View profile

AI Scholar Weekly - Issue #56

AI Scholar Weekly - Issue #56
By Educate AI  • Issue #56 • View online
40 Free Resources to Help You Learn Machine Learning on Your Own; Free Book on Neural Network and Deep Learning; Video Question Answering Datasets and Methods; Leveraging ML for game Development; A New Class of ResNet Architectures (and more)

Top AI Research Papers This Week!
#1 How I Failed Machine Learning in Medical Imaging
Medical imaging is an important research field with many opportunities for improving patients’ health. However, there are a number of challenges that are slowing down progress in the field.
In this paper, scholars reviewed several problems related to choosing datasets, methods, evaluation metrics, and publication strategies. Their work provides a broad overview of problems that may be slowing down medical imaging, as well related computational fields in general, based on both literature and their own analysis.
For instance, they show that dataset size is not everything, and while datasets are slowly getting larger predictive performance is not. Secondly, their analysis shows that the availability of datasets might influence what medical imaging choose to work on, possibly moving attention away from other unsolved problems. Thirdly, their analysis shows that outperforming a state-of-the-art method may not always be meaningful, and thus may create an illusion of progress.
The work also provides a broad range of strategies to address this situation. They hope that the suggestions will be useful to practitioners and medical imaging and related fields.
For reproducibility, data and code for our analyses are available on Github. Read more: How I Failed Machine Learning in Medical Imaging
#2 Stable ResNets — A New Class of ResNet Architectures
Deep ResNet architectures have attained state-of-the-art performance on numerous tasks.
Stable ResNets have the benefit of stabilizing the gradient and ensuring expressivity in the limit of infinite depth. However, they might suffer from gradient exploding as the depth becomes large. Moreover, recent results have shown that ResNet might lose expressivity as the depth goes to infinity according to previous research.
This recent paper seeks to address those challenges by introducing a new class of ResNet architectures, called Stable ResNet, that have the property of stabilizing the gradient while ensuring expressivity in the infinite depth limit. Researchers demonstrate that this type of scaling makes NNGP inference robust and improves test accuracy with SGD on modern ResNet architectures.
Nevertheless, while Stable ResNets outperform standard ResNet, the selection of optimal scaling remains an open question, which remains an open topic for future work. Read the full paper: Introducing Stable ResNet
#3 Video Prediction for Robot-assisted Surgery
This paper presents a new method based on VAE for frame prediction in robotic surgical video sequences. According to the researchers, this is the first time that the future frames of dual-arm robots are predicted considering their unique characteristics relative to general robotic videos.
The suggested model employs learned and intrinsic prior information as guidance to help generate future scene conditioning on the observed frames. The stochastic VAE based method is adapted as a deterministic approach by directly using the expectation of the distribution without sampling.
Results: this proposed method outperforms the baseline methods on the challenging dual-arm robotic surgical video dataset. Future work can be made to explore higher resolution generation and apply the predicted future frames to other advanced tasks. Read more: Future Frame Prediction for Robot-assisted Surgery
#4 A New Efficient and Lightweight Deep Classification Ensemble Structure
Researchers with Comcast AI Labs, USA propose an efficient and lightweight neural network ensemble structure, that’s specially designed for “high-accuracy” detections in images and videos with low false positives.
They designed, implemented, and evaluated a new approach for an explosion detection use case. However, the approach is general and can be applied to other similar object classification use cases as well.
Evaluations based on testing models on a large test set show significant accuracy improvement over the popular ResNet-50 model, while benefiting from the simplicity, lower computation time, and faster inference time.
Given the insights gained from the experiments, they propose a “think small, think many” philosophy when performing deep object classification tasks. Their argument is that transforming a single, large, monolithic deep model into a verification-based hierarchy of multiple small and lightweight models with contracting color channels can potentially lead to predictions with higher accuracy.
#5 Adaptive Pixel-wise Structured Sparse Network for Efficient CNNs
To accelerate deep CNN models, a group of researchers with Tsinghua University researchers have proposed a new spatially adaptive framework that can dynamically generate pixel-wise sparsity according to the input image.
The sparsity is guided by a unified importance map. An example of hardware implementation is also presented. The proposed method can control the sparsity level in one model, thus saving the burden to train multiple models for various devices or requirements. The method can also be widely utilized on various vision tasks.
Experimental results: show the proposed models achieve comparable or better performance with state-of-art methods.
Specifically, results show that this method efficiently improves the computing efficiency for both image classification using ResNet-18 and super-resolution using SRResNet. On image classification tasks, the new method can save 30%-70% MACs with a slight drop in top-1 and top-5 accuracy. On super-resolution task, the method can reduce more than 90% MACs while only causing around 0.1 dB and 0.01 decreasing in PSNR and SSIM. Hardware validation is also included. Read more: Adaptive Pixel-wise Structured Sparse Network for Efficient CNNs
Other Great AI Papers
This new research shows that the NLU tasks can be formulated as conditional generation tasks, and therefore solvable by autoregressive models. A Novel Pretraining Framework GLM (General Language Model)
Researchers show that the words in the NLU training set can be modeled as a long-tailed distribution. They propose a LTGR framework to reduce the model’s reliance on shortcut features, by suppressing the model from outputting overconfident prediction for samples with large shortcut degree. Experimental results validate the proposed method and improve generalization on OOD samples, without sacrificing the accuracy of in-distribution samples. Read more
Recent advances in Video Question Answering: A Review of Datasets and Methods
Google AI found that a relatively simple neural network was sufficient to reach high-level performance against humans and traditional game AI. Leveraging Machine Learning for Game Development
An interpretable hierarchical integrated decision and control framework for automated vehicles, which equips with high online computing efficiency and great adaptability to different tasks. The results show the potential of the method to be applied in real-world autonomous driving tasks. Read more
AI Resources
Introduction to Artificial Intelligence. Free AI Course on the Fundamentals of AI
How Not to Fail Your Machine Learning Interview. Go here
40 Free Resources to Help You Learn Machine Learning on Your Own. Go here
How to Build a Successful Machine Learning Portfolio. Read more
Free Book on Neural Network and Deep Learning. Get Access
Job Opportunities
Machine Learning Scientist Job at Apple
Top AI News
Researchers’ algorithm designs soft robots that sense. Read more
Alluxio Advances Analytics and AI with NVIDIA Accelerated Computing. Read more
About AI Scholar Weekly
Thanks for reading. Create a ripple effect by sharing this AI Scholar Newsletter with someone else, and they will also be lit up!
Did you enjoy this issue?
Educate AI

AI Scholar Weekly brings you everything new and exciting in the world of Artificial Intelligence, Machine Learning, and Deep Learning every week for free.

In order to unsubscribe, click here.
If you were forwarded this newsletter and you like it, you can subscribe here.
Powered by Revue