Large-scale GPU clusters have been widely used for effectively training both online and offline deep learning (DL) jobs. However, elastic scheduling in most cases of resource schedulers is dedicated for offline model training where resource adjustment is planned ahead of time. The native autoscaling policy is on the basis of pre-defined threshold and, if applied directly in online model training, often suffers from belated resource adjustment, leading to diminished model accuracy. In this paper, we present Kale, a novel elastic GPU scheduling system to improve the performance of online DL model training. Through traffic forecasting and resource-throughput modeling, Kale automatically pinpoints the number of required GPUs that best accommodate the on-the-fly data samples before performing stabilized autoscaling. An advanced data shuffling strategy is further employed for balancing uneven samples among different training workers, thereby improving the runtime efficacy. Experiments show that Kale substantially outperforms the state-of-the-art solutions. Compared with the default HPA autoscaling strategy, Kale reduces the accumulated lag and downtime by 69.2% and 33.1%, respectively, whilst lowering the SLO violation rate from 19.57% to just 2.6%. Kale has been deployed at Kuaishou's production-level GPU clusters and successfully underpins real-time video recommendation and advertisement at scale.

MOTIVATION: Large language models (LLMs) are being adopted at an unprecedented rate, yet still face challenges in knowledge-intensive domains such as biomedicine. Solutions such as pretraining and domain-specific fine-tuning add substantial computational overhead, requiring further domain-expertise. Here, we introduce a token-optimized and robust Knowledge Graph-based Retrieval Augmented Generation (KG-RAG) framework by leveraging a massive biomedical KG (SPOKE) with LLMs such as Llama-2-13b, GPT-3.5-Turbo, and GPT-4, to generate meaningful biomedical text rooted in established knowledge. RESULTS: Compared to the existing RAG technique for Knowledge Graphs, the proposed method utilizes minimal graph schema for context extraction and uses embedding methods for context pruning. This optimization in context extraction results in more than 50% reduction in token consumption without compromising the accuracy, making a cost-effective and robust RAG implementation on proprietary LLMs. KG-RAG consistently enhanced the performance of LLMs across diverse biomedical prompts by generating responses rooted in established knowledge, accompanied by accurate provenance and statistical evidence (if available) to substantiate the claims. Further benchmarking on human curated datasets, such as biomedical true/false and multiple-choice questions (MCQ), showed a remarkable 71% boost in the performance of the Llama-2 model on the challenging MCQ dataset, demonstrating the framework's capacity to empower open-source models with fewer parameters for domain-specific questions. Furthermore, KG-RAG enhanced the performance of proprietary GPT models, such as GPT-3.5 and GPT-4. In summary, the proposed framework combines explicit and implicit knowledge of KG and LLM in a token optimized fashion, thus enhancing the adaptability of general-purpose LLMs to tackle domain-specific questions in a cost-effective fashion. AVAILABILITY AND IMPLEMENTATION: SPOKE KG can be accessed at https://spoke.rbvi.ucsf.edu/neighborhood.html. It can also be accessed using REST-API (https://spoke.rbvi.ucsf.edu/swagger/). KG-RAG code is made available at https://github.com/BaranziniLab/KG_RAG. Biomedical benchmark datasets used in this study are made available to the research community in the same GitHub repository.

MOTIVATION: Knowledge graphs (KGs) are being adopted in industry, commerce and academia. Biomedical KG presents a challenge due to the complexity, size and heterogeneity of the underlying information. RESULTS: In this work, we present the Scalable Precision Medicine Open Knowledge Engine (SPOKE), a biomedical KG connecting millions of concepts via semantically meaningful relationships. SPOKE contains 27 million nodes of 21 different types and 53 million edges of 55 types downloaded from 41 databases. The graph is built on the framework of 11 ontologies that maintain its structure, enable mappings and facilitate navigation. SPOKE is built weekly by python scripts which download each resource, check for integrity and completeness, and then create a 'parent table' of nodes and edges. Graph queries are translated by a REST API and users can submit searches directly via an API or a graphical user interface. Conclusions/Significance: SPOKE enables the integration of seemingly disparate information to support precision medicine efforts. AVAILABILITY AND IMPLEMENTATION: The SPOKE neighborhood explorer is available at https://spoke.rbvi.ucsf.edu. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Intratumor cellular heterogeneity and non-genetic cell plasticity in tumors pose a recently recognized challenge to cancer treatment. Because of the dispersion of initial cell states within a clonal tumor cell population, a perturbation imparted by a cytocidal drug only kills a fraction of cells. Due to dynamic instability of cellular states the cells not killed are pushed by the treatment into a variety of functional states, including a "stem-like state" that confers resistance to treatment and regenerative capacity. This immanent stress-induced stemness competes against cell death in response to the same perturbation and may explain the near-inevitable recurrence after any treatment. This double-edged-sword mechanism of treatment complements the selection of preexisting resistant cells in explaining post-treatment progression. Unlike selection, the induction of a resistant state has not been systematically analyzed as an immanent cause of relapse. Here, we present a generic elementary model and analytical examination of this intrinsic limitation to therapy. We show how the relative proclivity towards cell death versus transition into a stem-like state, as a function of drug dose, establishes either a window of opportunity for containing tumors or the inevitability of progression following therapy. The model considers measurable cell behaviors independent of specific molecular pathways and provides a new theoretical framework for optimizing therapy dosing and scheduling as cancer treatment paradigms move from "maximal tolerated dose," which may promote therapy induced-stemness, to repeated "minimally effective doses" (as in adaptive therapies), which contain the tumor and avoid therapy-induced progression.

In a two-player deep reinforcement learning task, recent work shows an attacker could learn an adversarial policy that triggers a target agent to perform poorly and even react in an undesired way. However, its efficacy heavily relies upon the zero-sum assumption made in the two-player game. In this work, we propose a new adversarial learning algorithm. It addresses the problem by resetting the optimization goal in the learning process and designing a new surrogate optimization function. Our experiments show that our method significantly improves adversarial agents’ exploitability compared with the state-of-art attack. Besides, we also discover that our method could augment an agent with the ability to abuse the target game’s unfairness. Finally, we show that agents adversarially re-trained against our adversarial agents could obtain stronger adversary-resistance.

In Taobao, the largest e-commerce platform in China, billions of items are provided and typically displayed with their images.For better user experience and business effectiveness, Click Through Rate (CTR) prediction in online advertising system exploits abundant user historical behaviors to identify whether a user is interested in a candidate ad. Enhancing behavior representations with user behavior images will help understand user's visual preference and improve the accuracy of CTR prediction greatly. So we propose to model user preference jointly with user behavior ID features and behavior images. However, training with user behavior images brings tens to hundreds of images in one sample, giving rise to a great challenge in both communication and computation. To handle these challenges, we propose a novel and efficient distributed machine learning paradigm called Advanced Model Server (AMS). With the well-known Parameter Server (PS) framework, each server node handles a separate part of parameters and updates them independently. AMS goes beyond this and is designed to be capable of learning a unified image descriptor model shared by all server nodes which embeds large images into low dimensional high level features before transmitting images to worker nodes. AMS thus dramatically reduces the communication load and enables the arduous joint training process. Based on AMS, the methods of effectively combining the images and ID features are carefully studied, and then we propose a Deep Image CTR Model. Our approach is shown to achieve significant improvements in both online and offline evaluations, and has been deployed in Taobao display advertising system serving the main traffic.

It has been recently shown that deep neural networks (DNNs) are susceptible to a particular type of attack that exploits a fundamental flaw in their design. This attack consists of generating particular synthetic examples referred to as adversarial samples. These samples are constructed by slightly manipulating real data-points that change "fool" the original DNN model, forcing it to misclassify previously correctly classified samples with high confidence. Many believe addressing this flaw is essential for DNNs to be used in critical applications such as cyber security. Previous work has shown that learning algorithms that enhance the robustness of DNN models all use the tactic of "security through obscurity". This means that security can be guaranteed only if one can obscure the learning algorithms from adversaries. Once the learning technique is disclosed, DNNs protected by these defense mechanisms are still susceptible to adversarial samples. In this work, we investigate by examining how previous research dealt with this and propose a generic approach to enhance a DNN's resistance to adversarial samples. More specifically, our approach integrates a data transformation module with a DNN, making it robust even if we reveal the underlying learning algorithm. To demonstrate the generality of our proposed approach and its potential for handling cyber security applications, we evaluate our method and several other existing solutions on datasets publicly available, such as a large scale malware dataset and MNIST and IMDB datasets. Our results indicate that our approach typically provides superior classification performance and robustness to attacks compared with state-of-art solutions.

Understanding and interpreting how machine learning (ML) models make decisions have been a big challenge. While recent research has proposed various technical approaches to provide some clues as to how an ML model makes individual predictions, they cannot provide users with an ability to inspect a model as a complete entity. In this work, we propose a novel technical approach that augments a Bayesian non-parametric regression mixture model with multiple elastic nets. Using the enhanced mixture model, we can extract generalizable insights for a target model through a global approximation. To demonstrate the utility of our approach, we evaluate it on different ML models in the context of image recognition. The empirical results indicate that our proposed approach not only outperforms the state-of-the-art techniques in explaining individual decisions but also provides users with an ability to discover the vulnerabilities of the target ML models.

Online video chat services such as Chatroulette, Omegle, and vChatter that randomly match pairs of users in video chat sessions are quickly becoming very popular, with over a million users per month in the case of Chatroulette. A key problem encountered in such systems is the presence of flashers and obscene content. This problem is especially acute given the presence of underage minors in such systems. This article presents SafeVchat, a novel solution to the problem of flasher detection that employs an array of image detection algorithms. A key contribution of the article concerns how the results of the individual detectors are fused together into an overall decision classifying a user as misbehaving or not, based on Dempster-Shafer theory. The article introduces a novel, motion-based skin detection method that achieves significantly higher recall and better precision. The proposed methods have been evaluated over real-world data and image traces obtained from Chatroulette.com. SafeVchat has been deployed in Chatroulette. A combination of SafeVchat with human moderation has resulted in banning as many as 50,000 inappropriate users per day on Chatoulette. Furthermore, offensive content on Chatoulette has dropped significantly from 33.08% (before SafeVchat installation) to 3.49% (after SafeVchat installation).

The need for highly scalable and accurate detection and filtering of misbehaving users and obscene content in online video chat services has grown as the popularity of these services has exploded in popularity. This is a challenging problem because processing large amounts of video is compute intensive, decisions about whether a user is misbehaving or not must be made online and quickly, and moreover these video chats are characterized by low quality video, poorly lit scenes, diversity of users and their behaviors, diversity of the content, and typically short sessions. This paper presents EMeralD, a highly scalable system for accurately detecting and filtering misbehaving users in online video chat applications. EMeralD substantially improves upon the state-of-the-art filtering mechanisms by achieving much lower computational cost and higher accuracy. We demonstrate EMeralD's improvement via experimental evaluations on real-world data sets obtained from Chatroulette.com.

BACKGROUND: Gradual or sudden transitions among different states as exhibited by cell populations in a biological sample under particular conditions or stimuli can be detected and profiled by flow cytometric time course data. Often such temporal profiles contain features due to transient states that present unique modeling challenges. These could range from asymmetric non-Gaussian distributions to outliers and tail subpopulations, which need to be modeled with precision and rigor. RESULTS: To ensure precision and rigor, we propose a parametric modeling framework StateProfiler based on finite mixtures of skew t-Normal distributions that are robust against non-Gaussian features caused by asymmetry and outliers in data. Further, we present in StateProfiler a new greedy EM algorithm for fast and optimal model selection. The parsimonious approach of our greedy algorithm allows us to detect the genuine dynamic variation in the key features as and when they appear in time course data. We also present a procedure to construct a well-fitted profile by merging any redundant model components in a way that minimizes change in entropy of the resulting model. This allows precise profiling of unusually shaped distributions and less well-separated features that may appear due to cellular heterogeneity even within clonal populations. CONCLUSIONS: By modeling flow cytometric data measured over time course and marker space with StateProfiler, specific parametric characteristics of cellular states can be identified. The parameters are then tested statistically for learning global and local patterns of spatio-temporal change. We applied StateProfiler to identify the temporal features of yeast cell cycle progression based on knockout of S-phase triggering cyclins Clb5 and Clb6, and then compared the S-phase delay phenotypes due to differential regulation of the two cyclins. We also used StateProfiler to construct the temporal profile of clonal divergence underlying lineage selection in mammalian hematopoietic progenitor cells.

Accumulating experimental evidence suggests that the gene regulatory networks of living organisms operate in the critical phase, namely, at the transition between ordered and chaotic dynamics. Such critical dynamics of the network permits the coexistence of robustness and flexibility which are necessary to ensure homeostatic stability (of a given phenotype) while allowing for switching between multiple phenotypes (network states) as occurs in development and in response to environmental change. However, the mechanisms through which genetic networks evolve such critical behavior have remained elusive. Here we present an evolutionary model in which criticality naturally emerges from the need to balance between the two essential components of evolvability: phenotype conservation and phenotype innovation under mutations. We simulated the Darwinian evolution of random Boolean networks that mutate gene regulatory interactions and grow by gene duplication. The mutating networks were subjected to selection for networks that both (i) preserve all the already acquired phenotypes (dynamical attractor states) and (ii) generate new ones. Our results show that this interplay between extending the phenotypic landscape (innovation) while conserving the existing phenotypes (conservation) suffices to cause the evolution of all the networks in a population towards criticality. Furthermore, the networks produced by this evolutionary process exhibit structures with hubs (global regulators) similar to the observed topology of real gene regulatory networks. Thus, dynamical criticality and certain elementary topological properties of gene regulatory networks can emerge as a byproduct of the evolvability of the phenotypic landscape.

Online video chat services such as Chatroulette, Omegle, and vChatter that randomly match pairs of users in video chat sessions are fast becoming very popular, with over a million users per month in the case of Chatroulette. A key problem encountered in such systems is the presence of flashers and obscene content. This problem is especially acute given the presence of underage minors in such systems. This paper presents SafeVchat, a novel solution to the problem of flasher detection that employs an array of image detection algorithms. A key contribution of the paper concerns how the results of the individual detectors are fused together into an overall decision classifying the user as misbehaving or not, based on Dempster-Shafer Theory. The paper introduces a novel, motion-based skin detection method that achieves significantly higher recall and better precision. The proposed methods have been evaluated over real-world data and image traces obtained from Chatroulette.com.

Grouping genes by similarity of expression across multiple cellular conditions enables the identification of cellular modules. The known functions of genes enable the characterization of the aggregate biological functions of these modules. In this paper, we use a high-throughput approach to identify the effective mutual regulatory interactions between modules composed of mouse genes from the Alliance for Cell Signaling (AfCS) murine B-lymphocyte database which tracks the response of approximately 15,000 genes following chemokine perturbation. This analysis reveals principles of cellular organization that we discuss along four conceptual axes. (1) Regulatory implications: the derived collection of influences between any two modules quantifies intuitive as well as unexpected regulatory interactions. (2) Behavior across scales: trends across global networks of varying resolution (composed of various numbers of modules) reveal principles of assembly of high-level behaviors from smaller components. (3) Temporal behavior: tracking the mutual module influences over different time intervals provides features of regulation dynamics such as duration, persistence, and periodicity. (4) Gene Ontology correspondence: the association of modules to known biological roles of individual genes describes the organization of functions within coexpressed modules of various sizes. We present key specific results in each of these four areas, as well as derive general principles of cellular organization. At the coarsest scale, the entire transcriptional network contains five divisions: two divisions devoted to ATP production/biosynthesis and DNA replication that activate all other divisions, an "extracellular interaction" division that represses all other divisions, and two divisions (proliferation/differentiation and membrane infrastructure) that activate and repress other divisions in specific ways consistent with cell cycle control.

UNLABELLED: Genome-wide expression profiles contain global patterns that evade visual detection in current gene clustering analysis. Here, a Gene Expression Dynamics Inspector (GEDI) is described that uses self-organizing maps to translate high-dimensional expression profiles of time courses or sample classes into animated, coherent and robust mosaics images. GEDI facilitates identification of interesting patterns of molecular activity simultaneously across gene, time and sample space without prior assumption of any structure in the data, and then permits the user to retrieve genes of interest. Important changes in genome-wide activities may be quickly identified based on 'Gestalt' recognition and hence, GEDI may be especially useful for non-specialist end users, such as physicians. AVAILABILITY: GEDI v1.0 is written in Matlab, and binary Matlab.dll files which require Matlab to run can be downloaded for free by academic institutions at http://www.chip.org/~ge/gedihome.html SUPPLEMENTARY INFORMATION: http://www.chip.org/~ge/gedihome.html