Optimising the approximation of Average Precision (AP) has been widely studied for retrieval. Such methods consider both negative and positive instances before each target positive one according to the definition of AP. However, we argue that only penalizing negative instances before positive ones is enough, because the loss only comes from them. To this end, instead of following the AP-based loss, we propose a new loss, namely Penalizing Negative instances before Positive ones (PNP), which directly minimizes the number of negative instances before each positive one. Meanwhile, limited by the definition of AP, AP-based methods only adopt a specific gradient assignment strategy. We wonder whether there exists better ones. Instead, we systematically investigate different gradient assignment solutions via constructing derivative functions of the loss, resulting in PNP-I with increasing derivative functions and PNP-D with decreasing ones. Because of their gradient assignment strategies, PNP-I tries to make all the relevant instances together, while PNP-D only quickly corrects positive one with fewer negative instances before. Thus, PNP-D may be more suitable for real-world data, which usually contains several local clusters for one class. Extensive evaluations on three standard retrieval datasets also show that PNP-D achieves the state-of-the-art performance.

An original Apple computer with a wooden case is up for sale for £1.1 million ($1.5 million) on eBay -- some 2,250 times more than its original price tag in 1976. The'Apple-1' was the first product to be developed under the Apple name by company co-founders Steve Jobs and Steve Wozniak and launched in 1976. Around 175 of 200 Apple-1 machines were sold in total, each carrying a price tag of $666.66 (equivalent to some $3,126 today.) The fully-functional model is being sold by Krishna Blake of the US, who purchased the machine in 1978, and comes with its manuals and a cassette interface. Also included in the sale is a period Sony TV-115, which was the monitor model originally recommended by Mr Jobs to use to display the computer's output.

This paper proposes MCSSL, a self-supervised learning approach for building custom object detection models in multi-camera networks. MCSSL associates bounding boxes between cameras with overlapping fields of view by leveraging epipolar geometry and state-of-the-art tracking and reID algorithms, and prudently generates two sets of pseudo-labels to fine-tune backbone and detection networks respectively in an object detection model. To train effectively on pseudo-labels,a powerful reID-like pretext task with consistency loss is constructed for model customization. Our evaluation shows that compared with legacy selftraining methods, MCSSL improves average mAP by 5.44% and 6.76% on WildTrack and CityFlow dataset, respectively.

In many applications of machine learning, such as recommender systems, where an input characterizing a user should be matched with a target representing an ordering of a large number of items, accessing fully supervised data (,) is not an option. Instead, one should expect weak information on the target, which could be a list of previously taken (if items are online courses), watched (if items are plays), etc., items by a user characterized by the feature vector. This motivates weakly supervised learning, aiming at learning a mapping from inputs to targets in such a setting where tools from supervised learning can not be applied off-the-shelves. Recent applications of weakly supervised learning showcase impressive results in solving complex tasks such as action retrieval on instructional videos (Miech et al., 2019), image semantic segmentation (Papandreou et al., 2015), salient object detection (Wang et al., 2017), 3D pose estimation (Dabral et al., 2018), text-to-speech synthesis (Jia et al., 2018), to name a few. However, those applications of weakly supervised learning are usually based on clever heuristics, and theoretical foundations of learning from weakly supervised data are scarce, especially when compared to statistical learning literature on supervised learning (Vapnik, 1995; Boucheron et al., 2005; Steinwart and Christmann, 2008). We aim to provide a step in this direction. In this paper, we focus on partial labelling, a popular instance of weak supervision, approached with a structured prediction point of view Ciliberto et al. (2020). We detail this setup in Section 2. Our contributions are organized as follows.

Identifying user intents from natural language utterances is a crucial step in conversational systems that has been extensively studied as a supervised classification problem. However, in practice, new intents emerge after deploying an intent detection model. Thus, these models should seamlessly adapt and classify utterances with both seen and unseen intents -- unseen intents emerge after deployment and they do not have training data. The few existing models that target this setting rely heavily on the scarcely available training data and overfit to seen intents data, resulting in a bias to misclassify utterances with unseen intents into seen ones. We propose RIDE: an intent detection model that leverages commonsense knowledge in an unsupervised fashion to overcome the issue of training data scarcity. RIDE computes robust and generalizable relationship meta-features that capture deep semantic relationships between utterances and intent labels; these features are computed by considering how the concepts in an utterance are linked to those in an intent label via commonsense knowledge. Our extensive experimental analysis on three widely-used intent detection benchmarks shows that relationship meta-features significantly increase the accuracy of detecting both seen and unseen intents and that RIDE outperforms the state-of-the-art model for unseen intents.

Data-to-Text Generation (DTG) is a subfield of Natural Language Generation aiming at transcribing structured data in natural language descriptions. The field has been recently boosted by the use of neural-based generators which exhibit on one side great syntactic skills without the need of hand-crafted pipelines; on the other side, the quality of the generated text reflects the quality of the training data, which in realistic settings only offer imperfectly aligned structure-text pairs. Consequently, state-of-art neural models include misleading statements - usually called hallucinations - in their outputs. The control of this phenomenon is today a major challenge for DTG, and is the problem addressed in the paper. Previous work deal with this issue at the instance level: using an alignment score for each table-reference pair. In contrast, we propose a finer-grained approach, arguing that hallucinations should rather be treated at the word level. Specifically, we propose a Multi-Branch Decoder which is able to leverage word-level labels to learn the relevant parts of each training instance. These labels are obtained following a simple and efficient scoring procedure based on co-occurrence analysis and dependency parsing. Extensive evaluations, via automated metrics and human judgment on the standard WikiBio benchmark, show the accuracy of our alignment labels and the effectiveness of the proposed Multi-Branch Decoder. Our model is able to reduce and control hallucinations, while keeping fluency and coherence in generated texts. Further experiments on a degraded version of ToTTo show that our model could be successfully used on very noisy settings.

In many applications of machine learning, certain categories of examples may be underrepresented in the training data, causing systems to underperform on such "few-shot" cases at test time. A common remedy is to perform data augmentation, such as by duplicating underrepresented examples, or heuristically synthesizing new examples. But these remedies often fail to cover the full diversity and complexity of real examples. We propose a data augmentation approach that performs neural Example Extrapolation (Ex2). Given a handful of exemplars sampled from some distribution, Ex2 synthesizes new examples that also belong to the same distribution. The Ex2 model is learned by simulating the example generation procedure on data-rich slices of the data, and it is applied to underrepresented, few-shot slices. We apply Ex2 to a range of language understanding tasks and significantly improve over state-of-the-art methods on multiple few-shot learning benchmarks, including for relation extraction (FewRel) and intent classification + slot filling (SNIPS).

The aim of few-shot learning (FSL) is to learn how to recognize image categories from a small number of training examples. A central challenge is that the available training examples are normally insufficient to determine which visual features are most characteristic of the considered categories. To address this challenge, we organize these visual features into facets, which intuitively group features of the same kind (e.g. features that are relevant to shape, color, or texture). This is motivated from the assumption that (i) the importance of each facet differs from category to category and (ii) it is possible to predict facet importance from a pre-trained embedding of the category names. In particular, we propose an adaptive similarity measure, relying on predicted facet importance weights for a given set of categories. This measure can be used in combination with a wide array of existing metric-based methods. Experiments on miniImageNet and CUB show that our approach improves the state-of-the-art in metric-based FSL.

Discrete supervised learning problems such as classification are often tackled by introducing a continuous surrogate problem akin to regression. Bounding the original error, between estimate and solution, by the surrogate error endows discrete problems with convergence rates already shown for continuous instances. Yet, current approaches do not leverage the fact that discrete problems are essentially predicting a discrete output when continuous problems are predicting a continuous value. In this paper, we tackle this issue for general structured prediction problems, opening the way to "super fast" rates, that is, convergence rates for the excess risk faster than $n^{-1}$, where $n$ is the number of observations, with even exponential rates with the strongest assumptions. We first illustrate it for predictors based on nearest neighbors, generalizing rates known for binary classification to any discrete problem within the framework of structured prediction. We then consider kernel ridge regression where we improve known rates in $n^{-1/4}$ to arbitrarily fast rates, depending on a parameter characterizing the hardness of the problem, thus allowing, under smoothness assumptions, to bypass the curse of dimensionality.

The need to evaluate treatment effectiveness is ubiquitous in most of empirical science, and interest in flexibly investigating effect heterogeneity is growing rapidly. To do so, a multitude of model-agnostic, nonparametric meta-learners have been proposed in recent years. Such learners decompose the treatment effect estimation problem into separate sub-problems, each solvable using standard supervised learning methods. Choosing between different meta-learners in a data-driven manner is difficult, as it requires access to counterfactual information. Therefore, with the ultimate goal of building better understanding of the conditions under which some learners can be expected to perform better than others a priori, we theoretically analyze four broad meta-learning strategies which rely on plug-in estimation and pseudo-outcome regression. We highlight how this theoretical reasoning can be used to guide principled algorithm design and translate our analyses into practice by considering a variety of neural network architectures as base-learners for the discussed meta-learning strategies. In a simulation study, we showcase the relative strengths of the learners under different data-generating processes.