DOCSLIB.ORG

Fast and Scalable Polynomial Kernels Via Explicit Feature Maps *

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Fast and Scalable Polynomial Kernels Via Explicit Feature Maps * Fast and Scalable Polynomial Kernels via Explicit Feature Maps * Ninh Pham Rasmus Pagh IT University of Copenhagen IT University of Copenhagen Copenhagen, Denmark Copenhagen, Denmark [email protected] [email protected] ABSTRACT data space into high-dimensional feature space, where each Approximation of non-linear kernels using random feature coordinate corresponds to one feature of the data points. In mapping has been successfully employed in large-scale data that space, one can perform well-known data analysis al- analysis applications, accelerating the training of kernel ma- gorithms without ever interacting with the coordinates of chines. While previous random feature mappings run in the data, but rather by simply computing their pairwise in- O(ndD) time for n training samples in d-dimensional space ner products. This operation can not only avoid the cost and D random feature maps, we propose a novel random- of explicit computation of the coordinates in feature space ized tensor product technique, called Tensor Sketching, for but also handle general types of data (such as numeric data, approximating any polynomial kernel in O(n(d + D log D)) symbolic data). time. Also, we introduce both absolute and relative error While kernel methods have been used successfully in a va- bounds for our approximation to guarantee the reliability riety of data analysis tasks, their scalability is a bottleneck. of our estimation algorithm. Empirically, Tensor Sketching Kernel-based learning algorithms usually scale poorly with achieves higher accuracy and often runs orders of magni- the number of the training samples (a cubic running time tude faster than the state-of-the-art approach for large-scale and quadratic storage for direct methods). This drawback real-world datasets. is becoming more crucial with the rise of big data applica- tions [12, 3]. Recently, Joachims [9] proposed an efficient training algorithm for linear SVMs that runs in time lin- Categories and Subject Descriptors ear in the number of training examples. Since one can view I.5.2 [Pattern Recognition]: Design Methodology|Clas- non-linear SVMs as linear SVMs operating in an appro- sifier design and evaluation priate feature space, Rahimi and Recht [16] first proposed a random feature mapping to approximate shift-invariant kernels in order to combine the advantages of both linear General Terms and non-linear SVM approaches. This approach approxi- Algorithms, Performance, Experimentation mates kernels by an explicit data mapping into relatively low-dimensional random feature space. In this random fea- Keywords ture space, the kernel of any two points is well approximated by their inner product. Therefore, one can apply existing polynomial kernel; SVM; tensor product; Count Sketch; FFT fast linear learning algorithms to find data relations corre- sponding to non-linear kernel methods in the random feature 1. INTRODUCTION space. That leads to a substantial reduction in training time Kernel machines such as Support Vector Machines (SVMs) while obtaining similar testing error. have recently emerged as powerful approaches for many ma- Following up this line of work, many randomized approaches chine learning and data mining tasks. One of the key proper- to approximate kernels are proposed for accelerating the ties of kernel methods is the capability to efficiently find non- training of kernel machines [10, 12, 20, 21]. While the train- linear structure of data by the use of kernels. A kernel can be ing algorithm is linear, existing kernel approximation map- viewed as an implicit non-linear data mapping from original pings require time proportional to the product of the number of dimensions d and the number of random features D. This *This work is supported by the Danish National Research means that the mapping itself is a bottleneck whenever dD Foundation under the Sapere Aude program. is not small. In this paper we address this bottleneck, and present a near-linear time mapping for approximating any polynomial kernel. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not Particularly, given any two points of a dataset S of n d made or distributed for profit or commercial advantage and that copies bear points, x = fx1; ··· ; xdg ; y = fy1; ··· ; ydg 2 S ⊂ R and d this notice and the full citation on the first page. Copyrights for components an implicit feature space mapping φ : R 7! F, the inner of this work owned by others than ACM must be honored. Abstracting with product between these points in the feature space F can credit is permitted. To copy otherwise, or republish, to post on servers or to be quickly computed as hφ(x); φ(y)i = κ(x; y) where κ() is redistribute to lists, requires prior specific permission and/or a fee. Request an easily computable kernel. An explicit random feature permissions from [email protected]. d D KDD’13, August 11–14, 2013, Chicago, Illinois, USA. mapping f : R 7! R can efficiently approximate a kernel Copyright 2013 ACM 978-1-4503-2174-7/13/08 ...$15.00. κ() if it satisfies: of kernel methods' coefficients by performing coordinate as- cent on subsets of the training set until KKT conditions have E [hf(x); f(y)i] = hφ(x); φ(y)i = κ(x; y) : been satisfied to within a certain tolerance. Although such So we can transform data from the original data space into a approaches can handle the memory restrictions involving the low-dimensional explicit random feature space and use any dense kernel matrix, they still involve numerical solutions of linear learning algorithm to find non-linear data relations. optimization subproblems and therefore can be problematic Rahimi and Recht [16] introduced a random projection- and expensive to large-scale datasets. based algorithm to approximate shift-invariant kernels (e.g. In order to apply kernel methods to large-scale datasets, the Gaussian kernel κ(x; y) = exp(− kx − yk2 =2σ2), for σ > many approaches have been proposed for quickly approxi- 0). Vempati et al. [21] extended this work to approximate mating the kernel matrix, including the Nystr¨om methods generalized radial-basic function (RBF) kernels (e.g. the [5, 23], sparse greedy approximation [19] and low-rank kernel exponential-χ2 kernel κ(x; y) = exp(−χ2(x; y)=2σ2), where approximation [7]. These approximation schemes can reduce σ > 0 and χ2 is the Chi-squared distance measure). Re- the computational and storage costs of operating on a kernel cently, Kar and Karnick [10] made use of the Maclaurin se- matrix while preserving the quality of results. An assump- ries expansion to approximate inner product kernels (e.g. tion of these approaches is that the kernel matrix has many the polynomial kernel κ(x; y) = (hx; yi + c)p, for c ≥ 0 and zero eigenvalues. This might not be true in many datasets. an integer p). Furthermore, there is a lack of experiments to illustrate the These approaches have to maintain D random vectors efficiency of these approaches on large-scale datasets [16]. d !1; ··· !D 2 R in O(dD) space and need O(ndD) operations Instead of approximating the kernel matrix, recent ap- for computing D random feature maps. That incurs sig- proaches [10, 12, 16, 20, 21, 24] approximate the kernels nificant (quadratic) computational and storage costs when by explicitly mapping data into a relatively low-dimensional D = O(d) and d is rather large. When the decision bound- random feature space. The explicit mapping transforms ary of the problem is rather smooth, the computational cost data into a random feature space where the pairwise in- of random mapping might dominate the training cost. In ner products of transformed data points are approximately addition, the absolute error bounds of previous approaches equal to kernels in feature space. Therefore, we can apply are not tight. Particularly, the Maclaurin expansion based existing fast linear learning algorithms [6, 9, 18] to find non- approach [10] suffers from large error because it approxi- linear data relations in that random feature space. While mates the homogeneous polynomial kernel κ(x; y) = hx; yip previous such approaches can efficiently accelerate the train- Qp Qp d ing of kernel machines, they incur significant computational by i=1 h!i; xi i=1 h!i; yi where !i 2 f+1; −1g . Our experiments show that large estimation error results in ei- cost (quadratic in the dimensionality of data). That results ther accuracy degradation or negligible reduction in training in performance degradation on large-scale high-dimensional time. datasets. In this work, we consider the problem of approximating the commonly used polynomial kernel κ(x; y) = (hx; yi + c)p to accelerate the training of kernel machines. We develop 3. BACKGROUND AND PRELIMINARIES a fast and scalable randomized tensor product technique, named Tensor Sketching, to estimate the polynomial kernel 3.1 Count Sketch of any pair of points of the dataset. Our proposed approach works in O(np(d + D log D)) time and requires O(pd log D) Charikar et al. [2] described and analyzed a sketching ap- space for random vectors. The main technical insight is the proach, called Count Sketch, to estimate the frequency of connection between tensor product and fast convolution of all items in a stream. Recently, the machine learning com- Count Sketches [2, 14], which enables us to reduce the com- munity has used Count Sketch as a feature hashing tech- putational complexity and space usage. We introduce both nique for large-scale multitask learning [22] because Count absolute and relative error bounds for our approximation to Sketches can preserve the pairwise inner products within guarantee the reliability of our estimation algorithm. The an arbitrarily small factor.
Load more

Recommended publications
  • Short and Deep: Sketching and Neural Networks
    SHORT AND DEEP: SKETCHING AND NEURAL NETWORKS Amit Daniely, Nevena Lazic, Yoram Singer, Kunal Talwar∗ Google Brain ABSTRACT Data-independent methods for dimensionality reduction such as random projec- tions, sketches, and feature hashing have become increasingly popular in recent years. These methods often seek to reduce dimensionality while preserving the hypothesis class, resulting in inherent lower bounds on the size of projected data. For example, preserving linear separability requires Ω(1/γ2) dimensions, where γ is the margin, and in the case of polynomial functions, the number of required di- mensions has an exponential dependence on the polynomial degree. Despite these limitations, we show that the dimensionality can be reduced further while main- taining performance guarantees, using improper learning with a slightly larger hypothesis class. In particular, we show that any sparse polynomial function of a sparse binary vector can be computed from a compact sketch by a single-layer neu- ral network, where the sketch size has a logarithmic dependence on the polyno- mial degree. A practical consequence is that networks trained on sketched data are compact, and therefore suitable for settings with memory and power constraints. We empirically show that our approach leads to networks with fewer parameters than related methods such as feature hashing, at equal or better performance. 1 INTRODUCTION In many supervised learning problems, input data are high-dimensional and sparse. The high di- mensionality may be inherent in the domain, such as a large vocabulary in a language model, or the result of creating hybrid conjunction features. This setting poses known statistical and compu- tational challenges for standard supervised learning techniques, as high-dimensional inputs lead to models with a very large number of parameters.
    [Show full text]
  • Learning-Based Frequency Estimation Algorithms
    Published as a conference paper at ICLR 2019 LEARNING-BASED FREQUENCY ESTIMATION ALGORITHMS Chen-Yu Hsu, Piotr Indyk, Dina Katabi & Ali Vakilian Computer Science and Artificial Intelligence Lab Massachusetts Institute of Technology Cambridge, MA 02139, USA fcyhsu,indyk,dk,[email protected] ABSTRACT Estimating the frequencies of elements in a data stream is a fundamental task in data analysis and machine learning. The problem is typically addressed using streaming algorithms which can process very large data using limited storage. Today’s streaming algorithms, however, cannot exploit patterns in their input to improve performance. We propose a new class of algorithms that automatically learn relevant patterns in the input data and use them to improve its frequency estimates. The proposed algorithms combine the benefits of machine learning with the formal guarantees available through algorithm theory. We prove that our learning-based algorithms have lower estimation errors than their non-learning counterparts. We also evaluate our algorithms on two real-world datasets and demonstrate empirically their performance gains. 1.I NTRODUCTION Classical algorithms provide formal guarantees over their performance, but often fail to leverage useful patterns in their input data to improve their output. On the other hand, deep learning models are highly successful at capturing and utilizing complex data patterns, but often lack formal error bounds. The last few years have witnessed a growing effort to bridge this gap and introduce al- gorithms that can adapt to data properties while delivering worst case guarantees. Deep learning modules have been integrated into the design of Bloom filters (Kraska et al., 2018; Mitzenmacher, 2018), caching algorithms (Lykouris & Vassilvitskii, 2018), graph optimization (Dai et al., 2017), similarity search (Salakhutdinov & Hinton, 2009; Weiss et al., 2009) and compressive sensing (Bora et al., 2017).
    [Show full text]
  • Incremental Randomized Sketching for Online Kernel Learning
    Incremental Randomized Sketching for Online Kernel Learning Xiao Zhang 1 Shizhong Liao 1 Abstract updating time significantly (Wang et al., 2012). To solve the Randomized sketching has been used in offline problem, budgeted online kernel learning has been proposed kernel learning, but it cannot be applied directly to to maintain a buffer of support vectors (SVs) with a fixed online kernel learning due to the lack of incremen- budget, which limits the model size to reduce the computa- tal maintenances for randomized sketches with tional complexities (Crammer et al., 2003). Several budget regret guarantees. To address these issues, we maintenance strategies were introduced to bound the buffer propose a novel incremental randomized sketch- size of SVs. Dekel et al.(2005) maintained the budget by ing approach for online kernel learning, which discarding the oldest support vector. Orabona et al.(2008) has efficient incremental maintenances with theo- projected the new example onto the linear span of SVs in retical guarantees. We construct two incremental the feature space to reduce the buffer size. But the buffer of randomized sketches using the sparse transform SVs cannot be used for kernel matrix approximation directly. matrix and the sampling matrix for kernel ma- Besides, it is difficult to provide a lower bound of the buffer trix approximation, update the incremental ran- size that is key to the theoretical guarantees for budgeted domized sketches using rank-1 modifications, and online kernel learning. construct a time-varying explicit feature mapping Unlike the budget maintenance strategies of SVs, sketches for online kernel learning. We prove that the pro- of the kernel matrix have been introduced into online kernel posed incremental randomized sketching is statis- learning to approximate the kernel incrementally.
    [Show full text]
  • Compressing Gradient Optimizers Via Count-Sketches
    Compressing Gradient Optimizers via Count-Sketches Ryan Spring 1 * Anastasios Kyrillidis 1 Vijai Mohan 2 Anshumali Shrivastava 1 2 Abstract Training large-scale models efficiently is a challenging task. Many popular first-order optimization methods There are numerous publications that describe how to lever- accelerate the convergence rate of deep learning age multi-GPU data parallelism and mixed precision train- models. However, these algorithms require aux- ing effectively (Hoffer et al., 2017; Ott et al., 2018; Micike- iliary variables, which cost additional memory vicius et al., 2018). A key tool for improving training time proportional to the number of parameters in the is to increase the batch size, taking advantage of the massive model. The problem is becoming more severe as parallelism provided by GPUs. However, increasing the models grow larger to learn from complex, large- batch size also requires significant amounts of memory. A scale datasets. Our proposed solution is to main- practitioner will sometimes sacrifice their batch size for a tain a linear sketch to compress the auxiliary vari- larger, more expressive model. For example, (Puri et al., ables. Our approach has the same performance as 2018) showed that doubling the dimensionality of a mul- the full-sized baseline, while using less space for tiplicative LSTM (Krause et al., 2016) from 4096 to 8192 the auxiliary variables. Theoretically, we prove forced them to reduce the batch size per GPU by 4×. that count-sketch optimization maintains the SGD One culprit that aggravates the memory capacity issue is convergence rate, while gracefully reducing mem- the auxiliary parameters used by first-order optimization ory usage for large-models.
    [Show full text]
  • Communication-Efficient Federated Learning with Sketching
    Communication-Efficient Federated Learning with Sketching Ashwinee Panda Daniel Rothchild Enayat Ullah Nikita Ivkin Ion Stoica Joseph Gonzalez, Ed. Raman Arora, Ed. Vladimir Braverman, Ed. Electrical Engineering and Computer Sciences University of California at Berkeley Technical Report No. UCB/EECS-2020-57 http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-57.html May 23, 2020 Copyright © 2020, by the author(s). All rights reserved. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission. COMMUNICATION-EFFICIENT FEDERATED LEARNING WITH SKETCHING APREPRINT 1 1 2 3† Daniel Rothchild ∗ Ashwinee Panda ∗ Enayat Ullah Nikita Ivkin Vladimir Braverman2 Joseph Gonzalez1 Ion Stoica1 Raman Arora2 1 UC Berkeley, drothchild, ashwineepanda, jegonzal, istoica @berkeley.edu 2Johns Hopkinsf University, [email protected], vova, arora @cs.jhu.edug 3 Amazon, [email protected] g ABSTRACT Existing approaches to federated learning suffer from a communication bottleneck as well as conver- gence issues due to sparse client participation. In this paper we introduce a novel algorithm, called FedSketchedSGD, to overcome these challenges. FedSketchedSGD compresses model updates using a Count Sketch, and then takes advantage of the mergeability of sketches to combine model updates from many workers. A key insight in the design of FedSketchedSGD is that, because the Count Sketch is linear, momentum and error accumulation can both be carried out within the sketch.
    [Show full text]
  • Tensors in Modern Statistical Learning 1 Introduction
    Tensors in Modern Statistical Learning Will Wei Suny, Botao Haoz, and Lexin Li∗ yPurdue University, zDeepMind, and ∗University of California at Berkeley Abstract Tensor learning is gaining increasing attention in recent years. This survey provides an overview of tensor analysis in modern statistical learning. It consists of four main topics, including tensor supervised learning, tensor unsupervised learning, tensor reinforcement learning, and tensor deep learning. This review emphasizes statistical models and properties, as well as connections between tensors and other learning topics such as reinforcement learning and deep learning. Keywords: Tensor decomposition; Tensor regression; Tensor clustering; Tensor graphical model; Tensor reinforcement learning; Tensor deep learning. 1 Introduction Tensors, also known as multidimensional arrays, are generalizations of vectors and matrices to higher dimensions. In recent years, tensor data are fast emerging in a wide variety of scientific and business applications, including but not limited to recommendation systems (Rendle and Schmidt-Thieme, 2010; Bi et al., 2018), speech or facial recognitions (Vasilescu and Terzopoulos, 2002; Ma et al., 2019), networks analysis (Li et al., 2011; Ermis et al., 2015), knowledge graphs and relational learning (Trouillon et al., 2017; Liu et al., 2020), among many others. Tensor data analysis is thus gaining increasing attention in statistics and machine learning communities. In this survey, we provide an overview of tensor analysis in modern statistical learning.
    [Show full text]
  • (Learned) Frequency Estimation Algorithms Under Zipfian Distribution
    (Learned) Frequency Estimation Algorithms under Zipfian Distribution Anders Aamand∗ Piotr Indyky Ali Vakilianz Abstract The frequencies of the elements in a data stream are an important statistical measure and the task of estimating them arises in many applications within data analysis and machine learning. Two of the most popular algorithms for this problem, Count-Min and Count-Sketch, are widely used in practice. In a recent work [Hsu et al., ICLR'19], it was shown empirically that augmenting Count- Min and Count-Sketch with a machine learning algorithm leads to a significant reduction of the estimation error. The experiments were complemented with an analysis of the expected error incurred by Count-Min (both the standard and the augmented version) when the input frequencies follow a Zipfian distribution. Although the authors established that the learned version of Count-Min has lower estimation error than its standard counterpart, their analysis of the standard Count-Min algorithm was not tight. Moreover, they provided no similar analysis for Count-Sketch. In this paper we resolve these problems. First, we provide a simple tight analysis of the expected error incurred by Count-Min. Second, we provide the first error bounds for both the standard and the augmented version of Count-Sketch. These bounds are nearly tight and again demonstrate an improved performance of the learned version of Count-Sketch. In addition to demonstrating tight gaps between the aforementioned algorithms, we believe that our bounds for the standard versions of Count-Min and Count-Sketch are of independent interest. In particular, it is a typical practice to set the number of hash functions in those algorithms to Θ(log n).
    [Show full text]
  • Communication-Efficient Distributed SGD with Sketching
    Communication-efficient Distributed SGD with Sketching Nikita Ivkin ∗y Daniel Rothchild ∗ Enayat Ullah ∗ Amazon UC Berkeley Johns Hopkins University [email protected] [email protected] [email protected] Vladimir Braverman z Ion Stoica Raman Arora Johns Hopkins University UC Berkeley Johns Hopkins University [email protected] [email protected] [email protected] Abstract Large-scale distributed training of neural networks is often limited by network band- width, wherein the communication time overwhelms the local computation time. Motivated by the success of sketching methods in sub-linear/streaming algorithms, we introduce SKETCHED-SGD4, an algorithm for carrying out distributed SGD by communicating sketches instead of full gradients. We show that SKETCHED-SGD has favorable convergence rates on several classes of functions. When considering all communication – both of gradients and of updated model weights – SKETCHED- SGD reduces the amount of communication required compared to other gradient compression methods from O(d) or O(W ) to O(log d), where d is the number of model parameters and W is the number of workers participating in training. We run experiments on a transformer model, an LSTM, and a residual network, demonstrating up to a 40x reduction in total communication cost with no loss in final model performance. We also show experimentally that SKETCHED-SGD scales to at least 256 workers without increasing communication cost or degrading model performance. 1 Introduction Modern machine learning training workloads are commonly distributed across many machines using data-parallel synchronous stochastic gradient descent. At each iteration, W worker nodes split a mini-batch of size B; each worker computes the gradient of the loss on its portion of the data, and then a parameter server sums each worker’s gradient to yield the full mini-batch gradient.
    [Show full text]
  • Privacy for Free: Communication-Efficient Learning
    Privacy for Free: Communication-Efficient Learning with Differential Privacy Using Sketches Tian Li Zaoxing Liu Vyas Sekar Virginia Smith CMU CMU CMU CMU [email protected] [email protected] [email protected] [email protected] Abstract Communication and privacy are two critical concerns in distributed learning. Many existing works treat these concerns separately. In this work, we argue that a natural connection exists between methods for communication reduction and privacy preservation in the context of distributed machine learning. In particular, we prove that Count Sketch, a simple method for data stream summarization, has inherent differential privacy properties.1 Using these derived privacy guarantees, we propose a novel sketch-based framework (DiffSketch) for distributed learning, where we compress the transmitted messages via sketches to simultaneously achieve communication efficiency and provable privacy benefits. Our evaluation demonstrates that DiffSketch can provide strong differential privacy guarantees (e.g., ε = 1) and reduce communication by 20-50× with only marginal decreases in accuracy. Compared to baselines that treat privacy and communication separately, DiffSketch improves absolute test accuracy by 5%-50% while offering the same privacy guarantees and communication compression. 1 Introduction Communication costs and privacy concerns are two critical challenges in performing distributed machine learning with sensitive information. Indeed, these challenges are particularly relevant for the increasingly common problem of federated learning, in which data is collected and stored across a network of devices such as mobile phones or wearable devices [29, 34]. Communicating data across such a distributed network in order to learn an aggregate model is both costly and can potentially reveal sensitive user information [10, 35].
    [Show full text]
  • Frequency Estimation in Data Streams: Learning the Optimal Hashing Scheme
    IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 1, NO. 1, JULY 2020 1 Frequency Estimation in Data Streams: Learning the Optimal Hashing Scheme Dimitris Bertsimas and Vassilis Digalakis Jr. Abstract—We present a novel approach for the problem of frequency estimation in data streams that is based on optimization and machine learning. Contrary to state-of-the-art streaming frequency estimation algorithms, which heavily rely on random hashing to maintain the frequency distribution of the data steam using limited storage, the proposed approach exploits an observed stream prefix to near-optimally hash elements and compress the target frequency distribution. We develop an exact mixed-integer linear optimization formulation, as well as an efficient block coordinate descent algorithm, that enable us to compute near-optimal hashing schemes for elements seen in the observed stream prefix; then, we use machine learning to hash unseen elements. We empirically evaluate the proposed approach on real-world search query data and show that it outperforms existing approaches by one to two orders of magnitude in terms of its average (per element) estimation error and by 45-90% in terms of its expected magnitude of estimation error. Index Terms—Data streams, streaming frequency estimation, learning to hash, optimal hashing scheme. F 1 INTRODUCTION E consider a streaming model of computation [1], [2], query appears in the stream, in space much smaller than the total W where the input is represented as a finite sequence of number of unique queries. elements from some finite universe (domain) which is not available for random access, but instead arrives dynamically Massive data streams appear in a variety of applications.
    [Show full text]
  • Question Type Guided Attention in Visual Question Answering
    Question Type Guided Attention in Visual Question Answering Yang Shi1 ⋆, Tommaso Furlanello2, Sheng Zha3, and Animashree Anandkumar34 1 University of California, Irvine [email protected] 2 University of Southern California [email protected] 3 Amazon AI {zhasheng},{anima}@amazon.com 4 California Institute of Technology Abstract. Visual Question Answering (VQA) requires integration of feature maps with drastically different structures. Image descriptors have structures at mul- tiple spatial scales, while lexical inputs inherently follow a temporal sequence and naturally cluster into semantically different question types. A lot of previous works use complex models to extract feature representations but neglect to use high-level information summary such as question types in learning. In this work, we propose Question Type-guided Attention (QTA). It utilizes the information of question type to dynamically balance between bottom-up and top-down vi- sual features, respectively extracted from ResNet and Faster R-CNN networks. We experiment with multiple VQA architectures with extensive input ablation studies over the TDIUC dataset and show that QTA systematically improves the performance by more than 5% across multiple question type categories such as “Activity Recognition”, “Utility” and “Counting” on TDIUC dataset compared to the state-of-art. By adding QTA on the state-of-art model MCB, we achieve 3% improvement in overall accuracy. Finally, we propose a multi-task extension to predict question types which generalizes QTA to applications that lack question type, with a minimal performance loss. Keywords: Visual question answering, Attention, Question type, Feature selec- tion, Multi-task 1 Introduction The relative maturity and flexibility of deep learning allow us to build upon the suc- cess of computer vision [17] and natural language [13, 20] to face new complex and multimodal tasks.
    [Show full text]
  • Finding Heavily-Weighted Features with the Weight-Median Sketch
    Finding Heavily-Weighted Features with the Weight-Median Sketch Extended Abstract Kai Sheng Tai, Vatsal Sharan, Peter Bailis, Gregory Valiant Stanford University ABSTRACT update “foo” 2.5 We introduce the Weight-Median Sketch, a sub-linear space data “bar” -1.9 (xt ;yt ) rLˆt structure that captures the most heavily weighted features in linear “baz” 1.8 classifiers trained over data streams. This enables memory-limited query execution of several statistical analyses over streams, including on- streaming gradient sketched estimates of line feature selection, streaming data explanation, relative deltoid data estimates classifier largest weights detection, and streaming estimation of pointwise mutual informa- Figure 1: Streaming updates to a sketched classifier with ap- tion. On a standard binary classification benchmark, the Weight- proximations of the most heavily-weighted features. Median Sketch enables estimation of the top-100 weights with 10% relative error using two orders of magnitude less space than an We introduce a new sketching algorithm—the Weight-Median uncompressed classifier. Sketch (WM-Sketch)—for binary linear classifiers trained on stream- ing data.1 As each new labeled example, (x;y) with x 2 Rd , y 2 1 INTRODUCTION f−1;+1g, is observed in the stream, we update a fixed-size data struc- ture in memory with online gradient descent. This data structure Memory-efficient sketching algorithms are a well-established tool represents a compressed version of the full d-dimensional classifier in stream processing tasks such as frequent item identification [4, trained on the same data stream. This structure can be efficiently 5, 16], quantile estimation [10], and approximate counting of dis- queried for estimates of the top-K highest-magnitude weights in the tinct items [9].
    [Show full text]