Deep Sets

03/10/2017 ∙ by Manzil Zaheer, et al. ∙ Carnegie Mellon University 0

In this paper, we study the problem of designing objective functions for machine learning problems defined on finite sets. In contrast to traditional objective functions defined for machine learning problems operating on finite dimensional vectors, the new objective functions we propose are operating on finite sets and are invariant to permutations. Such problems are widespread, ranging from estimation of population statistics poczos13aistats, via anomaly detection in piezometer data of embankment dams Jung15Exploration, to cosmology Ntampaka16Dynamical,Ravanbakhsh16ICML1. Our main theorem characterizes the permutation invariant objective functions and provides a family of functions to which any permutation invariant objective function must belong. This family of functions has a special structure which enables us to design a deep network architecture that can operate on sets and which can be deployed on a variety of scenarios including both unsupervised and supervised learning tasks. We demonstrate the applicability of our method on population statistic estimation, point cloud classification, set expansion, and image tagging.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 8

page 18

page 19

page 20

Code Repositories

Point-clouds-evolution

Deep neural networks for point-clouds evolution


view repo

MNIST-nonlinear

Baby exercise to learn non-linear transformations on the MNIST dataset using deep sets.


view repo
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

A typical machine learning algorithm, like regression or classification, is designed for fixed dimensional data instances. Their extensions to handle the case when the inputs or outputs are permutation invariant sets rather than fixed dimensional vectors is not trivial and researchers have only recently started to investigate them Oliva et al. (2013); Szabo et al. (2016); Muandet et al. (2013, 2012). In this paper, we present a generic framework to deal with the setting where input and possibly output instances in a machine learning task are sets.

Similar to fixed dimensional data instances, we can characterize two learning paradigms in case of sets. In supervised learning, we have an output label for a set that is invariant or equivariant to the permutation of set elements. Examples include tasks like estimation of population statistics Poczos et al. (2013), where applications range from giga-scale cosmology Ntampaka et al. (2016); Ravanbakhsh et al. (2016a) to nano-scale quantum chemistry Faber et al. (2016).

Next, there can be the unsupervised setting, where the “set” structure needs to be learned, e.g. by leveraging the homophily/heterophily tendencies within sets. An example is the task of set expansion (a.k.a. audience expansion), where given a set of objects that are similar to each other (e.g. set of words {lion, tiger, leopard}), our goal is to find new objects from a large pool of candidates such that the selected new objects are similar to the query set (e.g. find words like jaguar or cheetah among all English words). This is a standard problem in similarity search and metric learning, and a typical application is to find new image tags given a small set of possible tags. Likewise, in the field of computational advertisement, given a set of high-value customers, the goal would be to find similar people. This is an important problem in many scientific applications, e.g. given a small set of interesting celestial objects, astrophysicists might want to find similar ones in large sky surveys.

Main contributions.

In this paper, (i) we propose a fundamental architecture, DeepSets, to deal with sets as inputs and show that the properties of this architecture are both necessary and sufficient (Sec. 2). (ii) We extend this architecture to allow for conditioning on arbitrary objects, and (iii) based on this architecture we develop a deep network that can operate on sets with possibly different sizes (Sec. 3). We show that a simple parameter-sharing scheme enables a general treatment of sets within supervised and semi-supervised settings. (iv) Finally, we demonstrate the wide applicability of our framework through experiments on diverse problems (Sec. 4).

2 Permutation Invariance and Equivariance

2.1 Problem Definition

A function transforms its domain into its range . Usually, the input domain is a vector space and the output response range is either a discrete space, e.g. in case of classification, or a continuous space in case of regression. Now, if the input is a set , i.e., the input domain is the power set , then we would like the response of the function to be “indifferent” to the ordering of the elements. In other words, A function acting on sets must be permutation invariant to the order of objects in the set, i.e. for any permutation . In the supervised setting, given examples of of as well as their labels

, the task would be to classify/regress (with variable number of predictors) while being permutation invariant w.r.t. predictors. Under unsupervised setting, the task would be to assign high scores to valid sets and low scores to improbable sets. These scores can then be used for set expansion tasks, such as image tagging or audience expansion in field of computational advertisement. In

transductive setting, each instance has an associated labeled . Then, the objective would be instead to learn a permutation equivariant function that upon permutation of the input instances permutes the output labels, i.e. for any permutation :

(1)

2.2 Structure

We want to study the structure of functions on sets. Their study in total generality is extremely difficult, so we analyze case-by-case. We begin by analyzing the invariant case when is a countable set and , where the next theorem characterizes its structure.

A function operating on a set having elements from a countable universe, is a valid set function, i.e., invariant to the permutation of instances in , iff it can be decomposed in the form , for suitable transformations and .

The extension to case when is uncountable, like , we could only prove that holds for sets of fixed size. The proofs and difficulties in handling the uncountable case, are discussed in Appendix  A. However, we still conjecture that exact equality holds in general.

Next, we analyze the equivariant case when and

is restricted to be a neural network layer. The standard neural network layer is represented as

where is the weight vector and

is a nonlinearity such as sigmoid function. The following lemma states the necessary and sufficient conditions for permutation-equivariance in this type of function. The function

defined above is permutation equivariant iff all the off-diagonal elements of are tied together and all the diagonal elements are equal as well. That is,

This result can be easily extended to higher dimensions, i.e., when can be matrices.

2.3 Related Results

The general form of Theorem 2.2 is closely related with important results in different domains. Here, we quickly review some of these connections.

de Finetti theorem.

A related concept is that of an exchangeable model in Bayesian statistics, It is backed by deFinetti’s theorem which states that any exchangeable model can be factored as

(2)

where is some latent feature and

are the hyper-parameters of the prior. To see that this fits into our result, let us consider exponential families with conjugate priors, where we can analytically calculate the integral of (

2). In this special case and . Now if we marginalize out , we get a form which looks exactly like the one in Theorem 2.2

(3)
Representer theorem and kernel machines.

Support distribution machines use as the prediction function Muandet et al. (2012); Poczos et al. (2012), where are distributions and . In practice, the distributions are never given to us explicitly, usually only i.i.d. sample sets are available from these distributions, and therefore we need to estimate kernel using these samples. A popular approach is to use , where is another kernel operating on the samples and . Now, these prediction functions can be seen fitting into the structure of our Theorem.

Spectral methods.

A consequence of the polynomial decomposition is that spectral methods Anandkumar et al. (2012) can be viewed as a special case of the mapping : in that case one can compute polynomials, usually only up to a relatively low degree (such as ), to perform inference about statistical properties of the distribution. The statistics are exchangeable in the data, hence they could be represented by the above map.

3 Deep Sets

3.1 Architecture

Invariant model. The structure of permutation invariant functions in Theorem 2.2 hints at a general strategy for inference over sets of objects, which we call DeepSets. Replacing and by universal approximators leaves matters unchanged, since, in particular, and can be used to approximate arbitrary polynomials. Then, it remains to learn these approximators, yielding in the following model:

  • [leftmargin=3mm, itemsep=0mm,partopsep=0pt,parsep=0pt]

  • Each instance is transformed (possibly by several layers) into some representation .

  • The representations are added up and the output is processed using the network in the same manner as in any deep network (e.g. fully connected layers, nonlinearities, etc.).

  • Optionally: If we have additional meta-information , then the above mentioned networks could be conditioned to obtain the conditioning mapping .

In other words, the key is to add up all representations and then apply nonlinear transformations.

Equivariant model. Our goal is to design neural network layers that are equivariant to the permutations of elements in the input . Based on Lemma 2.2, a neural network layer is permutation equivariant if and only if all the off-diagonal elements of are tied together and all the diagonal elements are equal as well, i.e., for . This function is simply a non-linearity applied to a weighted combination of (i) its input and; (ii) the sum of input values . Since summation does not depend on the permutation, the layer is permutation-equivariant. We can further manipulate the operations and parameters in this layer to get other variations, e.g.:

(4)

where the maxpooling operation over elements of the set (similar to sum) is commutative. In practice, this variation performs better in some applications. This may be due to the fact that for , the input to the non-linearity is max-normalized. Since composition of permutation equivariant functions is also permutation equivariant, we can build DeepSets by stacking such layers.

3.2 Other Related Works

Several recent works study equivariance and invariance in deep networks w.r.t. general group of transformations Gens and Domingos (2014); Cohen and Welling (2016); Ravanbakhsh et al. (2017). For example, Chen et al. (2014) construct deep permutation invariant features by pairwise coupling of features at the previous layer, where is invariant to transposition of and . Pairwise interactions within sets have also been studied in Chang et al. (2016); Guttenberg et al. (2016). Vinyals et al. (2015) approach unordered instances by finding “good” orderings.

The idea of pooling a function across set-members is not new. In Lopez-Paz et al. (2016), pooling was used binary classification task for causality on a set of samples. Shi et al. (2015) use pooling across a panoramic projection of 3D object for classification, while Su et al. (2015) perform pooling across multiple views. Hartford et al. (2016) observe the invariance of the payoff matrix in normal form games to the permutation of its rows and columns (i.e.

player actions) and leverage pooling to predict the player action. The need of permutation equivariance also arise in deep learning over sensor networks and multi-agent setings, where a special case of Lemma 

2.2 has been used as the architecture Sukhbaatar et al. (2016).

In light of these related works, we would like to emphasize our novel contributions: (i) the universality result of Theorem 2.2 for permutation invariance that also relates DeepSets to other machine learning techniques, see Sec. 3; (ii) the permutation equivariant layer of (4), which, according to Lemma 2.2 identifies necessary and sufficient form of parameter-sharing in a standard neural layer and; (iii) novel application settings that we study next.

4 Applications and Empirical Results

We present a diverse set of applications for DeepSets. For the supervised setting, we apply DeepSets to estimation of population statistics, sum of digits and classification of point-clouds, and regression with clustering side-information. The permutation-equivariant variation of DeepSets is applied to the task of outlier detection. Finally, we investigate the application of DeepSets to unsupervised set-expansion, in particular, concept-set retrieval and image tagging. In most cases we compare our approach with the state-of-the art and report competitive results.

4.1 Set Input Scalar Response

4.1.1 Supervised Learning: Learning to Estimate Population Statistics

(a) Entropy estimation for rotated of 2d Gaussian
(b) Mutual information estimation by varying correlation
(c) Mutual information estimation by varying rank-1 strength
(d) Mutual information on 32d random covariance matrices
Figure 1: Population statistic estimation: Top set of figures, show prediction of DeepSets vs SDM for case. Bottom set of figures, depict the mean squared error behavior as number of sets is increased. SDM has lower error for small and DeepSets requires more data to reach similar accuracy. But for high dimensional problems DeepSets easily scales to large number of examples and produces much lower estimation error. Note that the matrix inversion in SDM makes it prohibitively expensive for .

In the first experiment, we learn entropy and mutual information of Gaussian distributions, without providing any information about Gaussianity to DeepSets. The Gaussians are generated as follows:

  • [leftmargin=3mm, itemsep=0.7mm,partopsep=0mm,parsep=0mm]

  • Rotation: We randomly chose a covariance matrix , and then generated sample sets from of size for random values of . Our goal was to learn the entropy of the marginal distribution of first dimension. is the rotation matrix.

  • Correlation: We randomly chose a covariance matrix for , and then generated sample sets from of size for random values of . Goal was to learn the mutual information of among the first and last dimension.

  • Rank 1: We randomly chose and then generated a sample sets from of size for random values of . Goal was to learn the mutual information.

  • Random: We chose random covariance matrices for , and using each, generated a sample set from of size . Goal was to learn the mutual information.

We train using

loss with a DeepSets architecture having 3 fully connected layers with ReLU activation for both transformations

and . We compare against Support Distribution Machines (SDM) using a RBF kernel Poczos et al. (2012), and analyze the results in Fig. 1.

4.1.2 Sum of Digits

Figure 2: Accuracy of digit summation with text (left) and image (right) inputs. All approaches are trained on tasks of length 10 at most, tested on examples of length up to 100. We see that DeepSets generalizes better.

Next, we compare to what happens if our set data is treated as a sequence. We consider the task of finding sum of a given set of digits. We consider two variants of this experiment:

Text.

We randomly sample a subset of maximum digits from this dataset to build “sets” of training images, where the set-label is sum of digits in that set. We test against sums of digits, for starting from 5 all the way up to 100 over another examples.

Image.

MNIST8m Loosli et al. (2007) contains 8 million instances of grey-scale stamps of digits in . We randomly sample a subset of maximum images from this dataset to build “sets” of training and sets of test images, where the set-label is the sum of digits in that set (i.e. individual labels per image is unavailable). We test against sums of images of MNIST digits, for starting from 5 all the way up to 50.

We compare against recurrent neural networks – LSTM and GRU. All models are defined to have similar number of layers and parameters. The output of all models is a scalar, predicting the sum of

digits. Training is done on tasks of length 10 at most, while at test time we use examples of length up to 100. The accuracy, i.e. exact equality after rounding, is shown in Fig. 2. DeepSets generalize much better. Note for image case, the best classification error for single digit is around for MNIST8m, so in a collection of of images at least one image will be misclassified is , which is 40% for . This matches closely with observed value in Fig. 2(b).

4.1.3 Point Cloud Classification

Model Instance Size Representation Accuracy
3DShapeNets Wu et al. (2015) voxels (using convolutional deep belief net)
VoxNet Maturana and Scherer (2015) voxels (voxels from point-cloud + 3D CNN)
MVCNN Su et al. (2015) multi-vew images (2D CNN + view-pooling)
VRN Ensemble Brock et al. (2016)

voxels (3D CNN, variational autoencoder)

3D GAN Wu et al. (2016) voxels (3D CNN, generative adversarial training)
DeepSets point-cloud
DeepSets point-cloud
Table 1: Classification accuracy and the representation-size used by different methods on the ModelNet40.

A point-cloud is a set of low-dimensional vectors. This type of data is frequently encountered in various applications like robotics, vision, and cosmology. In these applications, existing methods often convert the point-cloud data to voxel or mesh representation as a preprocessing step, e.g. Maturana and Scherer (2015); Ravanbakhsh et al. (2016b); Lin et al. (2004). Since the output of many range sensors, such as LiDAR, is in the form of point-cloud, direct application of deep learning methods to point-cloud is highly desirable. Moreover, it is easy and cheaper to apply transformations, such as rotation and translation, when working with point-clouds than voxelized 3D objects.

As point-cloud data is just a set of points, we can use DeepSets to classify point-cloud representation of a subset of ShapeNet objects Chang et al. (2015), called ModelNet40 Wu et al. (2015). This subset consists of 3D representation of 9,843 training and 2,468 test instances belonging to 40 classes of objects. We produce point-clouds with 100, 1000 and 5000 particles each (-coordinates) from the mesh representation of objects using the point-cloud-library’s sampling routine Rusu and Cousins (2011)

. Each set is normalized by the initial layer of the deep network to have zero mean (along individual axes) and unit (global) variance. Tab. 

1 compares our method using three permutation equivariant layers against the competition; see Appendix  H for details.

4.1.4 Improved Red-shift Estimation Using Clustering Information

An important regression problem in cosmology is to estimate the red-shift of galaxies, corresponding to their age as well as their distance from us Binney and Merrifield (1998) based on photometric observations. One way to estimate the red-shift from photometric observations is using a regression model Connolly et al. (1995) on the galaxy clusters. The prediction for each galaxy does not change by permuting the members of the galaxy cluster. Therefore, we can treat each galaxy cluster as a “set” and use DeepSets to estimate the individual galaxy red-shifts. See Appendix  G for more details.

Method Scatter
MLP 0.026
redMaPPer 0.025
DeepSets 0.023
Table 2: Red-shift experiment. Lower scatter is better.

For each galaxy, we have photometric features from the redMaPPer galaxy cluster catalog Rozo and Rykoff (2014) that contains photometric readings for 26,111 red galaxy clusters. Each galaxy-cluster in this catalog has between galaxies – i.e. , where is the cluster-size. The catalog also provides accurate spectroscopic red-shift estimates for a subset of these galaxies.

We randomly split the data into 90% training and 10% test clusters, and minimize the squared loss of the prediction for available spectroscopic red-shifts. As it is customary in cosmology literature, we report the average scatter , where is the accurate spectroscopic measurement and is a photometric estimate in Tab. 2.

4.2 Set Expansion

In the set expansion task, we are given a set of objects that are similar to each other and our goal is to find new objects from a large pool of candidates such that the selected new objects are similar to the query set. To achieve this one needs to reason out the concept connecting the given set and then retrieve words based on their relevance to the inferred concept. It is an important task due to wide range of potential applications including personalized information retrieval, computational advertisement, tagging large amounts of unlabeled or weakly labeled datasets.

Going back to de Finetti’s theorem in Sec. 3.2

, where we consider the marginal probability of a set of observations, the marginal probability allows for very simple metric for scoring additional elements to be added to

. In other words, this allows one to perform set expansion via the following score

(5)

Note that is the point-wise mutual information between and . Moreover, due to exchangeability, it follows that regardless of the order of elements we have

(6)

When inferring sets, our goal is to find set completions for an initial set of query terms , such that the aggregate set is coherent. This is the key idea of the Bayesian Set algorithm Ghahramani and Heller (2005) (details in Appendix  D). Using DeepSets, we can solve this problem in more generality as we can drop the assumption of data belonging to certain exponential family.

For learning the score

, we take recourse to large-margin classification with structured loss functions

Taskar et al. (2004) to obtain the relative loss objective . In other words, we want to ensure that whenever should be added and should not be added to .

Conditioning. Often machine learning problems do not exist in isolation. For example, task like tag completion from a given set of tags is usually related to an object , for example an image, that needs to be tagged. Such meta-data are usually abundant, e.g. author information in case of text, contextual data such as the user click history, or extra information collected with LiDAR point cloud.

Conditioning graphical models with meta-data is often complicated. For instance, in the Beta-Binomial model we need to ensure that the counts are always nonnegative, regardless of

. Fortunately, DeepSets does not suffer from such complications and the fusion of multiple sources of data can be done in a relatively straightforward manner. Any of the existing methods in deep learning, including feature concatenation by averaging, or by max-pooling, can be employed. Incorporating these meta-data often leads to significantly improved performance as will be shown in experiments; Sec. 

4.2.2.

4.2.1 Text Concept Set Retrieval

Method LDA- (Vocab = ) LDA- (Vocab = ) LDA- (Vocab = )
Recall (%) MRR Med. Recall (%) MRR Med. Recall (%) MRR Med.
@10 @100 @1k @10 @100 @1k @10 @100 @1k
Random 0.06 0.6 5.9 0.001 8520 0.02 0.2 2.6 0.000 28635 0.01 0.2 1.6 0.000 30600
Bayes Set 1.69 11.9 37.2 0.007 2848 2.01 14.5 36.5 0.008 3234 1.75 12.5 34.5 0.007 3590
w2v Near 6.00 28.1 54.7 0.021 641 4.80 21.2 43.2 0.016 2054 4.03 16.7 35.2 0.013 6900
NN-max 4.78 22.5 53.1 0.023 779 5.30 24.9 54.8 0.025 672 4.72 21.4 47.0 0.022 1320
NN-sum-con 4.58 19.8 48.5 0.021 1110 5.81 27.2 60.0 0.027 453 4.87 23.5 53.9 0.022 731
NN-max-con 3.36 16.9 46.6 0.018 1250 5.61 25.7 57.5 0.026 570 4.72 22.0 51.8 0.022 877
DeepSets 5.53 24.2 54.3 0.025 696 6.04 28.5 60.7 0.027 426 5.54 26.1 55.5 0.026 616
Table 3: Results on Text Concept Set Retrieval on LDA-1k, LDA-3k, and LDA-5k. Our DeepSets model outperforms other methods on LDA-3k and LDA-5k. However, all neural network based methods have inferior performance to w2v-Near baseline on LDA-1k, possibly due to small data size. Higher the better for recall@k and mean reciprocal rank (MRR). Lower the better for median rank (Med.)

In text concept set retrieval, the objective is to retrieve words belonging to a ‘concept’ or ‘cluster’, given few words from that particular concept. For example, given the set of words {tiger, lion, cheetah}, we would need to retrieve other related words like jaguar, puma, etc, which belong to the same concept of big cats. This task of concept set retrieval can be seen as a set completion task conditioned on the latent semantic concept, and therefore our DeepSets form a desirable approach.

Dataset.

We construct a large dataset containing sets of related words by extracting topics from latent Dirichlet allocation Pritchard et al. (2000); Blei et al. (2003), taken out-of-the-box111github.com/dmlc/experimental-lda. To compare across scales, we consider three values of giving us three datasets LDA-, LDA-, and LDA-, with corresponding vocabulary sizes of and .

Methods.

We learn this using a margin loss with a DeepSets architecture having 3 fully connected layers with ReLU activation for both transformations and . Details of the architecture and training are in Appendix  E. We compare to several baselines: (a) Random picks a word from the vocabulary uniformly at random. (b) Bayes Set Ghahramani and Heller (2005). (c) w2v-Near computes the nearest neighbors in the word2vec Mikolov et al. (2013)

space. Note that both Bayes Set and w2v NN are strong baselines. The former runs Bayesian inference using Beta-Binomial conjugate pair, while the latter uses the powerful

dimensional word2vec trained on the billion word GoogleNews corpus222code.google.com/archive/p/word2vec/. (d) NN-max uses a similar architecture as our DeepSets but uses max pooling to compute the set feature, as opposed to sum pooling. (e) NN-max-con uses max pooling on set elements but concatenates this pooled representation with that of query for a final set feature. (f) NN-sum-con is similar to NN-max-con but uses sum pooling followed by concatenation with query representation.

Evaluation.

We consider the standard retrieval metrics – recall@K, median rank and mean reciprocal rank, for evaluation. To elaborate, recall@K measures the number of true labels that were recovered in the top K retrieved words. We use three values of K. The other two metrics, as the names suggest, are the median and mean of reciprocals of the true label ranks, respectively. Each dataset is split into TRAIN (), VAL () and TEST (

). We learn models using TRAIN and evaluate on TEST, while VAL is used for hyperparameter selection and early stopping.

Results and Observations.

As seen in Tab. 3: (a) Our DeepSets model outperforms all other approaches on LDA- and LDA- by any metric, highlighting the significance of permutation invariance property. (b) On LDA-, our model does not perform well when compared to w2v-Near. We hypothesize that this is due to small size of the dataset insufficient to train a high capacity neural network, while w2v-Near has been trained on a billion word corpus. Nevertheless, our approach comes the closest to w2v-Near amongst other approaches, and is only 0.5% lower by Recall@10.

4.2.2 Image Tagging

Method ESP game IAPRTC-12.5
P R F1 N+ P R F1 N+
Least Sq. 35 19 25 215 40 19 26 198
MBRM 18 19 18 209 24 23 23 223
JEC 24 19 21 222 29 19 23 211
FastTag 46 22 30 247 47 26 34 280
Least Sq.(D) 44 32 37 232 46 30 36 218
FastTag(D) 44 32 37 229 46 33 38 254
DeepSets 39 34 36 246 42 31 36 247
Table 4: Results of image tagging on ESPgame and IAPRTC-12.5 datasets. Performance of our DeepSets approach is roughly similar to the best competing approaches, except for precision. Refer text for more details. Higher the better for all metrics – precision (P), recall (R), f1 score (F1), and number of non-zero recall tags (N+).

We next experiment with image tagging, where the task is to retrieve all relevant tags corresponding to an image. Images usually have only a subset of relevant tags, therefore predicting other tags can help enrich information that can further be leveraged in a downstream supervised task. In our setup, we learn to predict tags by conditioning DeepSets on the image, i.e., we train to predict a partial set of tags from the image and remaining tags. At test time, we predict tags from the image alone.

Datasets.

We report results on the following three datasets - ESPGame, IAPRTC-12.5 and our in-house dataset, COCO-Tag. We refer the reader to Appendix  F, for more details about datasets.

Methods.

The setup for DeepSets to tag images is similar to that described in Sec. 4.2.1. The only difference being the conditioning on the image features, which is concatenated with the set feature obtained from pooling individual element representations.

Baselines.

We perform comparisons against several baselines, previously reported in Chen et al. (2013)

. Specifically, we have Least Sq., a ridge regression model, MBRM

Feng et al. (2004), JEC Makadia et al. (2008) and FastTag Chen et al. (2013)

. Note that these methods do not use deep features for images, which could lead to an unfair comparison. As there is no publicly available code for MBRM and JEC, we cannot get performances of these models with Resnet extracted features. However, we report results with deep features for FastTag and Least Sq., using code made available by the authors

333http://www.cse.wustl.edu/~mchen/.

Evaluation.

For ESPgame and IAPRTC-12.5, we follow the evaluation metrics as in

Guillaumin et al. (2009)–precision (P), recall (R), F1 score (F1), and number of tags with non-zero recall (N+). These metrics are evaluate for each tag and the mean is reported (see Guillaumin et al. (2009) for further details). For COCO-Tag, however, we use recall@K for three values of K, along with median rank and mean reciprocal rank (see evaluation in Sec. 4.2.1 for metric details).

Method Recall MRR Med.
@10 @100 @1k
w2v NN (blind) 5.6 20.0 54.2 0.021 823
DeepSets (blind) 9.0 39.2 71.3 0.044 310
DeepSets 31.4 73.4 95.3 0.131 28
Table 5: Results on COCO-Tag dataset. Clearly, DeepSets outperforms other baselines significantly. Higher the better for recall@K and mean reciprocal rank (MRR). Lower the better for median rank (Med).
Results and Observations.

Tab. 4 shows results of image tagging on ESPgame and IAPRTC-12.5, and Tab. 5 on COCO-Tag. Here are the key observations from Tab. 4: (a) performance of our DeepSets model is comparable to the best approaches on all metrics but precision, (b) our recall beats the best approach by 2% in ESPgame. On further investigation, we found that the DeepSets model retrieves more relevant tags, which are not present in list of ground truth tags due to a limited tag annotation. Thus, this takes a toll on precision while gaining on recall, yet yielding improvement on F1. On the larger and richer COCO-Tag, we see that the DeepSets approach outperforms other methods comprehensively, as expected. Qualitative examples are in Appendix  F.

4.3 Set Anomaly Detection

The objective here is to find the anomalous face in each set, simply by observing examples and without any access to the attribute values. CelebA dataset Liu et al. (2015) contains 202,599 face images, each annotated with 40 boolean attributes. We build sets of stamps, using these attributes each containing images (on the training set) as follows: randomly select 2 attributes, draw 15 images having those attributes, and a single target image where both attributes are absent. Using a similar procedure we build sets on the test images. No individual person‘s face appears in both train and test sets.

Figure 3: Each row shows a set, constructed from CelebA dataset, such that all set members except for an outlier, share at least two attributes (on the right). The outlier is identified with a red frame. The model is trained by observing examples of sets and their anomalous members, without access to the attributes. The probability assigned to each member by the outlier detection network is visualized using a red bar at the bottom of each image. The probabilities in each row sum to one.

Our deep neural network consists of 9 2D-convolution and max-pooling layers followed by 3 permutation-equivariant layers, and finally a softmax layer that assigns a probability value to each set member (Note that one could identify arbitrary number of outliers using a sigmoid activation at the output). Our trained model successfully finds the anomalous face in

75% of test sets. Visually inspecting these instances suggests that the task is non-trivial even for humans; see Fig. 3.

As a baseline, we repeat the same experiment by using a set-pooling layer after convolution layers, and replacing the permutation-equivariant layers with fully connected layers of same size, where the final layer is a 16-way softmax. The resulting network shares the convolution filters for all instances within all sets, however the input to the softmax is not equivariant to the permutation of input images. Permutation equivariance seems to be crucial here as the baseline model achieves a training and test accuracy of ; the same as random selection. See Appendix  I for more details.

5 Summary

In this paper, we develop DeepSets, a model based on powerful permutation invariance and equivariance properties, along with the theory to support its performance. We demonstrate the generalization ability of DeepSets across several domains by extensive experiments, and show both qualitative and quantitative results. In particular, we explicitly show that DeepSets outperforms other intuitive deep networks, which are not backed by theory (Sec. 4.2.1, Sec. 4.1.2). Last but not least, it is worth noting that the state-of-the-art we compare to is a specialized technique for each task, whereas our one model, i.e., DeepSets, is competitive across the board.

References

Appendix A Proofs and Discussion Related to Theorem 2

A function transforms its domain into its range . Usually, the input domain is a vector space and the output response range is either a discrete space, e.g. in case of classification, or a continuous space in case of regression.

Now, if the input is a set , i.e. , then we would like the response of the function not to depend on the ordering of the elements in the set. In other words,

Property 1   A function acting on sets must be permutation invariant to the order of objects in the set, i.e.

(7)

for any permutation .

Now, roughly speaking, we claim that such functions must have a structure of the form for some functions and . Over the next two sections we try to formally prove this structure of the permutation invariant functions.

a.1 Countable Case

Theorem 2   Assume the elements are countable, i.e. . A function operating on a set can be a valid set function, i.e. it is permutation invariant to the elements in , if and only if it can be decomposed in the form , for suitable transformations and .

Proof.

Permutation invariance follows from the fact that sets have no particular order, hence any function on a set must not exploit any particular order either. The sufficiency follows by observing that the function satisfies the permutation invariance condition.

To prove necessity, i.e. that all functions can be decomposed in this manner, we begin by noting that there must be a mapping from the elements to natural numbers functions, since the elements are countable. Let this mapping be denoted by . Now if we let then constitutes an unique representation for every set . Now a function can always be constructed such that . ∎

a.2 Uncountable Case

The extension to case when is uncountable, e.g. , is not so trivial. We could only prove in case of fixed set size, e.g. instead of , that any permutation invariant continuous function can be expressed as . Also, we show that there is a universal approximator of the same form. These results are discussed below.

To illustrate the uncountable case, we assume a fixed set size of . Without loss of generality we can let . Then the domain becomes . Also, to handle ambiguity due to permutation, we often define the domain to be the set for some ordering of the elements in .

The proof builds on the famous Newton-Girard formulae which connect moments of a sample set (sum-of-power) to the elementary symmetric polynomials. But first we present some results needed for the proof. The first result establishes that sum-of-power mapping is injective.

Lemma 4 Let . The sum-of-power mapping defined by the coordinate functions

(8)

is injective.

Proof.

Suppose for some , we have . We will now show that it must be the case that . Construct two polynomials as follows:

(9)

If we expand the two polynomials we obtain:

(10)

with coefficients being elementary symmetric polynomials in and respectively, i.e.

(11)

These elementary symmetric polynomials can be uniquely expressed as a function of and respectively, by Newton-Girard formula. The -th coefficient is given by the determinant of matrix having terms from and respectively:

(12)

Since we assumed implying , which in turn implies that the polynomials and are the same. Therefore, their roots must be the same, which shows that . ∎

The second result we borrow from Ćurgus and Mascioni [2006] which establishes a homeomorphism between coefficients and roots of a polynomial.

Theorem 5 Ćurgus and Mascioni [2006] The function , which associates every to the multiset of roots, , of the monic polynomial formed using as the coefficient i.e. , is a homeomorphism.

Among other things, this implies that (complex) roots of a polynomial depends continuously on the coefficients. We will use this fact for our next lemma.

Finally, we establish a continuous inverse mapping for the sum-of-power function.

Lemma 6 Let . We define the sum-of-power mapping by the coordinate functions

(13)

where is the range of the function. The function has a continuous inverse mapping.

Proof.

First of all note that , the range of , is a compact set. This follows from following observations:

  • The domain of is a bounded polytope (i.e. a compact set),

  • is a continuous function, and

  • image of a compact set under a continuous function is a compact set.

To show the continuity of inverse mapping, we establish connection to the continuous dependence of roots of polynomials on its coefficients.

As in Lemma 4, for any , let and construct the polynomial:

(14)

If we expand the polynomial we obtain:

(15)

with coefficients being elementary symmetric polynomials in , i.e.

(16)

These elementary symmetric polynomials can be uniquely expressed as a function of by Newton-Girard formula:

(17)

Since determinants are just polynomials, is a continuous function of . Thus to show continuity of inverse mapping of , it remains to show continuity from back to the roots . In this regard, we invoke Theorem 5. Note that homeomorphism implies the mapping as well as its inverse is continuous. Thus, restricting to the compact set where the map from coefficients to roots only goes to the reals, the desired result follows. To explicitly check the continuity, note that limit of , as approaches from inside , always exists and is equal to since it does so in the complex plane. ∎

With the lemma developed above we are in a position to tackle the main theorem.

Figure 4: Outline of the proof strategy for Theorem 2.1. The proof consists of two parts. First, we desire to show that we can find unique embeddings for each possible input, i.e. we show that there exists a homeomorphism of the form between original domain and some higher dimensional space . The second part of the proof consists of showing we can map the embedding to desired target value, i.e. to show the existence of the continuous map between and original target space such that .

Theorem 7   Let be a permutation invariant continuous function iff it has the representation

(18)

for some continuous outer and inner function and respectively. The inner function is independent of the function .

Proof.

The sufficiency follows by observing that the function satisfies the permutation invariance condition.

To prove necessity, i.e. that all permutation invariant continuous functions over the compact set can be expressed in this manner, we divide the proof into two parts, with outline in Fig. 4. We begin by looking at the continuous embedding formed by the inner function: . Consider defined as . Now as is a polynomial, the image of in under is a compact set as well, denote it by . Then by definition, the embedding is surjective. Using Lemma 4 and 6, we know that upon restricting the permutations, i.e. replacing with , the embedding is injective with a continuous inverse. Therefore, combining these observation we get that is a homeomorphism between and . Now it remains to show that we can map the embedding to desired target value, i.e. to show the existence of the continuous map such that . In particular consider the map . The continuity of follows directly from the fact that composition of continuous functions is continuous. Therefore we can always find continuous functions and to express any permutation invariant function as . ∎

A very similar but more general results holds in case of any continuous function (not necessarily permutation invariant). The result is known as Kolmogorov-Arnold representation theorem [Khesin and Tabachnikov, 2014, Chap. 17] which we state below:

Theorem 8 (Kolmogorov–Arnold representation)   Let be an arbitrary multivariate continuous function iff it has the representation

(19)

with continuous outer and inner functions and . The inner function is independent of the function .

This theorem essentially states a representation theorem for any multivariate continuous function. Their representation is very similar to the one we are proved, except for the dependence of inner transformation on the co-ordinate through . Thus it is reassuring that behind all the beautiful mathematics something intuitive is happening. If the function is permutation invariant, this dependence on co-ordinate of the inner transformation gets dropped!

Further we can show that arbitrary approximator having the same form can be obtained for continuous permutation-invariant functions.

Theorem 9   Assume the elements are from a compact set in , i.e. possibly uncountable, and the set size is fixed to . Then any continuous function operating on a set , i.e. which is permutation invariant to the elements in can be approximated arbitrarily close in the form of , for suitable transformations and .

Proof.

Permutation invariance follows from the fact that sets have no particular order, hence any function on a set must not exploit any particular order either. The sufficiency follows by observing that the function satisfies the permutation invariance condition.

To prove necessity, i.e. that all continuous functions over the compact set can be approximated arbitrarily close in this manner, we begin noting that polynomials are universal approximators by Stone–Weierstrass theorem [Marsden and Hoffman, 1993, sec. 5.7]. In this case the Chevalley-Shephard-Todd (CST) theorem [Bourbaki, 1990, chap. V, theorem 4], or more precisely, its special case, the Fundamental Theorem of Symmetric Functions states that symmetric polynomials are given by a polynomial of homogeneous symmetric monomials. The latter are given by the sum over monomial terms, which is all that we need since it implies that all symmetric polynomials can be written in the form required by the theorem. ∎

Finally, we still conjecture that even in case of sets of all sizes, i.e. when the domain is , a representation of the form should exist for all “continuous” permutation invariant functions for some suitable transformations and . However, in this case even what a “continuous” function means is not clear as the space does not have any natural topology. As a future work, we want to study further by defining various topologies, like using Fréchet distance as used in Ćurgus and Mascioni [2006] or MMD distance. Our preliminary findings in this regards hints that using MMD distance if the representation is allowed to be in , instead of being finite dimensional, then the conjecture seems to be provable. Thus, clearly this direction needs further exploration. We end this section by providing some examples:

Examples:
  • , Consider and , then is the desired function.

  • , Consider and , then is the desired function.

  • , Consider and , then is the desired function.

  • , Consider and , then as , then we have approaching the desired function.

  • Second largest among , Consider and , then as , we have approaching the desired function.

Appendix B Proof of Lemma 3

Our goal is to design neural network layers that are equivariant to permutations of elements in the input . The function is equivariant to the permutation of its inputs iff

where the symmetric group is the set of all permutation of indices .

Consider the standard neural network layer

(20)

where is the weight vector and is a nonlinearity such as sigmoid function. The following lemma states the necessary and sufficient conditions for permutation-equivariance in this type of function.

Lemma 3   The function as defined in (20) is permutation equivariant if and only if all the off-diagonal elements of are tied together and all the diagonal elements are equal as well. That is,

where is the identity matrix.

Proof.

From definition of permutation equivariance and definition of in (20), the condition becomes , which (assuming sigmoid is a bijection) is equivalent to . Therefore we need to show that the necessary and sufficient conditions for the matrix to commute with all permutation matrices is given by this proposition. We prove this in both directions:

  • To see why commutes with any permutation matrix, first note that commutativity is linear – that is

    Since both Identity matrix , and constant matrix , commute with any permutation matrix, so does their linear combination .

  • We need to show that in a matrix that commutes with “all” permutation matrices

    • All diagonal elements are identical: Let for , be a transposition (i.e. a permutation that only swaps two elements). The inverse permutation matrix of is the permutation matrix of . We see that commutativity of with the transposition implies that :

      Therefore, and commute for any permutation , they also commute for any transposition and therefore .

    • All off-diagonal elements are identical: We show that since commutes with any product of transpositions, any choice two off-diagonal elements should be identical. Let and be the index of two off-diagonal elements (i.e. and ). Moreover for now assume and . Application of the transposition , swaps the rows in . Similarly, switches the column with column. From commutativity property of and we have

      where in the last step we used our assumptions that , , and . In the cases where either or , we can use the above to show that and , for some and , and conclude .

Appendix C More Details on the architecture

c.1 Invariant model

The structure of permutation invariant functions in Theorem 2.2 hints at a general strategy for inference over sets of objects, which we call deep sets. Replacing and by universal approximators leaves matters unchanged, since, in particular, and can be used to approximate arbitrary polynomials. Then, it remains to learn these approximators. This yields in the following model:

Figure 5: Architecture of DeepSets: Invariant

Each instance is transformed (possibly by several layers) into some representation .

The addition of these representations processed using the network very much in the same manner as in any deep network (e.g. fully connected layers, nonlinearities, etc).

Optionally: If we have additional meta-information , then the above mentioned networks could be conditioned to obtain the conditioning mapping .

In other words, the key to deep sets is to add up all representations and then apply nonlinear transformations.

The overall model structure is illustrated in Fig. 7.

This architecture has a number of desirable properties in terms of universality and correctness. We assume in the following that the networks we choose are, in principle, universal approximators. That is, we assume that they can represent any functional mapping. This is a well established property (see e.g. Micchelli [1986]

for details in the case