Raman Kannan

Followers

Following

Public Views

Uploads

Papers by Raman Kannan

Benchmarking Supervised Learning Frameworks for Engineering Highly Scalable Intelligent Systems

Procedia Computer Science

There are numerous freely available machine learning frameworks, built using R[1], Python . Howev... more There are numerous freely available machine learning frameworks, built using R[1], Python . However, business applications continue to be dominated by java/j2ee, requiring cross-language environment to leverage machine learning capabilities in R or python. Apache SparkML© [3], a java based M/L infrastructure, with native hooks for Java, has the potential to fulfill this need in the public domain. Unfortunately, a rigorous comparative analysis of machine learning capabilities of Spark vs python/R is not available. Furthermore, R and python continue to be hampered by different design constructs including in-memory requirement and interpreted execution. These two factors inhibit the adoption rate and with an environment "naturally" interoperable with java/j2ee, such problems can be eliminated. We believe an objective comparative analysis can provide required decision matrix, allowing engineering teams to determine if SparkML can meet their needs. As a consequence, many hard/soft benefits that come with SparkML, can be availed. In this experiment our quest was to understand if SparkML is ready for supporting ML Classification Projects and if not identify the gaps, in comparison to R. We performed numerous classification experiments, in a controlled environment, using SparkML and R. SparkML was run in single/local mode. We compared the performance of the standard algorithms using AUC. One would expect very little difference but to our surprise there were differences and some of them were significant that cannot be easily accounted for by floating point or other factors impacting numerical precision. In the process we have identified algorithms yet to be implemented in SparkML. We present our method, results and future work..

Download

Next Generation Information Comprehension Aids

Assisted Cognition

No Free Lunch: Free at Last!

Augmented Cognition. Theoretical and Technological Approaches, 2020

No Free Lunch (NFL) Theorem in M/L is rigid and inflexible, which states that “No particular clas... more No Free Lunch (NFL) Theorem in M/L is rigid and inflexible, which states that “No particular classifier can outperform all the other classifiers for every dataset”. In this paper we present a MISD machine that runs multiple classifiers in parallel against a given dataset, implementing a “Swiss Army Knife” to combine many different classifiers and review their performance, effortlessly. The service will be hosted as a public service over the internet for any Machine Learning practitioner to experiment with datasets.

Kantian Computational Linguistics

Language is a defining characteristic of higher order species-particularly of us, homo sapiens. F... more Language is a defining characteristic of higher order species-particularly of us, homo sapiens. For that reason, it is only appropriate to conclude that linguistic capability-the ability to understand and generate natural language-is essential in any tools that we build for cognitive assistance. Fast and superior automated solutions exist in most other competencies, including numeracy and vision, but parity in linguistic skills remains elusive [1] for many reasons. In this chapter, we seek to identify why we find ourselves here and propose an alternative approach to overcome identified deficiencies. We do not know yet if this approach is useful or effective, but we proceed to share our thoughts because the biggest risk is doing nothing. First, we review seminal work on linguistics and the structure of language. We then examine the most popular contemporary solutions, identify their drawbacks and propose alternatives to address them.

The Way We Think About Ourselves

In the new normal of fake news, wide-scale disinformation and alternate facts, the need for fact-... more In the new normal of fake news, wide-scale disinformation and alternate facts, the need for fact-checks and bot detection is real and immediate. It is generally accepted that altering the psyche, our thinking, is the most potent form of controlling and shaping human behavior. Fake news, disinformation and alternate reality are all aimed at shaping our beliefs. In this experiment, we set out to understand if the stream of news articles is itself designed to influence society at large, either to think positively or negatively – in other words can dictate how society views itself – disconnected from reality. In this exercise we are seeking to identify if there is a systematic prevalence of positive/negative sentiment in a given stream of news articles, using standard NLP techniques.

Configurable Parallel Induction Machines

Machine Learning practice in general offers significant opportunities for parallel computing and ... more Machine Learning practice in general offers significant opportunities for parallel computing and practicing sound software engineering. More often than not, practitioners routinely write dataset specific scripts and learners focus on model building and refining. Focusing on particular models is not consistent with NFL, a fundamental theorem in Machine Learning. Not minding time-honored software engineering principles is inefficient. In this paper, we present our implementation of MISD machine, consistent with No Free Lunch Theorem, problems we encountered and our approach to solve those problems.

Variation of the Neel temperature with magnetic dilution in CopMg1-pO

Journal of Physics C: Solid State Physics, 1986

... Phys. 52 1785 Kumar D, Pandy RB and Barma M 1981 Phys. Rev. B 23 2269 McGum AR and Tahir-Khel... more

NEUTRON DIFFRACTION IN Co p Mg 1-p O SOLID SOLUTIONS

Le Journal de Physique Colloques, 1988

Magnetic Properties of Randomly Diluted Antiferromagnetic System: COBALT(P)MAGNESIUM(1-P)OXYGEN

In this work, the effect of randomly diluting CoO (a fcc antiferromagnet with Neel temperature T ... more In this work, the effect of randomly diluting CoO (a fcc antiferromagnet with Neel temperature T _{rm N} = 289 K) with MgO has been investigated using temperature dependent dc magnetic susceptibility measurements. About twenty samples of Co _{rm p}Mg _{rm 1-p}O with p = 0.10, 0.13, 0.17, 0.22, 0.23, 0.31, 0.33, 0.36, 0.41, 0.46, 0.49, 0.53, 0.56, 0.60, 0.64, 0.70, 0.80 and 0.87, were investigated in the temperature range of 1.6 K to 300 K. These powder samples were prepared starting from the nitrates, Co(NO _3)_2cdot6H _2O and Mg(NO_3)_2 cdot6H_2O. The samples were characterized by magnetic measurements for the presence of ferromagnetic impurities, by X-ray diffraction technique for the determination of crystal structure and lattice constants, and by atomic absorption spectroscopy for the determination of the composition parameter p. The lattice constant is found to vary linearly with p, in accordance with Vegard's Law. The magnetic susceptibility, chi , measurements were carrie...

Dynamic RPC for extensibility

Eleventh Annual International Phoenix Conference on Computers and Communication [1992 Conference Proceedings], 1992

call in form (i.e. setting up IPC abstractions, sending and receiving data are hidden). For examp... more call in form (i.e. setting up IPC abstractions, sending and receiving data are hidden). For example, the RPC based NFS call is the same as the usual system call to acquire a file descriptor for a local file; and Remote Procedure Call (RPC) is admittedly a valuable mechanism for implementing distributed systems. However, certain inherent features of RPC limit the level of extensibility and dynamism one can achieve in distributing a computational effort. In this article, we propose an enhancement to RPC, called open(filename, ...) Dynamic GC@RPC), to facilitate dynamic clientserver relationships, extensibility, and protocol transparency. A prototype DRPC has been implemented as a component of a distributed computing infrastructure, Communications Manager (CM), for the DARPA Initiative in Concurrent Engineering (DICE) program. DRPC currently supports two popular commercial implementations of RPC-Sun ONC/RPC and Apollo NCS/RPC. A preliminary comparative performance analysis is presented. (2) when the client makes a RPC call, the client is suspended until the remote procedure is completed, just as in a local call. The main difference, however, lies in the mechanics of the execution of the procedure. In the case of a normal procedure, the calling process is the same as the executing process, and the U 0 parameters are passed via a local stack segment according to some ianguage level parameter passing semantics. In the case of RPC, however, the procedure is executed by a different process which provides the service [See Birrel and Nelson 1985 or Mullender 1990 for other details], and the U 0 parameters are passed via some IPC abstraction (sockets or streams) exclusively by value. The salient features that render RPC as an effective tool for building distributed systems are: (1) €&mote Procedure call (RPc) is a popular the transparency, especially the location mechanism for distributing control and information independence, and (2) the abstraction in terms of within a heterogenous network of computing pc. elements. In its simplest form, RPC involves a client and a server, and, for most purposes, RPC looks and behaves like a conventional procedure call:

Download

Percolation effects and magnetic properties of the randomly diluted fcc systemCopMg1−pO

Physical Review B, 1987

Temperature-dependent, low-field (50 Oe) dc magnetic susceptibility +) studies are reported for t... more Temperature-dependent, low-field (50 Oe) dc magnetic susceptibility +) studies are reported for the randomly diluted fcc system Co~Mgl~O. The studies, using powder samples, covered the temperature range 1.6-300 K and the composition range 0. 10&p & 1. Variations of the Neel temperature T&, the Curie-Weiss temperature 6, the Curie constant C, the low-temperature magnetization M, and the lattice constant ao with p are reported. The observed percolation threshold p, =0.13 agrees well with the predicted value of p, =0.136 for fcc system with nearest-neighbor and next-nearestneighbor interactions included. The variation of 6(p) and CM (p) with p is found to be linear whereas that of t = T~(p)/T&(1) is more complex, showing a crossover near p =0.45, where d t/dp changes sign. Calculations are needed to explain this observed t versus p variation. For p, &p &0.45, g below T& depends upon how the sample is cooled through T~, and the hysteresis and time dependence (relaxation) of P are observed at 4.2 K if the sample is cycled through a higher field (6 kOe). These observations are not fully understood. An estimate of the exchange constants for the Co~Mgl~O system is given.

Download

Application Message Interface

Eleventh Annual International Phoenix Conference on Computers and Communication [1992 Conference Proceedings], 1992

Magnetic specific heat and critical magnetic susceptibility of the diluted antiferromagnet CopMg1−pO

Journal of Applied Physics, 1993

For CopMg1−pO, a prototype randomly diluted fcc antiferromagnet, measurements of the specific hea... more For CopMg1−pO, a prototype randomly diluted fcc antiferromagnet, measurements of the specific heat versus temperature show that the size of the λ anomaly at TN decreases as p decreases. These results show that with increasing dilution, a progressively larger fraction of the magnetic entropy is removed by short-range order processes. The critical susceptibility χc at TN increases as p decreases, supporting the idea that near Mg sites, the balance of the two antiferromagnetic sublattices is destroyed. After correcting for single ion contribution, χc(p)/χc(1) is found to vary as (p−pc)−τ for p<0.47 (pc=0.136). The experimental value of τ≊0.7 is in reasonable agreement with the numerical estimate of τ=0.5±0.2 by Harris et al. for cubic systems.

DAIS: a distributed AI programming shell

IEEE Expert, 1990

The design and implementation of a distributed AI programming shell, DAIS, which provides an inte... more The design and implementation of a distributed AI programming shell, DAIS, which provides an integrated, domain-independent programming environment for building distributed AI programs on PCs, are described. The target platform is the DOS environment, and DAIS has been implemented in C. The DAIS distribution characteristics, architecture, and knowledge-representation scheme are discussed. Its rule-based features and communication functions are examined. A

PAcceptor and SConnector frameworks

ACM Computing Surveys, 2000

In this article we present the PAcceptor and SConnector frameworks that facilitate rapid deployme... more In this article we present the PAcceptor and SConnector frameworks that facilitate rapid deployment of applications that require both concurrency and inter-process communication system services. The PAcceptor and SConnector frameworks encapsulate both threads -primitive for concurrency --and sockets -primitive for inter-process communication --. These frameworks present a single interface to these disparate but essential services and thus promote the use of these abstractions in combination with ease. These frameworks may be customized to fabricate arbitrary applications by providing appropriate hotspot definitions. The frameworks obviate the need for explicit understanding and expertise over the underlying system servicessockets and threads -as well as the intricacies involved in combining the native socket and thread level services. The PAcceptors accept incoming connect requests using socket and provide service in parallel using threads while the SConnectors permit one or more threads to share a single connection to send data over the connection. Robust infrastructures for distributed and concurrent computing have been available for well over a decade and yet significant effort is required to exploit these services. As evidenced by a recent Technical Opinion column "Parallel Computing Still Not Ready for the Mainstream", in CACM the reason is neither due to the lack of infrastructures nor due to our cognitive ability to manage parallelism and concurrency. To the contrary parallelism and concurrency are quite innate and we thrive based on our ability to solve problems concurrently. The reason for the lack of use of these powerful technologies, we posit, is due to the combinatorial explosion of the conceptual difficulty in leveraging these abstract albeit powerful system services. Simplicity and maintainability is chosen over --better performance-and overwhelming complexity. Thus, in this article we present the PAcceptor and SConnector frameworks --which encapsulate these system level abstractions, hiding many of the difficult conceptual aspects--such that these high performance services may be leveraged sans their associated conceptual barriers. Programming with threads --an abstraction for concurrency --and sockets --an abstraction for inter-process communication --is anything but simple. Both conceptually { timing errors, race conditions, synchronization } and programmatically { thread attributes, socket options, and notational nuances } programming with these services is highly distractive in nature for application developers. Combining these two abstractions is even more complex. Understandably, more often than not engineers trade performance for simplicity and choose not to use these high performance devices solely due to the complexity involved. It is in this context that frameworks become highly relevant. Frameworks mitigate the barriers and promote the use of such complex abstractions by encapsulating them, while preserving the associated semantics [See 2 for examples].

Download

Disha: A Direction Toward Software Engineering on the Net

World Conference on the WWW and Internet, 1997

Assisted software engineering over the net with a goal to help engineers discover components avai... more

Raman Kannan

Uploads

Papers by Raman Kannan

Log In