Meta-Transfer Learning for Few-Shot Learning

Sun, Quansen; Liu, Yaoyao; Chua, Tat-Seng; Schiele, Bernt

doi:10.1109/cvpr.2019.00049

Cited by 981 publications

(549 citation statements)

References 33 publications

Supporting

Mentioning

495

Contrasting

Unclassified

Order By: Relevance

“…Modulo empirical fluctuations, our method performs at the state-of-the art and in some cases exceeds it. We wish to point out that SNAIL [9], TADAM [10,18], LEO [14], MTLF [18] pre-train the network for a 64 way classification task on miniImagenet and 351 way classification on tieredImagenet. However, all the models trained for our method are trained from scratch and use no form of pre-training.…”

Section: Comparison To the State-of-the-artmentioning

confidence: 99%

Few-Shot Learning With Embedded Class Models and Shot-Free Meta Training

Ravichandran

Bhotika

Soatto

2019

2019 IEEE/CVF International Conference on Computer Vision (ICCV)

132

View full text Add to dashboard Cite

We propose a method for learning embeddings for fewshot learning that is suitable for use with any number of ways and any number of shots (shot-free). Rather than fixing the class prototypes to be the Euclidean average of sample embeddings, we allow them to live in a higherdimensional space (embedded class models) and learn the prototypes along with the model parameters. The class representation function is defined implicitly, which allows us to deal with a variable number of shots per each class with a simple constant-size architecture. The class embedding encompasses metric learning, that facilitates adding new classes without crowding the class representation space. Despite being general and not tuned to the benchmark, our approach achieves state-of-the-art performance on the standard few-shot benchmark datasets. Figure 1. One image of a mushroom (Muscaria) may be enough to recognize it in the wild (left); in other cases, there may be more subtle differences between an edible (Russula, shown in the center) and a deadly one (Phalloides, shown on the right), but still few samples are enough for humans.

show abstract

Section: Comparison To the State-of-the-artmentioning

confidence: 99%

Few-Shot Learning With Embedded Class Models and Shot-Free Meta Training

Ravichandran

Bhotika

Soatto

2019

2019 IEEE/CVF International Conference on Computer Vision (ICCV)

132

View full text Add to dashboard Cite

show abstract

“…Few-shot learning methods can be roughly categorized into two classes: data augmentation and task-based meta-learning. For example, in [15] the proposed model gave state-of-the-art results and paved the path for more sophisticated meta-transfer learning methods.…”

Section: Discussion 41 Current Challengesmentioning

confidence: 96%

Collaborative Intelligent Cross-Camera Video Analytics at Edge

Pasandi

Nadeem

2019

Proceedings of the First International Workshop on Challenges in Artificial Intelligence and Machine Learning for Internet of T

View full text Add to dashboard Cite

Today, video cameras are deployed in dense for monitoring physical places e.g., city, industrial, or agricultural sites. In the current systems, each camera node sends its feed to a cloud server individually. However, this approach suffers from several hurdles including higher computation cost, large bandwidth requirement for analyzing the enormous data, and privacy concerns. In dense deployment, video nodes typically demonstrate a significant spatio-temporal correlation. To overcome these obstacles in current approaches, this paper introduces CONVINCE, a new approach to look at the network cameras as a collective entity that enables collaborative video analytics pipeline among cameras. CON-VINCE aims at 1) reducing the computation cost and bandwidth requirements by leveraging spatio-temporal correlations among cameras in eliminating redundant frames intelligently, and ii) improving vision algorithms' accuracy by enabling collaborative knowledge sharing among relevant cameras. Our results demonstrate that CONVINCE achieves an object identification accuracy of ∼91%, by transmitting only about ∼25% of all the recorded frames.

show abstract

“…Adaptive batch normalization [16] performs domain adaptation for the segmentation task by replacing the statistics of the source domain with those of the target domain, achieving performance accuracy competitive with other deep-learning based methods. Meta-transfer learning [31] performs few-shot learning by updating only the scale and shift parameters, and achieves better performance than when fine-tuning all kernel parameters. These methods show that the scaling and shifting operation is effective for transferring knowledge to feature extractor models.…”

Section: Transfer Learning With Scale and Shiftmentioning

confidence: 99%

Image Generation From Small Datasets via Batch Statistics Adaptation

Noguchi

Harada

2019

2019 IEEE/CVF International Conference on Computer Vision (ICCV)

157

159

View full text Add to dashboard Cite

Figure 1: Interpolation results generated by BigGAN adapted to only 25 human and anime face images. The left-and rightmost images are the generated images that correspond to training samples, and the other images are generated by linearly changing the latent vector. In spite of the small amount of training data, our method achieves a smooth interpolation. AbstractThanks to the recent development of deep generative models, it is becoming possible to generate high-quality images with both fidelity and diversity. However, the training of such generative models requires a large dataset. To reduce the amount of data required, we propose a new method for transferring prior knowledge of the pre-trained generator, which is trained with a large dataset, to a small dataset in a different domain. Using such prior knowledge, the model can generate images leveraging some common sense that cannot be acquired from a small dataset. In this work, we propose a novel method focusing on the parameters for batch statistics, scale and shift, of the hidden layers in the generator. By training only these parameters in a supervised manner, we achieved stable training of the generator, and our method can generate higher quality images compared to previous methods without collapsing, even when the dataset is small (∼100). Our results show that the diversity of the filters acquired in the pre-trained generator is important for the performance on the target domain. Our method makes it possible to add a new class or domain to a pre-trained generator without disturbing the performance on the original domain. Code is available at github.com/nogu-atsu/small-dataset-image-generation

show abstract

Meta-Transfer Learning for Few-Shot Learning

Cited by 981 publications

References 33 publications

Few-Shot Learning With Embedded Class Models and Shot-Free Meta Training

Few-Shot Learning With Embedded Class Models and Shot-Free Meta Training

Collaborative Intelligent Cross-Camera Video Analytics at Edge

Image Generation From Small Datasets via Batch Statistics Adaptation

Contact Info

Product

Resources

About