Set Of New Classes Research Articles

Prompt tuning for parameter-efficient medical image segmentation

Neural networks pre-trained on a self-supervision scheme have become the standard when operating in data rich environments with scarce annotations. As such, fine-tuning a model to a downstream task in a parameter-efficient but effective way, e.g. for a new set of classes in the case of semantic segmentation, is of increasing importance. In this work, we propose and investigate several contributions to achieve a parameter-efficient but effective adaptation for semantic segmentation on two medical imaging datasets. Relying on the recently popularized prompt tuning approach, we provide a prompt-able UNETR (PUNETR) architecture, that is frozen after pre-training, but adaptable throughout the network by class-dependent learnable prompt tokens. We pre-train this architecture with a dedicated dense self-supervision scheme based on assignments to online generated prototypes (contrastive prototype assignment, CPA) of a student teacher combination. Concurrently, an additional segmentation loss is applied for a subset of classes during pre-training, further increasing the effectiveness of leveraged prompts in the fine-tuning phase. We demonstrate that the resulting method is able to attenuate the gap between fully fine-tuned and parameter-efficiently adapted models on CT imaging datasets. To this end, the difference between fully fine-tuned and prompt-tuned variants amounts to 7.81 pp for the TCIA/BTCV dataset as well as 5.37 and 6.57 pp for subsets of the TotalSegmentator dataset in the mean Dice Similarity Coefficient (DSC, in %) while only adjusting prompt tokens, corresponding to 0.51% of the pre-trained backbone model with 24.4M frozen parameters. The code for this work is available on https://github.com/marcdcfischer/PUNETR.

On new difference sequence spaces via Cesàro mean

In the present article, we define certain a set of new classes of sequence spaces by using Ces\`{a}ro mean and difference operator $\Delta^r\;\;r\in\mathbb{N}_0=\{0,1,2,3,\dots\}$. Also, we study the topological structures of the defined classes and determine their $\alpha-,\beta-$ and $\gamma-$ duals. Matrix transformations of given classes with their basic sequence spaces are characterized.

Reorientation Rates of Structural and Kinematic Axes in Simulated Massive Galaxies and the Origins of Prolate Rotation

In this work, we analyze a sample of ∼4000 massive (M * ≥ 1011 M ⊙ at z = 0) galaxies in TNG300, the (300Mpc)3 box of the IllustrisTNG simulation suite. We characterize the shape and kinematics of these galaxies with a focus on the kinematic misalignment (Ψint) between the angular momentum (AM) and morphological major axis. We find that the traditional purely shape- or kinematics-based classifications are insufficient to characterize the diversity of our sample and define a new set of classes based on the rates of change of the galaxies’ morphological and kinematic axes. We show that these classes are mostly stable over time and correspond to six distinct populations of galaxies: the rapid AM reorienters (58% of our sample), unsettled galaxies (10%), spinning disks (10%), twirling cigars (16%), misaligned slow reorienters (3%), and regular prolate rotators (galaxies that display major axis rotation; 2%). We demonstrate that the most-recent significant (mass-ratio μ > 1/10) mergers of these galaxies are the primary cause for their present-day properties and find that these mergers are best characterized at the point of the satellite’s final infall—that is, much closer to the final coalescence than has been previously thought. We show that regular prolate rotators evolve from spinning disk progenitors that experience a radial merger along their internal AM direction. Finally, we argue that these regular prolate rotators are distinct from the similarly sized population of rapid AM reorienters with large Ψint, implying that a large Ψint is not a sufficient condition for major axis rotation.

Few-Shot Lifelong Learning

Many real-world classification problems often have classes with very few labeled training samples. Moreover, all possible classes may not be initially available for training, and may be given incrementally. Deep learning models need to deal with this two-fold problem in order to perform well in real-life situations. In this paper, we propose a novel Few-Shot Lifelong Learning (FSLL) method that enables deep learning models to perform lifelong/continual learning on few-shot data. Our method selects very few parameters from the model for training every new set of classes instead of training the full model. This helps in preventing overfitting. We choose the few parameters from the model in such a way that only the currently unimportant parameters get selected. By keeping the important parameters in the model intact, our approach minimizes catastrophic forgetting. Furthermore, we minimize the cosine similarity between the new and the old class prototypes in order to maximize their separation, thereby improving the classification performance. We also show that integrating our method with self-supervision improves the model performance significantly. We experimentally show that our method significantly outperforms existing methods on the miniImageNet, CIFAR-100, and CUB-200 datasets. Specifically, we outperform the state-of-the-art method by an absolute margin of 19.27% for the CUB dataset.

Incremental Learning in Deep Convolutional Neural Networks Using Partial Network Sharing

Deep convolutional neural network (DCNN) based supervised learning is a widely practiced approach for large-scale image classification. However, retraining these large networks to accommodate new, previously unseen data demands high computational time and energy requirements. Also, previously seen training samples may not be available at the time of retraining. We propose an efficient training methodology and incrementally growing DCNN to learn new tasks while sharing part of the base network. Our proposed methodology is inspired by transfer learning techniques, although it does not forget previously learned tasks. An updated network for learning new set of classes is formed using previously learned convolutional layers (shared from initial part of base network) with addition of few newly added convolutional kernels included in the later layers of the network. We employed a `clone-and-branch' technique which allows the network to learn new tasks one after another without any performance loss in old tasks. We evaluated the proposed scheme on several recognition applications. The classification accuracy achieved by our approach is comparable to the regular incremental learning approach (where networks are updated with new training samples only, without any network sharing), while achieving energy efficiency, reduction in storage requirements, memory access and training time.

Data science can be an attractive career for physicists

Boot camps can help researchers transition into nonacademic jobs that use big data.

Enablement of Consumer-Oriented Interoperable Systems With Integration of CIM and Green Button Standards

Future grid evolutions promise to promote the interoperability in the consumer domain. CIM (mainly IEC61970 and IEC61968) is an established standard in the industry to enable the interoperable data exchanges at the upstream levels like generation, transmission and distribution, while the Green Button standards (OpenADE, NAESB ESPI, NIST SGIP-PAP 20, PAP10) are the emerging smart grid standards to standardize the information on energy usage, interfaces at the downstream end-consumer level. This paper presents the various modeling issues, and techniques to move toward integrating these standards to enable the interoperable services at the consumer levels. The challenging task is to identify the options to promote the interoperability between the end-consumer and upstream levels for future grid evolutions. This paper begins to address this issue by identifying the requirements like integrating the external services data from the end-consumer oriented third party tools to EMS/DMS, hierarchical tagging of consumers, combined semantics for third party applications, inter-utility migration of consumers/prosumers, and dynamic consumer/prosumer open access interactions. However, this involves extensions and mappings between the existing CIM, and the green button standards. Without affecting much the base packages, it should be either incorporated in the CIM or Green Button standards at the domain modeling as a separate package. A set of new classes and attributes to the existing CIM and the corresponding XML tags for Green Button standards to realize the integration is proposed. The authors' contributions also strengthen the arguments for creating a separate green button profile for the existing upstream standards.

LUXSim: A component-centric approach to low-background simulations

Geant4 has been used throughout the nuclear and high-energy physics community to simulate energy depositions in various detectors and materials. These simulations have mostly been run with a source beam outside the detector. In the case of low-background physics, however, a primary concern is the effect on the detector from radioactivity inherent in the detector parts themselves. From this standpoint, there is no single source or beam, but rather a collection of sources with potentially complicated spatial extent. LUXSim is a simulation framework used by the LUX collaboration that takes a component-centric approach to event generation and recording. A new set of classes allows for multiple radioactive sources to be set within any number of components at run time, with the entire collection of sources handled within a single simulation run. Various levels of information can also be recorded from the individual components, with these record levels also being set at run time. This flexibility in both source generation and information recording is possible without the need to recompile, reducing the complexity of code management and the proliferation of versions. Within the code itself, casting geometry objects within this new set of classes rather than as the default Geant4 classes automatically extends this flexibility to every individual component. No additional work is required on the part of the developer, reducing development time and increasing confidence in the results. We describe the guiding principles behind LUXSim, detail some of its unique classes and methods, and give examples of usage.

Reader/Writer for Analyze Object Maps for ITK

This document describes an addition to the Image IO Library of the Insight Toolkit (ITK). ITK has been able to read in Analyze image files but not the Analyze Object Maps that correspond to the images. Analyze Object maps define regions of interest that are helpful in designating locations, giving a name to those locations, as well as display characteristics. Also, Analyze Object Maps are useful to tell someone where to look in an image. With the lack of not being able to read/write Analyze Object Maps, a new set of classes have been developed to read/write Analyze Object Maps and manipulate the object maps.

Label object representation and manipulation with ITK

Richard Beare has recently introduced a new filter to efficiently labelize the connected components with ITK, and has also proposed to use the run-length encoding used in that filter to implement some of the most useful binary mathematical morphology operators: the opening by attribute. Following that idea, and after have searched a way to use the ITK’s spatial objects for this task, a new set of classes have been developed to represent and manipulate the label images and the objects within them in ITK. Those new classes have been used to implement several label images manipulation based on object attributes, as well as the binary and label specialization of some mathematical morphology filters based on the morphological reconstruction. This contribution comes with 65 new classes, and should greatly enhance the binary mathematical morphology in ITK.All the source codes are provided, as well as a full set of tests and several usage examples of the new classes.

Achieving knowledge interoperability: An XML/XSLT approach

Development of an intelligent system requires not only profound understanding of the problem under study, but also employment of different knowledge representation techniques and tools often based on a variety of paradigms and technological platforms. In this context automation of knowledge sharing between different systems becomes increasingly important. One solution might be to extend a knowledge modeling tool by implementing a set of new classes or functions for importing other knowledge formats (using, e.g. Java, C++, etc.). But, this can be a rather difficult and time consuming task. Since XML is now widely accepted as knowledge representation syntax, we believe that a more suitable solution would be to use eXtensible Stylesheet Language Transformation (XSLT) a W3C standard for transforming XML documents. A special advantage of this approach is that even though an XSLT is written independently of any programming language, it can be executed by a program written in almost any up-to-date programming language. We experiment on an XSLT-based infrastructure for sharing knowledge between three knowledge modeling and acquisition tools that use different conceptual models for knowledge representation in order to evaluate cons and pros of the proposed XSLT approach. Two of these tools, JessGUI and JavaDON are ongoing efforts of the GOOD OLD AI research group to develop interoperable development tools for building intelligent systems, while the third one is [email protected][email protected]?-2000, a broadly accepted ontology development tool.