We introduce a new convolutional neural network architecture with the ability to adapt dynamically to computational resource limits at test time. Our network architecture uses progressively growing multi-scale convolutions and dense connectivity, which allows for the training of multiple classifiers at intermediate layers of the network. We evaluate our approach in two settings: (1) anytime classification, where the network's prediction for a test example is progressively updated, facilitating the output of a prediction at any time; and (2) budgeted batch classification, where a fixed amount of computation is available to classify a set of examples that can be spent unevenly across"easier"and"harder"inputs. Experiments on three image-classification datasets demonstrate that our proposed framework substantially improves the state-of-the-art in both settings.
Multi-Scale Dense Convolutional Networks for Efficient Prediction
Gao Huang,Tianhong Li,Felix Wu,L. Maaten,Kilian Q. Weinberger
Published 2017 in arXiv.org
ABSTRACT
PUBLICATION RECORD
- Publication year
2017
- Venue
arXiv.org
- Publication date
2017-03-29
- Fields of study
Computer Science
- Identifiers
- External record
- Source metadata
Semantic Scholar
CITATION MAP
EXTRACTION MAP
CLAIMS
CONCEPTS
- anytime classification
A prediction setting in which the model can output an updated classification at any point during computation.
Aliases: anytime prediction
- budgeted batch classification
A batch prediction setting in which a fixed computation budget is allocated unevenly across examples.
Aliases: budgeted prediction
- dense connectivity
A layer-connection pattern in which later layers receive dense connections from earlier layers.
Aliases: dense connections
- intermediate layer classifiers
Classifiers attached to intermediate network layers so the model can produce predictions before the final layer.
Aliases: intermediate classifiers, early-exit classifiers
- multi-scale dense convolutional networks
A convolutional neural network architecture that combines multi-scale convolutions with dense connections and intermediate classifiers.
Aliases: MSDNet
- progressively growing multi-scale convolutions
Convolutional layers whose scale or receptive field grows progressively across depth in the network.
Aliases: progressive multi-scale convolutions
- test-time computational resource limits
The compute or latency constraints available when a model is being used for inference.
Aliases: resource limits at test time, computational budget at test time
- three image-classification datasets
The three benchmark image datasets used to evaluate the proposed network in the abstract.
Aliases: three datasets
REFERENCES
Showing 1-45 of 45 references · Page 1 of 1