements of its sequence, where the quality of such measurements can be controlled via a cost-constrained “action” sequence; (ii) Measurement at the decoder may fail in a way that is unpredictable to the encoder, thus requiring robust encoding. The considered scenario generalizes known settings such as the Heegard-Berger-Kaspi and the “source coding with a vending machine” problems. The rate-distortion-cost function is derived in relevant special cases, along with general upper and lower bounds. Numerical examples are also worked out to obtain further insight into the optimal system design.