Matrix capsules with EM routing

By Geoffrey E. Hinton, Sara Sabour, Nicholas Frosst. 2018

1,142 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Model Architectures

Citations in the corpus

Pathways: Asynchronous Distributed Dataflow for ML (2022, background): “Models like Mixture of Experts (MoE) (Shazeer et al., 2017) and routed capsule networks (Hinton et al., 2018; Barham and Isard, 2019) exploit computational sparsity by “routing” different (sub-)examples to the accelerators hosting different subsets of model weights based on learned functions that…”
Object-Centric Learning with Slot Attention (2020, methodology): “Iterative routing Our iterative attention mechanism shares similarlities with iterative routing mechanisms typically employed in variants of Capsule Networks [12, 13, 62].”
Local Relation Networks for Image Recognition (2019, methodology): “A representative work is Capsule Networks [23, 11], in which composability is computed by an iterative routing process.”
Dynamic Neural Networks: A Survey (2021, background): “[14], [95] also perform dynamic routing between capsules, i.”
Greedy AutoAugment (2019, background): “Another example of this approach is CapsuleNet, which tries to find the relevant pose information automatically [17], [18], [19].”
On the Binding Problem in Artificial Neural Networks (2020, background): “The main example of category slots are capsules (Hinton et al., 2011, 2018), although other approaches such as Recurrent Entity Networks (Henaff et al., 2017) can also be viewed from this perspective.”
Recurrent Independent Mechanisms (2019, background): “Capsules and Dynamic Routing: EM Capsules (Hinton et al., 2018) and the preceding Dynamic Capsules (Sabour et al.”
How to Represent Part-Whole Hierarchies in a Neural Network (2021, methodology): “The EM capsule model (Hinton et al., 2018) attempts to activate capsules that represent wholes (e.”
General E(2)-Equivariant Steerable CNNs (2019, methodology): “First, we try to use squashing nonlinearities [45] (row 37) instead of norm-ReLUs on each non-trivial irrep.”
Automatically Composing Representation Transformations as a Means for Generalization (2018, background): “For example, [24, 42, 67] have begun to examine how to route through an ﬁxed architecture to encapsulate computation, but the hard limitation that prevents ﬁxed architectures from scaling to problems with complexity beyond that which they were trained is their limited number of computation steps.”
A General Theory of Equivariant CNNs on Homogeneous Spaces (2018, methodology): “Our use of fields (with ρ block-diagonal) can be viewed as a formalization of convolutional capsules [22, 23].”
CubeNet: Equivariance to 3D Rotation and Translation (2018, background): “[24] achieved highly impressive results on the MNIST dataset with capsule networks by learning approximations to affine equivariance.”
Capsule-forensics: Using Capsule Networks to Detect Forged Images and Videos (2018, background): “Now, with the introduction of the dynamic routing algorithm [15] and the expectationmaximization routing algorithm [16], capsule networks have been implemented with remarkable initial results.”
Fast dynamic routing based on weighted kernel density estimation (2018, neutral): “algorithm is not well ensured. During our experiments, we also observed that stack of algorithm 3 can easily lead to failures of training. So we remove the method from comparison. FREM and EM routing [6]. The proposed FREM can be treated as a simplied version of EM routing. The detailed analysis of ”
A Survey of Transformers (2021, background): “These two works both utilizes two routing mechanisms, namely dynamic routing[111] and EM routing[53].”
COVID-CAPS: A capsule network-based framework for identification of COVID-19 cases from X-ray images (2020, positive): “Capsule Networks (CapsNets) [12] are alternative models that are capable of capturing spatial information using routing by agreement, through which Capsules try to reach a mutual agreement on the existence of the objects.”
Reinforcement Learning based Recommender Systems: A Survey (2021, methodology): “The main contribution in CapDRL [95] is to use CapsNet [128] for feature extraction.”
Multi-Interest Network with Dynamic Routing for Recommendation at Tmall (2019, methodology): “ng representation vectors for resulted clusters. Since the design of multi-interest extractor layer is inspired by the recently proposed dynamic routing for representation learning in capsule network [13, 14, 21], we firstly revisit essential basics in order to make this paper self-contained. 3.3.1 ”