Milos Nikolic

Followers

Following

Public Views

Prof. dr Vladimir M . Cvetković

University of Belgrade

Magistar Struke Bezbjednosti I Kriminalistike

Jasna Vukovic

University of Belgrade

Dragan Stanic

Vidan Dankovic

University of Belgrade

Katarina Lisavac

Interests

Uploads

Papers by Milos Nikolic

Извод из Евиденција и нотификација о повредама на раду и професионалним обољењима -Међународна организација рада--МОР Кодекси праксе

Download

BitPruning: Learning Bitlengths for Aggressive and Accurate Quantization

Cornell University - arXiv, Feb 7, 2020

Neural networks have demonstrably achieved stateof-the art accuracy using low-bitlength integer q... more Neural networks have demonstrably achieved stateof-the art accuracy using low-bitlength integer quantization, yielding both execution time and energy benefits on existing hardware designs that support short bitlengths. However, the question of finding the minimum bitlength for a desired accuracy remains open. We introduce a training method for minimizing inference bitlength at any granularity while maintaining accuracy. Namely, we propose a regularizer that penalizes large bitlength representations throughout the architecture and show how it can be modified to minimize other quantifiable criteria, such as number of operations or memory footprint. We demonstrate that our method learns thrifty representations while maintaining accuracy. With ImageNet, the method produces an average per layer bitlength of 4.13, 3.76 and 4.36 bits on AlexNet, ResNet18 and MobileNet V2 respectively, remaining within 2.0%, 0.5% and 0.5% of the base TOP-1 accuracy.

Download

Laboratory exercises for embedded engineering learning platform

ABSTRACT This paper presents the set of laboratory exercises developed for the embedded engineeri... more ABSTRACT This paper presents the set of laboratory exercises developed for the embedded engineering learning platform. The platform used is based on FPGA device which is programmed by students. The laboratory exercises cover the fundamental topics in digital design: (1) design of combinational circuits, (2) design of sequential circuits, (3) design of finite state machines, (4) design of complex digital systems, (5) processor design - components, arithmetic logic unit and control unit and (6) programming of the designed processor. The designed set of laboratory exercises was used in Digital Systems Design Principles course. At the end of the semester the students were asked to fill the survey about the laboratory exercises and they gave positive feedback. The new laboratory exercises demand more effort than the previous version, but the end result was increased percentage of students who passed the course, even though the exercises are more challenging.

Web Server and QR Decoder Applications for Xilinx FPGA Boards

2018 Zooming Innovation in Consumer Technologies Conference (ZINC), 2018

The microelectronic industry is undergoing dramatic changes nowadays. Every year faster and more ... more The microelectronic industry is undergoing dramatic changes nowadays. Every year faster and more advanced systems are released. They are capable of running more complex tasks than before for a fraction of a PCs price. The task of this paper is realization of two applications that run on an embedded system. The first one is a QR (Quick Response) image decoder and the second one is a web server. The requirement was for the QR decoder to decode a stream of images and for the web server to process certain HTTP requests. Most existing versions of these applications are designed to run on PCs. Since only a part of their functionalities were needed, the best solution was to develop the applications in C programming language and implement them on a system as bare-metal programs. With this approach the performance can be maximized and the development time of the software shortened.

An insight into antimicrobial activity of the freshwater bryozoan <i>Pectinatella magnifica</i>

The antimicrobial activity of five crude extracts of the freshwater bryozoan Pectinatella magnifi... more The antimicrobial activity of five crude extracts of the freshwater bryozoan Pectinatella magnifica (Leidy, 1851) was evaluated in vitro for the first time. P. magnifica acetone extract exhibited the highest antibacterial activity (MIC 0.004-0.350 mg/mL and MBC 0.007-0.500 mg/mL), while its methanol extract showed the most promising antifungal activity (MIC 0.03-0.12 mg/mL and MFC 0.06-0.25 mg/mL). Furthermore, at a concentration of 0.25 MIC, the methanol extract reduced biofilm formation of the bacterial strain Pseudomonas aeruginosa PAO1 in a considerable extent (59.14%). FTIR spectra of the most active extracts indicate the presence of carbonyl compounds, long-chain alcohols and/or sterols. According to the experimental data obtained, P. magnifica methanol extract may be considered as a good resource of novel natural products with potent anti-biofilm activity against the bacterium well known for its resistance.

Download

Distributed Incremental View Maintenance

Encyclopedia of Big Data Technologies, 2018

The Variable Neighborhood Search Heuristic for the Containers Drayage Problem with Time Windows

Advances in Intelligent Systems and Computing, 2013

The containers drayage problem studied here arise in International Standards Organization (ISO) c... more The containers drayage problem studied here arise in International Standards Organization (ISO) container distribution and collecting processes, in regions which are oriented to container sea ports or inland terminals. Containers of different sizes, but mostly 20ft, and 40ft empty and loaded should be delivered to, or collected from the customers. Therefore, the problem studied here is closely related to the vehicle routing problem with the time windows where an optimal set of routes is obtained. Both delivery and pickup demands can be satisfied in a single route. The specificity of the containers drayage problem analyzed here lies in the fact that a truck may simultaneously carry one 40ft, or two 20ft containers, using an appropriate trailer type. This means that in one route there can be one, two, three or four nodes, which is equivalent to the problem of matching nodes in single routes. This paper presents the Variable Neighborhood Search (VNS) heuristic for solving the Containers Drayage Problem with Time Windows (CDPTW). The results from the VNS heuristic are compared with the two optimal MIP mathematical formulations that were introduced in our previous research papers.

Download

2 Counting Triangles under Updates in Worst-Case Optimal Time 1

We consider the problem of incrementally maintaining the triangle count query under single-tuple ... more We consider the problem of incrementally maintaining the triangle count query under single-tuple updates to the input relations. We introduce an approach that exhibits a space-time tradeoff such that the space-time product is quadratic in the size of the input database and the update time can be as low as the square root of this size. This lowest update time is worst-case optimal conditioned on the Online Matrix-Vector Multiplication conjecture. The classical and factorized incremental view maintenance approaches are recovered as special cases of our approach within the space-time tradeoff. In particular, they require linear-time maintenance under updates, which is suboptimal. Our approach can also count all triangles in a static database in the worst-case optimal time needed for enumerating them. 2012 ACM Subject Classification Theory of computation → Database query processing and optimization (theory); Information systems → Database views; Information systems → Data streams

Download

Enabling Deep Analytics in Stream Processing Systems

Real-time applications often analyze data coming from sensor networks using relational and domain... more Real-time applications often analyze data coming from sensor networks using relational and domain-specific operations such as signal processing and machine learning algorithms. To support such increasingly important scenarios, many data management systems integrate with numerical frameworks like R. Such solutions, however, incur significant performance penalties as relational engines and numerical tools operate on fundamentally different data models with expensive inter-communication mechanisms. In addition, none of these solutions supports efficient real-time and incremental analysis. In this work, we advocate a deep integration of domain-specific operations into general-purpose query processors with the goal of providing unified query and data models for both online and offline processing. Our proof-of-concept system tightly integrates relational and digital signal processing operations and achieves orders of magnitude better performance than existing loosely-coupled data manageme...

Download

Boveda: Building an On-Chip Deep Learning Memory Hierarchy Brick by Brick

Data access between onand off-chip memories account for a large fraction of overall energy consum... more Data access between onand off-chip memories account for a large fraction of overall energy consumption during inference with deep learning networks. On-chip memory compression can greatly reduce this energy cost as long as it balances the simplicity and low cost of the compression/decompression implementation and its effectiveness in data size reduction. We present Boveda, a simple and effective on-chip lossless memory compression technique for fixed-point precision networks. It reduces data widths by exploiting the value distribution deep learning applications naturally exhibit. Boveda can increase the effective on-chip capacity, reduce off-chip traffic, and/or achieve a desired performance/energy target while using smaller on-chip memories. Boveda can be placed after any memory block in the on-chip memory hierarchy and can work with any data-parallel processing units such as the vector-like or the tensorcore units of modern graphics processors (Durant et al., 2017; Jia et al., 201...

Download

DPRed: Making Typical Activation Values Matter In Deep Learning Computing

ArXiv, 2018

We show that selecting a fixed precision for all activations in Convolutional Neural Networks, ev... more We show that selecting a fixed precision for all activations in Convolutional Neural Networks, even if that precision is different per layer, amounts to worst case design. We show that much lower precisions can be used, if we could target the common case instead by tailoring the precision at a much finer granularity than that of a layer. We propose Dynamic Prediction Reduction (DPRed) where hardware on-the-fly detects the precision activations need and at a much finer granularity than a whole layer. We demonstrate a practical implementation of DPRed with DPRed Stripes (DPRS), a data-parallel hardware accelerator that adjusts precision on-the-fly to accommodate the values of the activations it processes concurrently. DPRS accelerates convolutional layers and executes unmodified convolutional neural networks. DPRS is 2.61x faster and 1.84x more energy efficient than a fixed-precision accelerator for a set of convolutional neural networks. We further extend DPRS to exploit activation a...

Efficient Incremental Data Analysis

Many data-intensive applications require real-time analytics over streaming data. In a growing nu... more Many data-intensive applications require real-time analytics over streaming data. In a growing number of domains -- sensor network monitoring, social web applications, clickstream analysis, high-frequency algorithmic trading, and fraud detections to name a few -- applications continuously monitor stream events to promptly react to certain data conditions. These applications demand responsive analytics even when faced with high volume and velocity of incoming changes, large numbers of users, and complex processing requirements. Developing suitable online analytics engine that meets these requirements is challenging. In this thesis, we study techniques for efficient online processing of complex analytical queries, ranging from standard database queries to complex machine learning and digital signal processing workflows. First, we focus on the problem of efficient incremental computation for database queries. We have developed a system, called DBToaster, that compiles declarative queri...

Download

2 Domain-specific Computation in Stream Processing Engines

Download

DPRed: Making Typical Activation and Weight Values Matter In Deep Learning Computing

arXiv: Neural and Evolutionary Computing, 2018

We show that selecting a single data type (precision) for all values in Deep Neural Networks, eve... more We show that selecting a single data type (precision) for all values in Deep Neural Networks, even if that data type is different per layer, amounts to worst case design. Much shorter data types can be used if we target the common case by adjusting the precision at a much finer granularity. We propose Dynamic Precision Reduction (DPRed), where we group weights and activations and encode them using a precision specific to each group. The per group precisions are selected statically for the weights and dynamically by hardware for the activations. We exploit these precisions to reduce: 1) off-chip storage and off- and on-chip communication, and 2) execution time. DPRed compression reduces off-chip traffic to nearly 35% and 33% on average compared to no compression respectively for 16b and 8b models. This makes it possible to sustain higher performance for a given off-chip memory interface while also boosting energy efficiency. We also demonstrate designs where the time required to proc...

Download

Penicillin-induced Cutaneous Necrotizing Eosinophilic Vasculitis with Cryofibrinogenemia

Acta dermatovenerologica Croatica : ADC, 2019

Cutaneous necrotizing eosinophilic vasculitis (CNEV) is a rare type of vasculitis. Eosinophilic v... more Cutaneous necrotizing eosinophilic vasculitis (CNEV) is a rare type of vasculitis. Eosinophilic vasculitis is a necrotizing vasculitis with eosinophilic vascular infiltration, in which eosinophils mediate vascular damage in the disease process. We present a case of an 18-year-old girl who developed palpable purpura and hemorrhagic bullae over the lower extremities associated with itching, 7 days after the commencement of penicillin therapy. Plasma cryofibrinogen was positive. Histopathology showed an infiltration of eosinophils within and around the vessel walls and a complete absence of nuclear dust and neutrophils. Oral prednisone at 1 mg/kg induced remission in 2 weeks; the prednisone dose was tapered and discontinued after 2.5 months. There was no evidence of recurrence after 37 months of follow-up. Our patient represents a rare case of drug/penicillin-induced CNEV associated with cryofibrinogenemia, without systemic organ involvement.

Download

Enabling Signal Processing over Data Streams

Proceedings of the 2017 ACM International Conference on Management of Data, 2017

Internet of Things applications analyze the data coming from large networks of sensor devices usi... more Internet of Things applications analyze the data coming from large networks of sensor devices using relational and signal processing operations and running the same query logic over groups of sensor signals. To support such increasingly important scenarios, many data management systems integrate with numerical frameworks like R. Such solutions, however, incur significant performance penalties as relational data processing engines and numerical tools operate on fundamentally different data models with expensive intercommunication mechanisms. In addition, none of these solutions supports efficient real-time and incremental analysis. In this paper, we advocate a deep integration of signal processing operations and general-purpose query processors. We aim to reconcile the disparate data models and provide a common query language that allows users to seamlessly interleave tempo-relational and signal operations for both online and offline processing. Our approach is extensible and offers frameworks for quick and easy integration of user-defined operations while supporting incremental computation. Our system that deeply integrates relational and signal operations, called TRILLDSP, achieves up to two orders of magnitude better performance than popular loosely-coupled data management systems on grouped signal processing workflows.

Download

Characterizing Sources of Ineffectual Computations in Deep Learning Networks

2019 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), 2019

Hardware accelerators for inference with neural networks can take advantage of the properties of ... more Hardware accelerators for inference with neural networks can take advantage of the properties of data they process. Performance gains and reduced memory bandwidth during inference have been demonstrated by using narrower data types [1] [2] and by exploiting the ability to skip and compress values that are zero [3] [4] [5] [6]. Similarly useful properties have been identified at a lower-level such as varying precision requirements [7] and bit-level sparsity [8] [9]. To date, the analysis of these potential sources of superfluous computation and communication has been constrained to a small number of older Convolutional Neural Networks (CNNs) used for image classification. It is an open question as to whether they exist more broadly. This paper aims to determine whether these properties persist in: (1) more recent and thus more accurate and better performing image classification networks, (2) models for image applications other than classification such as image segmentation and low-level computational imaging, (3) Long-Short-Term-Memory (LSTM) models for non-image applications such as those for natural language processing, and (4) quantized image classification models. We demonstrate that such properties persist and discuss the implications and opportunities for future accelerator designs.

Download

DNA methylation changes during long-term in vitro cell culture are caused by epigenetic drift

Communications Biology, 2021

Culture expansion of primary cells evokes highly reproducible DNA methylation (DNAm) changes. We ... more Culture expansion of primary cells evokes highly reproducible DNA methylation (DNAm) changes. We have identified CG dinucleotides (CpGs) that become continuously hyper- or hypomethylated during long-term culture of mesenchymal stem cells (MSCs) and other cell types. Bisulfite barcoded amplicon sequencing (BBA-seq) demonstrated that DNAm patterns of neighboring CpGs become more complex without evidence of continuous pattern development and without association to oligoclonal subpopulations. Circularized chromatin conformation capture (4C) revealed reproducible changes in nuclear organization between early and late passages, while there was no enriched interaction with other genomic regions that also harbor culture-associated DNAm changes. Chromatin immunoprecipitation of CTCF did not show significant differences during long-term culture of MSCs, however culture-associated hypermethylation was enriched at CTCF binding sites and hypomethylated CpGs were devoid of CTCF. Taken together, o...

Download

Improving design and operating parameters of the recuperator for waste heat recovery from rotary kilns

Thermal Science, 2021

Depending on their applications, heat losses from the shells of rotary kilns account for 3-25% of... more Depending on their applications, heat losses from the shells of rotary kilns account for 3-25% of the total heat input. Over the hottest zone of the kiln shell, an annular duct with a variable diameter is formed. Two air streams entering the annulus at both ends flow to a common extraction point to receive the thermal power equal to the ambient heat loss of the bare kiln. The design does not require airtightness, utilizes the entire heat loss, and by the variation of the air-flow can be used over the kilns with variable operating parameters (?20% heat loss), which show similar surface temperature patterns. The main disadvantage of the design is the approaching of the surfaces of the kiln and the recuperator at the outlet of preheated air. This means that for a given heat loss and a surface temperature pattern, the rotational eccentricity of the kiln shell is the most influencing parameter that defines the air preheating temperature and the efficiency of the recuperator. To solve the...

$Unlike the RPP design, the introduction of radiation fins (see fig. 2 b)) has a positive effect o: the critical diameter. Compared with the basic design with the minimal diameter of 2.84 m (see fig. - a), RF (see fig. 2 b) and tab. 6) with the fins at the diameter of 2.84 m has more than 30 °C highe preheating temperature. A larger surface area for the convective heat transfer causes the improvement Compared with the basic design, the surface area is more than 60% larger, but the benefit is modes because the majority of the heat loss is transferred to the air by convection at the kiln surface. The kil shell temperature limits the impact of thermal radiation. It decreases with the rise of air temperature Tab. 6 shows that the distance between the radiation fin and the recuperator becomes very small nea the exit. That suggests that instead of the use of radiation fins, the recuperator surface should be finne« in the outlet sections. In fact, in the air outlet sections preferable option is the application of fins on th kiln surface. and the ratio of heat loss transferred to the air in the annuluses. After mixing all 4 streams, the air preheating temperature would be 210 °C. Compared with the basic design (fig. 1 and 5 a)) for the same critical diameter, the preheating temperature for RPP is lower. The reason for this is the temperature of the kiln shell, which cannot transfer the majority of the heat loss to the recuperator surface by thermal radiation. The positive features of the RPP design that allow very small gaps and usage of fins between the surfaces in the outer passage cannot be utilized except to prevent overheating of the outlet sections. The majority of the heat loss can be transferred to the surface of the recuperator and the second annulus only at the air exit sections because there, thermal radiation from the kiln surface is accompanied with the cooling of already preheated air. The isothermal mixing of 4 streams is possible and is achievable only when a small fraction of air flows through the outer passage. E.g., for the minimum diameter in the inner annulus of 2.88 m, isothermal mixing at 138 °C is achieved if two air streams of 6.926 kg/s and 0.606 kg flow through the inner and outer annuluses, respectively.$

Download

TraNCE

Proceedings of the VLDB Endowment, 2021

Nested relational query languages have long been seen as an attractive tool for scenarios involvi... more Nested relational query languages have long been seen as an attractive tool for scenarios involving large hierarchical datasets. There has been a resurgence of interest in nested relational languages. One driver has been the affinity of these languages for large-scale processing platforms such as Spark and Flink. This demonstration gives a tour of TraNCE, a new system for processing nested data on top of distributed processing systems. The core innovation of the system is a compiler that processes nested relational queries in a series of transformations; these include variants of two prior techniques, shredding and unnesting, as well as a materialization transformation that customizes the way levels of the nested output are generated. The TraNCE platform builds on these techniques by adding components for users to create and visualize queries, as well as data exploration and notebook execution targets to facilitate the construction of large-scale data science applications. The demon...

Download