Universal representations: The missing link between faces, text, planktons, and cat breeds

By Hakan Bilen, Andrea Vedaldi. 2017

154 citations recorded by the source corpus.

With the advent of large labelled datasets and high-capacity models, the performance of machine vision systems has been improving rapidly. However, the technology has still major limitations, starting from the fact that different vision problems are still solved by different models, trained from scratch or fine-tuned on the target data. The human visual system, in stark contrast, learns a universal representation for vision in the early life of an individual. This representation works well for an enormous variety of vision problems, with little or no change, with the major advantage of requiring little training data to solve any of them.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

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

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Incremental Learning Through Deep Adaptation (2017, methodology): “Like [2] , we select 80% for training and 20% for validation in datasets where no fixed split is provided.”
GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks (2017, background): “…processing (Hashimoto et al. (2016); Collobert & We-ston (2008); Søgaard & Goldberg (2016)) to speech synthesis (Wu et al. (2015); Seltzer & Droppo (2013)), from very domain-speciﬁc applications like trafﬁc prediction (Huang et al. (2014)) to very general cross-domain work (Bilen & Vedaldi (2017)).”
Beyond Shared Hierarchies: Deep Multitask Learning through Soft Layer Ordering (2017, methodology): “Such techniques have been applied across settings such as vision (Bilen and Vedaldi, 2016; 2017; Jou and Chang, 2016; Lu et al., 2017; Misra et al., 2016; Ranjan et al., 2016; Yang and Hospedales, 2017; Zhang et al., 2014), natural language (Collobert and Weston, 2008; Dong et al., 2015; Hashimoto…”
Meta-Learning with Warped Gradient Descent (2019, background): “Intriguingly, the approximation hints at the close relationship with multi-task learning (Li & Hoiem, 2016; Bilen & Vedaldi, 2017; Rebufﬁ et al., 2017); it differs in the expectation over parameter space.”
MSeg: A Composite Dataset for Multi-Domain Semantic Segmentation (2020, background): “Bilen and Vedaldi [2] propose to learn a unified representation and eliminate domain-specific scaling factors using instance normalization.”
Side-Tuning: A Baseline for Network Adaptation via Additive Side Networks (2019, background): “Other works like [24,1,25] learn taskdependent parameters (e.”
NeurAll: Towards a Unified Model for Visual Perception in Automated Driving (2019, methodology): “Other papers [24], [25] demonstrated that universality not only applies to different domains but also to different modalities.”
Improving Task Adaptation for Cross-domain Few-shot Learning (2021, background): “Learning domain-agnostic or universal representations [3] has been key to the success of cross-domain generalization.”
Universal Representation Learning from Multiple Domains for Few-shot Classification (2021, background): “Alternatively, [3, 41, 42] propose to learn a single network to perform image classification on very different domains by sharing a large majority of parameters across domains and encoding domain-specific information via normalization layers [3], light-weight residual adapters [41, 42], Feature-wise”
What and Where: Learn to Plug Adapters via NAS for Multidomain Learning (2020, methodology): “with different adapter structures, including RA [1], BN [16], DAN [17], PA [2].”
Episodic Training for Domain Generalization (2019, background): “Multi-Domain Learning (MDL) MDL aims to learn several domains simultaneously using a single model [3, 28, 29, 39].”
Side-Tuning: Network Adaptation via Additive Side Networks (2019, background): “[1, 25] learn across different tasks with additional task-speciﬁc parameters.”
Generalization in multitask deep neural classifiers: a statistical physics approach (2019, background): “Further, multitask learning also provides an interesting strategy for learning a single universal representation for many tasks possibly across multiple domains [19, 20, 21].”