DOCSLIB.ORG

On the Calibration and Uncertainty of Neural Learning to Rank Models for Conversational Search

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
On the Calibration and Uncertainty of Neural Learning to Rank Models for Conversational Search On the Calibration and Uncertainty of Neural Learning to Rank Models for Conversational Search Gustavo Penha Claudia Hauff Delft University of Technology Delft University of Technology Delft Delft Netherlands Netherlands [email protected] [email protected] Abstract Deterministic ranker Stochastic ranker According to the Probability Ranking Princi- ple (PRP), ranking documents in decreasing or- 0.1 der of their probability of relevance leads to an 0.9 optimal document ranking for ad-hoc retrieval. 0.7 The PRP holds when two conditions are met: [C1] the models are well calibrated, and, [C2] the probabilities of relevance are reported Figure 1: While deterministic neural rankers output with certainty. We know however that deep a point estimate probability (magenta values) of rele- neural networks (DNNs) are often not well cal- vance for a combination of query (blue bars) and doc- ibrated and have several sources of uncertainty, ument (grey bars), stochastic neural rankers output a and thus [C1] and [C2] might not be satisfied predictive distribution (orange curves). The dispersion by neural rankers. Given the success of neu- of the predictive distribution provides an estimation of ral Learning to Rank (L2R) approaches—and the model uncertainty. here, especially BERT-based approaches—we first analyze under which circumstances de- terministic neural rankers are calibrated for for the PRP to hold, ranking models must at least conversational search problems. Then, moti- meet the following conditions: [C1] assign well cal- vated by our findings we use two techniques to ibrated probabilities of relevance, i.e. if we gather model the uncertainty of neural rankers lead- all documents for which the model predicts rele- ing to the proposed stochastic rankers, which output a predictive distribution of relevance as vance with a probability of e.g. 30%, the amount opposed to point estimates. Our experimental of relevant documents should be 30%, and [C2] results on the ad-hoc retrieval task of conver- report certain predictions, i.e. only point estimates sation response ranking1 reveal that (i) BERT- such as, for example, 80% probability of relevance. based rankers are not robustly calibrated and DNNs have been shown to outperform classic that stochastic BERT-based rankers yield bet- Information Retrieval (IR) ranking models over the ter calibration; and (ii) uncertainty estimation past few years in setups where considerable train- is beneficial for both risk-aware neural rank- ing, i.e. taking into account the uncertainty ing data is available. It has been shown that DNNs when ranking documents, and for predicting are not well calibrated in the context of computer vi- unanswerable conversational contexts. sion (Guo et al., 2017). If the same is true for neural L2R models for IR, e.g. transformer-based mod- 1 Introduction els for ranking (Nogueira and Cho, 2019), [C1] According to the Probability Ranking Principle is not met. Additionally, there are a number of (PRP) (Robertson, 1977), ranking documents in sources of uncertainty in the training process of decreasing order of their probability of relevance neural networks (Gal, 2016) that make it unreason- leads to an optimal document ranking for ad-hoc able to assume that neural ranking models fulfill retrieval2. Gordon and Lenk(1991) discussed that [C2]: parameter uncertainty (different combina- 1The source code and data are available at tions of weights that explain the data equally well), https://github.com/Guzpenha/transformer_ structural uncertainty (which neural architecture to rankers/tree/uncertainty_estimation. 2Standard retrieval task where the user specifies his infor- system for documents that are likely relevant (Baeza-Yates mation need through a query which initiates a search by the et al., 1999). 160 Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 160–170 April 19 - 23, 2021. ©2021 Association for Computational Linguistics use for neural ranking), and aleatoric uncertainty ber of studies about calibration (Ovadia et al., 2019; (noisy data). Given these sources of uncertainty, Maddox et al., 2019), due to the larger decision using point estimate predictions and ranking ac- making pipelines DNNs are often part of and their cording to the PRP might not achieve the optimal importance for model interpretability (Thiagarajan ranking for retrieval. While the effectiveness bene- et al., 2020). For instance, in the automated medi- fits of risk-aware models (Wang, 2009; Wang and cal domain it is important to provide a calibrated Zhu, 2009), which take into account the risk3, i.e. confidence measure besides the prediction of a dis- the uncertainty of the document’s prediction scores, ease diagnosis to provide clinicians with sufficient have been shown for non-neural IR approaches, this information (Jiang et al., 2012). Guo et al.(2017) has not yet been explored for neural L2R models. have shown that DNNs are not well calibrated in the In this paper we first analyze the calibration of context of computer vision, motivating our study neural rankers, specifically BERT-based rankers of the calibration of neural L2R models. for IR tasks to evaluate how calibrated they are. The second condition ([C2]) for optimal retrieval Then, to model the uncertainty of BERT-based when ranking according to the PRP (Gordon and rankers, we propose stochastic neural ranking mod- Lenk, 1991) is that models report predictions with els (see Figure1), by applying different techniques certainty. While the (un)certainty has not been to model the uncertainty of DNNs, namely MC studied in neural L2R models, there are classic ap- Dropout (Gal and Ghahramani, 2016) and Deep proaches in IR that model the uncertainty. Such ap- Ensembles (Lakshminarayanan et al., 2017) which proaches have been mostly inspired by economics are agnostic to the particular DNN. theory, treating variance as a measure of uncer- In our experiments, we test models under distri- tainty (Varian, 1999). Following such ideas, non- butional shift, i.e. the test data distribution is differ- neural ranking models that take uncertainty into ent from the training data, also referred to as out-of- account (i.e. risk-aware models), and thus do not distribution (OOD) examples (Lee et al., 2018). In follow the PRP (Robertson, 1977), have been pro- real-world settings, there are often inputs that are posed (Zhu et al., 2009; Wang and Zhu, 2009), shifted due to factors such as non-stationarity and showing significant effectiveness improvements sample bias. Additionally, this experimental setup compared to the models that do not model uncer- provides a way of measuring whether the DNN tainty. Uncertainty estimation is a difficult task that ”knows what it knows” (Ovadia et al., 2019), e.g. has other applications in IR besides improving the by outputting high uncertainty for OOD examples. ranking effectiveness: it can be employed to decide We find that BERT-based rankers are not ro- between asking clarifying questions and providing bustly calibrated. Stochastic BERT-based rankers a potential answer in conversational search (Alian- have 14% less calibration error on average than nejadi et al., 2019); to perform dynamic query re- BERT-based rankers. Uncertainty estimation from formulation (Lin et al., 2020) for queries where the stochastic BERT-based rankers is advantageous for intent is uncertain; and to predict questions with no downstream applications as shown by our exper- correct answers (Feng et al., 2020). iments for risk-aware neural ranking (2% more effective on average relative to a model without Bayesian Neural Networks risk-awareness) and for predicting unanswerable Unlike standard algorithms to train neural net- conversational contexts (improves classification by works, e.g. SGD, that fit point estimate weights 33% on average of all conditions). given the observed data, Bayesian Neural Networks (BNNs) infer a distribution over the weights given 2 Related Work the observed data. Denker et al.(1987) contains one of the earliest mentions of choosing probabil- Calibration and Uncertainty in IR ity over weights of a model. An advantage of the Even though to optimally rank documents accord- Bayesian treatment of neural networks (MacKay, ing to the PRP (Robertson, 1977) requires the 1992; Neal, 2012; Blundell et al., 2015) is that they model to be calibrated (Gordon and Lenk, 1991) are better at representing existing uncertainties in ([C1]), the calibration of ranking models has re- the training procedure. One limitation of BNNs is ceived little attention in IR. In contrast, in the ma- that they are computationally expensive compared chine learning community there have been a num- to DNNs. This has lead to the development of 3In this paper we use risk and uncertainty interchangeably. techniques that scale well, and do not require mod- 161 ifications of the neural net architecture and training estimates from stochastic BERT-based rankers use- procedure. Gal and Ghahramani(2016) proposed a ful for risk-aware ranking? RQ3 Are the uncer- way to approximate Bayesian inference by relying tainty estimates obtained from stochastic BERT- on dropout (Srivastava et al., 2014). While dropout based rankers useful for identifying unanswerable is a regularization technique that ignores units with queries? We first describe how to measure the cal- probability p during every training iteration and is ibration of neural rankers ([C1]), followed by our disabled at test time, Dropout (Gal and Ghahra- approach for modeling and ranking under uncer- mani, 2016) employs dropout at both train and test tainty ([C2]), and then we describe how we evalu- time and generates a predictive distribution after ate their robustness to distributional shift. a number of forward passes. Lakshminarayanan et al.(2017) proposed an alternative: they employ 3.1 Measuring Calibration ensembles of models (Ensemble) to obtain a pre- To evaluate the calibration of neural rankers dictive distribution. Ovadia et al.(2019) showed (RQ1) we resort to the Empirical Calibration Error that Ensemble are able to produce well calibrated (ECE) (Naeini et al., 2015).
Load more

Recommended publications
  • Context-Aware Learning to Rank with Self-Attention
    Context-Aware Learning to Rank with Self-Attention Przemysław Pobrotyn Tomasz Bartczak Mikołaj Synowiec ML Research at Allegro.pl ML Research at Allegro.pl ML Research at Allegro.pl [email protected] [email protected] [email protected] Radosław Białobrzeski Jarosław Bojar ML Research at Allegro.pl ML Research at Allegro.pl [email protected] [email protected] ABSTRACT engines, recommender systems, question-answering systems, and Learning to rank is a key component of many e-commerce search others. engines. In learning to rank, one is interested in optimising the A typical machine learning solution to the LTR problem involves global ordering of a list of items according to their utility for users. learning a scoring function, which assigns real-valued scores to Popular approaches learn a scoring function that scores items in- each item of a given list, based on a dataset of item features and dividually (i.e. without the context of other items in the list) by human-curated or implicit (e.g. clickthrough logs) relevance labels. optimising a pointwise, pairwise or listwise loss. The list is then Items are then sorted in the descending order of scores [23]. Perfor- sorted in the descending order of the scores. Possible interactions mance of the trained scoring function is usually evaluated using an between items present in the same list are taken into account in the IR metric like Mean Reciprocal Rank (MRR) [33], Normalised Dis- training phase at the loss level. However, during inference, items counted Cumulative Gain (NDCG) [19] or Mean Average Precision are scored individually, and possible interactions between them are (MAP) [4].
    [Show full text]
  • An Introduction to Neural Information Retrieval
    The version of record is available at: http://dx.doi.org/10.1561/1500000061 An Introduction to Neural Information Retrieval Bhaskar Mitra1 and Nick Craswell2 1Microsoft, University College London; [email protected] 2Microsoft; [email protected] ABSTRACT Neural ranking models for information retrieval (IR) use shallow or deep neural networks to rank search results in response to a query. Tradi- tional learning to rank models employ supervised machine learning (ML) techniques—including neural networks—over hand-crafted IR features. By contrast, more recently proposed neural models learn representations of language from raw text that can bridge the gap between query and docu- ment vocabulary. Unlike classical learning to rank models and non-neural approaches to IR, these new ML techniques are data-hungry, requiring large scale training data before they can be deployed. This tutorial intro- duces basic concepts and intuitions behind neural IR models, and places them in the context of classical non-neural approaches to IR. We begin by introducing fundamental concepts of retrieval and different neural and non-neural approaches to unsupervised learning of vector representations of text. We then review IR methods that employ these pre-trained neural vector representations without learning the IR task end-to-end. We intro- duce the Learning to Rank (LTR) framework next, discussing standard loss functions for ranking. We follow that with an overview of deep neural networks (DNNs), including standard architectures and implementations. Finally, we review supervised neural learning to rank models, including re- cent DNN architectures trained end-to-end for ranking tasks. We conclude with a discussion on potential future directions for neural IR.
    [Show full text]
  • Metric Learning for Session-Based Recommendations - Preprint
    METRIC LEARNING FOR SESSION-BASED RECOMMENDATIONS A PREPRINT Bartłomiej Twardowski1,2 Paweł Zawistowski Szymon Zaborowski 1Computer Vision Center, Warsaw University of Technology, Sales Intelligence Universitat Autónoma de Barcelona Institute of Computer Science 2Warsaw University of Technology, [email protected] Institute of Computer Science [email protected] January 8, 2021 ABSTRACT Session-based recommenders, used for making predictions out of users’ uninterrupted sequences of actions, are attractive for many applications. Here, for this task we propose using metric learning, where a common embedding space for sessions and items is created, and distance measures dissimi- larity between the provided sequence of users’ events and the next action. We discuss and compare metric learning approaches to commonly used learning-to-rank methods, where some synergies ex- ist. We propose a simple architecture for problem analysis and demonstrate that neither extensively big nor deep architectures are necessary in order to outperform existing methods. The experimental results against strong baselines on four datasets are provided with an ablation study. Keywords session-based recommendations · deep metric learning · learning to rank 1 Introduction We consider the session-based recommendation problem, which is set up as follows: a user interacts with a given system (e.g., an e-commerce website) and produces a sequence of events (each described by a set of attributes). Such a continuous sequence is called a session, thus we denote sk = ek,1,ek,2,...,ek,t as the k-th session in our dataset, where ek,j is the j-th event in that session. The events are usually interactions with items (e.g., products) within the system’s domain.
    [Show full text]
  • Learning Embeddings: Efficient Algorithms and Applications
    Learning embeddings: HIÀFLHQWDOJRULWKPVDQGDSSOLFDWLRQV THÈSE NO 8348 (2018) PRÉSENTÉE LE 15 FÉVRIER 2018 À LA FACULTÉ DES SCIENCES ET TECHNIQUES DE L'INGÉNIEUR LABORATOIRE DE L'IDIAP PROGRAMME DOCTORAL EN GÉNIE ÉLECTRIQUE ÉCOLE POLYTECHNIQUE FÉDÉRALE DE LAUSANNE POUR L'OBTENTION DU GRADE DE DOCTEUR ÈS SCIENCES PAR Cijo JOSE acceptée sur proposition du jury: Prof. P. Frossard, président du jury Dr F. Fleuret, directeur de thèse Dr O. Bousquet, rapporteur Dr M. Cisse, rapporteur Dr M. Salzmann, rapporteur Suisse 2018 Abstract Learning to embed data into a space where similar points are together and dissimilar points are far apart is a challenging machine learning problem. In this dissertation we study two learning scenarios that arise in the context of learning embeddings and one scenario in efficiently estimating an empirical expectation. We present novel algorithmic solutions and demonstrate their applications on a wide range of data-sets. The first scenario deals with learning from small data with large number of classes. This setting is common in computer vision problems such as person re-identification and face verification. To address this problem we present a new algorithm called Weighted Approximate Rank Component Analysis (WARCA), which is scalable, robust, non-linear and is independent of the number of classes. We empirically demonstrate the performance of our algorithm on 9 standard person re-identification data-sets where we obtain state of the art performance in terms of accuracy as well as computational speed. The second scenario we consider is learning embeddings from sequences. When it comes to learning from sequences, recurrent neural networks have proved to be an effective algorithm.
    [Show full text]
  • An Empirical Study of Embedding Features in Learning to Rank
    An Empirical Study of Embedding Features in Learning to Rank ABSTRACT use of entities can recognize which part of the query is the most is paper explores the possibility of using neural embedding fea- informative; (2) the second aspect that we consider is the ‘type tures for enhancing the eectiveness of ad hoc document ranking of embeddings’ that can be used. We have included both word based on learning to rank models. We have extensively introduced embedding [8] and document embedding [4] models for building and investigated the eectiveness of features learnt based on word dierent features. and document embeddings to represent both queries and docu- Based on these two cross-cuing aspects, namely type of content ments. We employ several learning to rank methods for document and type of embedding, we have dened ve classes of features as ranking using embedding-based features, keyword-based features shown in Table 1. Briey stated, our features are query-dependent as well as the interpolation of the embedding-based features with features and hence, we employ the embedding models to form vec- keyword-based features. e results show that embedding features tor representations of words, chunks and entities in both documents have a synergistic impact on keyword based features and are able and queries that can then be employed for measuring the distance to provide statistically signicant improvement on harder queries. between queries and documents. It should be noted that we do not use entity embeddings [3] to represent entities, but rather use the ACM Reference format: same word embedding and document embedding models on the .
    [Show full text]
  • Learning to Rank Using Gradient Descent
    Learning to Rank using Gradient Descent Keywords: ranking, gradient descent, neural networks, probabilistic cost functions, internet search Chris Burges [email protected] Tal Shaked∗ [email protected] Erin Renshaw [email protected] Microsoft Research, One Microsoft Way, Redmond, WA 98052-6399 Ari Lazier [email protected] Matt Deeds [email protected] Nicole Hamilton [email protected] Greg Hullender [email protected] Microsoft, One Microsoft Way, Redmond, WA 98052-6399 Abstract that maps to the reals (having the model evaluate on We investigate using gradient descent meth- pairs would be prohibitively slow for many applica- ods for learning ranking functions; we pro- tions). However (Herbrich et al., 2000) cast the rank- pose a simple probabilistic cost function, and ing problem as an ordinal regression problem; rank we introduce RankNet, an implementation of boundaries play a critical role during training, as they these ideas using a neural network to model do for several other algorithms (Crammer & Singer, the underlying ranking function. We present 2002; Harrington, 2003). For our application, given test results on toy data and on data from a that item A appears higher than item B in the out- commercial internet search engine. put list, the user concludes that the system ranks A higher than, or equal to, B; no mapping to particular rank values, and no rank boundaries, are needed; to 1. Introduction cast this as an ordinal regression problem is to solve an unnecessarily hard problem, and our approach avoids Any system that presents results to a user, ordered this extra step. We also propose a natural probabilis- by a utility function that the user cares about, is per- tic cost function on pairs of examples.
    [Show full text]
  • Multi-Agent Reinforced Learning to Rank
    MarlRank: Multi-agent Reinforced Learning to Rank Shihao Zou∗ Zhonghua Li Mohammad Akbari [email protected] [email protected] [email protected] University of Alberta Huawei University College London Jun Wang Peng Zhang [email protected] [email protected] University College London Tianjin University ABSTRACT 1 INTRODUCTION When estimating the relevancy between a query and a document, Learning to rank, which learns a model for ranking documents ranking models largely neglect the mutual information among doc- (items) based on their relevance scores to a given query, is a key uments. A common wisdom is that if two documents are similar task in information retrieval (IR) [18]. Typically, a query-document in terms of the same query, they are more likely to have similar pair is represented as a feature vector and many classical models, in- relevance score. To mitigate this problem, in this paper, we propose cluding SVM [8], AdaBoost [18] and regression tree [2] are applied a multi-agent reinforced ranking model, named MarlRank. In partic- to learn to predict the relevancy. Recently, deep learning models has ular, by considering each document as an agent, we formulate the also been applied in learning to rank task. For example, CDSSM [13], ranking process as a multi-agent Markov Decision Process (MDP), DeepRank [12] and their variants aim at learning better semantic where the mutual interactions among documents are incorporated representation of the query-document pair for more accurate pre- in the ranking process. To compute the ranking list, each docu- diction of the relevancy.
    [Show full text]
  • Yahoo! Learning to Rank Challenge Overview
    JMLR: Workshop and Conference Proceedings 14 (2011) 1–24 Yahoo! Learning to Rank Challenge Yahoo! Learning to Rank Challenge Overview Olivier Chapelle⇤ [email protected] Yi Chang [email protected] Yahoo! Labs Sunnyvale, CA Abstract Learning to rank for information retrieval has gained a lot of interest in the recent years but there is a lack for large real-world datasets to benchmark algorithms. That led us to publicly release two datasets used internally at Yahoo! for learning the web search ranking function. To promote these datasets and foster the development of state-of-the-art learning to rank algorithms, we organized the Yahoo! Learning to Rank Challenge in spring 2010. This paper provides an overview and an analysis of this challenge, along with a detailed description of the released datasets. 1. Introduction Ranking is at the core of information retrieval: given a query, candidates documents have to be ranked according to their relevance to the query. Learning to rank is a relatively new field in which machine learning algorithms are used to learn this ranking function. It is of particular importance for web search engines to accurately tune their ranking functions as it directly a↵ects the search experience of millions of users. A typical setting in learning to rank is that feature vectors describing a query-document pair are constructed and relevance judgments of the documents to the query are available. A ranking function is learned based on this training data, and then applied to the test data. Several benchmark datasets, such as letor, have been released to evaluate the newly proposed learning to rank algorithms.
    [Show full text]
  • Sodeep: a Sorting Deep Net to Learn Ranking Loss Surrogates
    SoDeep: a Sorting Deep net to learn ranking loss surrogates Martin Engilberge1,2, Louis Chevallier2, Patrick Perez´ 3, Matthieu Cord1,3 1Sorbonne Universite,´ Paris, France, 2Technicolor, Cesson Sevign´ e,´ France, 3Valeo.ai, Paris, France {martin.engilberge, matthieu.cord}@lip6.fr [email protected] [email protected] Abstract Several tasks in machine learning are evaluated us- ing non-differentiable metrics such as mean average pre- cision or Spearman correlation. However, their non- differentiability prevents from using them as objective func- tions in a learning framework. Surrogate and relaxation methods exist but tend to be specific to a given metric. In the present work, we introduce a new method to learn approximations of such non-differentiable objective func- tions. Our approach is based on a deep architecture that approximates the sorting of arbitrary sets of scores. It is Figure 1: Overview of SoDeep, the proposed end-to- trained virtually for free using synthetic data. This sorting end trainable deep architecture to approximate non- deep (SoDeep) net can then be combined in a plug-and-play differentiable ranking metrics. A pre-trained differen- manner with existing deep architectures. We demonstrate tiable sorter (deep neural net [DNN] ΘB) is used to con- the interest of our approach in three different tasks that vert into ranks the raw scores given by the model (DNN require ranking: Cross-modal text-image retrieval, multi- ΘA) being trained to a collection of inputs. A loss is then label image classification and visual memorability ranking. applied to the predicted rank and the error can be back- Our approach yields very competitive results on these three propagated through the differentiable sorter and used to up- tasks, which validates the merit and the flexibility of SoDeep date the weights ΘA.
    [Show full text]
  • Spectrum-Enhanced Pairwise Learning to Rank
    Spectrum-enhanced Pairwise Learning to Rank Wenhui Yu Zheng Qin∗ Tsinghua University Tsinghua University Beijing, China Beijing, China [email protected] [email protected] ABSTRACT 1 INTRODUCTION To enhance the performance of the recommender system, side in- Recommender systems have been widely used in online services formation is extensively explored with various features (e.g., visual such as E-commerce and social media sites to predict users’ prefer- features and textual features). However, there are some demerits ence based on their interaction histories. Modern recommender sys- of side information: (1) the extra data is not always available in all tems uncover the underlying latent factors that encode the prefer- recommendation tasks; (2) it is only for items, there is seldom high- ence of users and properties of items. In recent years, to strengthen level feature describing users. To address these gaps, we introduce the presentation ability of the models, various features are incor- the spectral features extracted from two hypergraph structures of porated for additional information, such as visual features from the purchase records. Spectral features describe the similarity of product images [14, 44, 48], textual features from review data [8, 9], users/items in the graph space, which is critical for recommendation. and auditory features from music [5]. However, these features are We leverage spectral features to model the users’ preference and not generally applicable, for example, visual features can only be items’ properties by incorporating them into a Matrix Factorization used in product recommendation, while not in music recommenda- (MF) model. tion. Also, these features are unavailable in some recommendation In addition to modeling, we also use spectral features to optimize.
    [Show full text]
  • Embedding Meta-Textual Information for Improved Learning to Rank
    Embedding Meta-Textual Information for Improved Learning to Rank Toshitaka Kuwa∗, Shigehiko Schamoniy;z, Stefan Riezlery;z yDepartment of Computational Linguistics, Heidelberg, Germany. zInterdisciplinary Center for Scientific Computing (IWR), Heidelberg, Germany fkuwa, schamoni, [email protected] Abstract Neural approaches to learning term embeddings have led to improved computation of similarity and ranking in information retrieval (IR). So far neural representation learning has not been ex- tended to meta-textual information that is readily available for many IR tasks, for example, patent classes in prior-art retrieval, topical information in Wikipedia articles, or product categories in e-commerce data. We present a framework that learns embeddings for meta-textual categories, and optimizes a pairwise ranking objective for improved matching based on combined embed- dings of textual and meta-textual information. We show considerable gains in an experimental evaluation on cross-lingual retrieval in the Wikipedia domain for three language pairs, and in the Patent domain for one language pair. Our results emphasize that the mode of combining different types of information is crucial for model improvement. 1 Introduction The recent success of neural methods in IR rests on bridging the gap between query and document vocabulary by learning term embeddings that allow for improved computation of similarity and ranking (Mitra and Craswell, 2018). So far vector representations in neural IR have been confined to textual information, neglecting information beyond the raw text that is potentially useful for retrieval and is readily available in many data situations. For example, meta-textual information is available in the form of patent classes in prior-art search, topic classes in search over Wikipedia articles, or product classes in retrieval in the e-commerce domain.
    [Show full text]
  • Self-Supervised Learning from Permutations Via Differentiable Ranking
    Under review as a conference paper at ICLR 2021 SHUFFLE TO LEARN:SELF-SUPERVISED LEARNING FROM PERMUTATIONS VIA DIFFERENTIABLE RANKING Anonymous authors Paper under double-blind review ABSTRACT Self-supervised pre-training using so-called “pretext” tasks has recently shown impressive performance across a wide range of tasks. In this work, we advance self-supervised learning from permutations, that consists in shuffling parts of in- put and training a model to reorder them, improving downstream performance in classification. To do so, we overcome the main challenges of integrating permu- tation inversions (a discontinuous operation) into an end-to-end training scheme, heretofore sidestepped by casting the reordering task as classification, fundamen- tally reducing the space of permutations that can be exploited. We make two main contributions. First, we use recent advances in differentiable ranking to integrate the permutation inversion flawlessly into a neural network, enabling us to use the full set of permutations, at no additional computing cost. Our experiments validate that learning from all possible permutations improves the quality of the pre-trained representations over using a limited, fixed set. Second, we successfully demon- strate that inverting permutations is a meaningful pretext task in a diverse range of modalities, beyond images, which does not require modality-specific design. In particular, we improve music understanding by reordering spectrogram patches in the time-frequency space, as well as video classification by reordering frames along the time axis. We furthermore analyze the influence of the patches that we use (vertical, horizontal, 2-dimensional), as well as the benefit of our approach in different data regimes.
    [Show full text]