Taskonomy: Disentangling Task Transfer Learning

Zamir, Amir; Sax, Alexander F.; Shen, William; Guibas, Leonidas J.; Malik, Jitendra; Savarese, Silvio

doi:10.24963/ijcai.2019/871

Cited by 328 publications

(543 citation statements)

References 1 publication

Supporting

Mentioning

530

Contrasting

Unclassified

Order By: Relevance

“…Studies have shown that using models pretrained on large datasets can effectively improve network performance and reduce training time in the target domain . Therefore, we evaluated the pretrained ResNet50 network on our dataset.…”

Section: Resultsmentioning

confidence: 99%

See 1 more Smart Citation

A deep‐learning‐based approach for adenoid hypertrophy diagnosis

Shen

Liang

et al. 2020

Medical Physics

View full text Add to dashboard Cite

Purpose: Adenoid hypertrophy is a pathological hyperplasia of adenoids and may cause snoring, apnea, and impede breathing during sleep. In clinical practice, radiologists diagnose the severity of adenoid hypertrophy by measuring the ratio of adenoid width (A) to nasopharyngeal width (N) according to the lateral cephalogram, which indicates the locations of four keypoints. The entire diagnostic process is tedious and time-consuming due to the acquisition of A and N. Thus, there is an urgent need to develop computer-aided diagnostic tools for adenoid hypertrophy. Methods: In this paper, we first propose the use of deep learning to solve the problem of adenoid hypertrophy classification. Deep learning driven by big data has developed greatly in the image processing field. However, obtaining a large amount of training data is hard, making the application of deep learning to medical images more difficult. This paper proposes a keypoint localization method to incorporate more prior information to improve the performance of the model under limited data. Furthermore, we design a novel regularized term called VerticalLoss to capture the vertical relationship between keypoints to provide prior information to strengthen the network performance. Results: To evaluate the performance of our proposed method, we conducted experiments with a clinical dataset from the First Affiliated Hospital of Anhui Medical University consisting of a total of 688 patients. As our results show, we obtained a classification accuracy of 95.6%, a macro F1-score of 0.957, and an average AN ratio error of 0.026. Furthermore, we obtained a macro F1-score of 0.89, a classification accuracy of 94%, and an average AN ratio error of 0.027 while using only half of the data for training. Conclusions: The study shows that our proposed method can achieve satisfactory results in the task of adenoid hypertrophy classification. Our approach incorporates more prior information, which is especially important in the field of medical imaging, where it is difficult to obtain large amounts of training data.

show abstract

Section: Resultsmentioning

confidence: 99%

“…On the other hand, the labeling of medical images requires substantial professional knowledge, making it difficult to obtain high‐quality labeled data. Using a pretrained network, the dilemma caused by insufficient data can be solved to some extent . However, the pretrained network has its own limitations.…”

Section: Discussionmentioning

confidence: 99%

A deep‐learning‐based approach for adenoid hypertrophy diagnosis

Shen

Liang

et al. 2020

Medical Physics

View full text Add to dashboard Cite

show abstract

“…8 Because the anterior layers in a deep convolutional neural network only learn some contour features, such as boundaries, shapes, and colors, the anterior layers in same convolutional neural network have almost the same parameters, no matter what visual tasks are applied. 23 The input to our convolutional network is a fixed-size 224 × 224 red-green-blue (RGB) fabric image, and the only preprocessing we do is subtracting the mean RGB value from each pixel during training. The mean RGB value comes from the per-pixel average value of all 6000 fabric images in the train set.…”

Section: Fine-tuning Vgg16mentioning

confidence: 99%

Fabric defect recognition using optimized neural networks

Liu

Zhang

et al. 2019

Journal of Engineered Fibers and Fabrics

View full text Add to dashboard Cite

Fabric defect recognition is an important measure for quality control in a textile factory. This article utilizes a deep convolutional neural network to recognize defects in fabrics that have complicated textures. Although convolutional neural networks are very powerful, a large number of parameters consume considerable computation time and memory bandwidth. In real-world applications, however, the fabric defect recognition task needs to be carried out in a timely fashion on a computation-limited platform. To optimize a deep convolutional neural network, a novel method is introduced to reveal the input pattern that originally caused a specific activation in the network feature maps. Using this visualization technique, this study visualizes the features in a fully trained convolutional model and attempts to change the architecture of original neural network to reduce computational load. After a series of improvements, a new convolutional network is acquired that is more efficient to the fabric image feature extraction, and the computation load and the total number of parameters in the new network is 23% and 8.9%, respectively, of the original model. The proposed neural network is specifically tailored for fabric defect recognition in resource-constrained environments. All of the source code and pretrained models are available online at https://github.com/ZCmeteor .

show abstract

“…Furthermore, these gradient-based methods can often be implemented to efficiently run on specialized hardware such as GPUs. Importantly, it is natural for gradient-based optimization methods to be combined for the joint optimization of multiple objectives or components of a model for end-to-end training [46,52]. In these joint optimization settings, signal from the underlying task can help inform the clustering algorithm and vice-versa.…”

Section: Introductionmentioning

confidence: 99%

Gradient-based Hierarchical Clustering using Continuous Representations of Trees in Hyperbolic Space

Monath

Zaheer

Silva

et al. 2019

Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery &Amp; Data Mining

View full text Add to dashboard Cite

Hierarchical clustering is typically performed using algorithmicbased optimization searching over the discrete space of trees. While these optimization methods are often effective, their discreteness restricts them from many of the benefits of their continuous counterparts, such as scalable stochastic optimization and the joint optimization of multiple objectives or components of a model (e.g. end-to-end training). In this paper, we present an approach for hierarchical clustering that searches over continuous representations of trees in hyperbolic space by running gradient descent. We compactly represent uncertainty over tree structures with vectors in the Poincaré ball. We show how the vectors can be optimized using an objective related to recently proposed cost functions for hierarchical clustering [16, 49]. Using our method with a mini-batch stochastic gradient descent inference procedure, we are able to outperform prior work on clustering millions of ImageNet images by 15 points of dendrogram purity. Further, our continuous tree representation can be jointly optimized in multi-task learning applications offering a 9 point improvement over baseline methods. CCS CONCEPTS • Computing methodologies → Cluster analysis.

show abstract

Taskonomy: Disentangling Task Transfer Learning

Cited by 328 publications

References 1 publication

A deep‐learning‐based approach for adenoid hypertrophy diagnosis

A deep‐learning‐based approach for adenoid hypertrophy diagnosis

Fabric defect recognition using optimized neural networks

Gradient-based Hierarchical Clustering using Continuous Representations of Trees in Hyperbolic Space

Contact Info

Product

Resources

About