Predicting Application Resilience Using Machine Learning (PARIS) 
Excessive-scale scientific functions may be extra susceptible to gentle errors (transient faults) as high-performance computing methods enhance in scale. The widespread observe to guage the resilience to faults of an software is random fault injection, a way that may be extremely time consuming. Whereas resilience prediction modeling has been lately proposed to foretell software resilience in a quicker means than fault injection, it could possibly solely predict a single class of fault manifestation (SDC) and there’s no proof demonstrating that it could possibly work on beforehand unseen packages, which drastically limits its re-usability. We current PARIS, a resilience prediction technique that addresses the issues of current prediction strategies utilizing machine studying. Utilizing carefully-selected options and a machine studying mannequin, our technique is ready to make resilience predictions of three courses of fault manifestations (success, SDC, and interruption) versus one class like in present resilience prediction modeling. The generality of our method permits us to make prediction on new functions, i.e., beforehand unseen functions, offering massive applicability to our mannequin. Our analysis on 125 packages exhibits that PARIS offers excessive prediction accuracy, 82% and 77% on common for predicting the speed of success and interruption, respectively, whereas the state-of-the-art resilience prediction mannequin can’t predict them. When predicting the speed of SDC, PARIS offers significantly better accuracy than the state-of-the-art (38% vs. -273%). PARIS is way quicker (as much as 450x speedup) than the standard technique (random fault injection). …
Semi-Supervised Semantic Matching
Convolutional neural networks (CNNs) have been efficiently utilized to unravel the issue of correspondence estimation between semantically associated pictures. Because of non-availability of enormous coaching datasets, current strategies resort to self-supervised or unsupervised coaching paradigm. On this paper we suggest a semi-supervised studying framework that imposes cyclic consistency constraint on unlabeled picture pairs. Along with the supervised loss the proposed mannequin achieves state-of-the-art on a benchmark semantic matching dataset. …
Avro2TF
Deep studying has been efficiently utilized to a number of AI methods at LinkedIn which are associated to advice and search. One of many essential classes that we’ve realized throughout this journey is to supply good deep studying platforms that assist our modeling engineers turn out to be extra environment friendly and productive. Avro2TF is a part of this effort to scale back the complexity of information processing and enhancing velocity of superior modeling. Along with superior deep studying methods, LinkedIn has been on the forefront of Machine Studying innovation for years now. We’ve got many various ML approaches that devour great amount of information on a regular basis. Effectivity and accuracy are a very powerful measurements for these approaches. To successfully help deep studying at LinkedIn, we have to first deal with the info processing points. Many of the datasets utilized by our ML algorithms (e.g., LinkedIn’s massive scale personalization engine Photon-ML) are in Avro format. Every report in a Avro dataset is basically a sparse vector, and may be simply consumed by many of the fashionable classifiers. Nevertheless, the format can’t be instantly utilized by TensorFlow — the main deep studying package deal. The principle blocker is that the sparse vector will not be in the identical format as Tensor. We imagine that this isn’t solely a LinkedIn downside, many corporations have huge quantity of ML knowledge in related sparse vector format, and Tensor format remains to be comparatively new to many corporations. Avro2TF bridges this hole by offering scalable Spark primarily based transformation and extensions mechanism to effectively convert the info into TF data that may be readily consumed by TensorFlow. With this expertise, builders can enhance their productiveness by specializing in mannequin constructing somewhat than knowledge conversion. …
Tiny SSD
Object detection is a significant problem in laptop imaginative and prescient, involving each object classification and object localization inside a scene. Whereas deep neural networks have been proven lately to yield very highly effective methods for tackling the problem of object detection, one of many largest challenges with enabling such object detection networks for widespread deployment on embedded units is excessive computational and reminiscence necessities. Not too long ago, there was an rising focus in exploring small deep neural community architectures for object detection which are extra appropriate for embedded units, comparable to Tiny YOLO and SqueezeDet. Impressed by the effectivity of the Fireplace microarchitecture launched in SqueezeNet and the article detection efficiency of the single-shot detection macroarchitecture launched in SSD, this paper introduces Tiny SSD, a single-shot detection deep convolutional neural community for real-time embedded object detection that’s composed of a extremely optimized, non-uniform Fireplace sub-network stack and a non-uniform sub-network stack of extremely optimized SSD-based auxiliary convolutional function layers designed particularly to reduce mannequin measurement whereas sustaining object detection efficiency. The ensuing Tiny SSD possess a mannequin measurement of two.3MB (~26X smaller than Tiny YOLO) whereas nonetheless reaching an mAP of 61.3% on VOC 2007 (~4.2% increased than Tiny YOLO). These experimental outcomes present that very small deep neural community architectures may be designed for real-time object detection which are well-suited for embedded situations. …
