D Roy Choudhury Networks And Systems Pdf 110
Edelira Longinotti
Nature is fond of geometry, and different molecular structures show unique geometric traits. However, training neural networks to predict properties of such geometric structures comes with a unique set of challenges. Our ongoing work focuses on accelerating the training of such neural networks on emerging AI accelerators, heterogeneous computing that integrates simulations and machine-learning on supercomputers, and creative combination of different machine-learning techniques to support wider range of chemical systems.
What do chemical compounds, power grid and Wikipedia have in common? All of them are a manifestation of diffrent entities coming togeher, something that is commonly modeled as graphs, or networks. A significant part of my career is dedicated to "connecting the dots": learning and building tools that often map into diverse fields in computer science: data mining techniques to extract structural patterns, searching a graph database for complex patterns, semantic reasoning using knowledge graphs, predictive and generative machine learning methods for graphs, and explaining decisions on complex multi-modal data in natural language. StreamWorks, a system we developed for continuous pattern detection and reasoning received a R&D100 award in 2018 for it's application into cyber-security.
However, sampling-based kinetic modelling frameworks frequently produce large subpopulations of kinetic models inconsistent with the experimentally observed physiology. For instance, the constructed models can be locally unstable or display too fast or too slow time evolution of metabolic states compared with the experimental data (Fig. 1). This entails a considerable loss of computational efficiency, especially for the low incidence of subpopulations with desirable properties. For example, the generation rate of locally stable large-scale kinetic models can be lower than 1% (ref. 20). Requiring other model properties such as experimentally observed time evolution of metabolic states further reduces the incidence of desired models. Indeed, just a tiny fraction of the parameter space satisfies all desirable model properties simultaneously, and our observations suggest that this subspace is not contiguous. Moreover, none of these methods guarantees that the sampling process, often implemented as unbiased samplifng, will produce the desirable parameter sets. These drawbacks become amplified with increasing size of the kinetic models, and finding regions in the parameter space that satisfy the desired properties and observed physiology becomes challenging. Additionally, the structure of these regions is so complex that nonlinear function approximators such as neural networks are required to map them (Supplementary Notes 1 and 2).
Conditional GANs consist of two feedforward neural networks, the generator and the discriminator, which are conditioned on class labels during training. The goal of the training procedure is to obtain a good generator that generates kinetic models (Fig. 1a, step 3) from a specific predefined class that are indistinguishable from the kinetic models of the same class in the training data (Methods).
Comprehensive analyses of metabolic networks require large populations of parameter sets. However, a trade-off exists between generating large datasets and computational requirements, which can limit the scope of studies depending on the efficiency of the methods employed.
We conclude that transfer learning successfully captures the specificities of the physiologies (Supplementary Note 7). With only a few kinetic parameter sets, transfer learning allows the generation of kinetic models that possess the desired properties of biological relevance, robustness and parametric diversity. We anticipate that this approach could help to derive new methods for high-throughput analysis of metabolic networks.
The scarceness of experimentally verified information about intracellular metabolic fluxes, metabolite concentration and kinetic properties of enzymes leads to an underdetermined system with multiple models capable of capturing the experimental data. Due to the requirement of intense computational resources aiming to quantify the involved uncertainties, researchers often end up using only one out of the many alternative solutions, leading to unreliable analysis and misguided predictions about metabolic behaviour of cells. This is one of the reasons for the limited use of kinetic models in studies of metabolism, despite their widely acknowledged capabilities. REKINDLE offers a highly efficient way of sampling the parameter space and creating kinetic models, thus enabling an unprecedented level of comprehensiveness for analysing these networks and offering a much broader scope of applicability of kinetic models. In general, sampling of nonlinear parameter spaces has emerged as a standard method in addressing underdeterminedness in computational physics, biology and chemistry43,44.
In GANs, two neural networks, the generator that we train to generate new data and the discriminator that tries to distinguish generated new data from real data, are pitted against each other in a zero-sum game. The end goal of this game is to obtain the generator that generates new data of such a quality that the discriminator cannot distinguish it from real data (Fig. 1a, step 3). We train the generator and discriminator networks in turn. To train the discriminator, we freeze the generator by fixing its network weights. Then, we alternate the training by freezing the discriminator and train the generator. In the first part of each learning step, we provide the discriminator with (1) a random batch of kinetic parameter sets from the training data with labels indicating the class of models and (2) a batch of kinetic parameter sets that have been generated by the generator (fake data). The discriminator then classifies the models it is presented with as real (from the training set) or fake (from the generator). In the second part of a learning step, the discriminator is frozen and the generator generates a batch of fake kinetic parameter sets using as inputs (1) random Gaussian noise and (2) sampled labels. The discriminator and the generator improve their performance with training. The generator becomes better at deceiving the discriminator, and the discriminator becomes better at classifying between training and generated data (Fig. 1a, step 2). The training continues until we reach equilibrium between the two neural networks, and no further improvement is possible.
The minimum number of data required to train a randomly initialized GAN for generating relevant models depends on the sizes of the neural networks used and the metabolic system studied. Determining the minimal number of data for training as a function of the size of the studied metabolic system remains an open problem.
Tanzeem Choudhury is a professor at the Jacobs Technion-Cornell Institute at Cornell Tech in Information Sciences and a co-founder of HealthRhythms Inc. She directs the People-Aware Computing group, which works on developing machine learning techniques for systems that can reason about human activities, interactions, and social networks in everyday environments.
Dr. Choudhury received her Ph.D. degree from the Media Laboratory at the Massachusetts Institute of Technology (MIT). As part of her doctoral work, she created the sociometer and conducted the first experiment that uses mobile sensors to model social networks, which led to a new field of research referred to as Reality Mining. She holds a B.S. in electrical engineering from the University of Rochester, and M.S. from the MIT Media Laboratory.
Chowdhury leads SymbioticLab, which tailors Big Data and AI applications to coexist with their underlying networks and applies networking principles in designing new data systems. His research supports a wide variety of data systems, ranging from those running on networks with single-microsecond latency to those running over the Internet.
He recently received an NSF CAREER award in support of a project aimed at memory disaggregation, which takes advantage of emerging low-latency networks to expose all the unused memory in a data cluster as a single, massive memory unit in order to improve the performance of memory-intensive applications that require low latency and the use of massive amounts of data. His recent work on remote memory prefetching has resulted in a system called Leap, which earned a Best Paper Award at the 2020 USENIX Annual Technical Conference. His other ongoing projects include federated computation, systems for AI, big data systems, and datacenter networking.
Millimeter Wave (mmWave) networks can deliver multi-Gbps wireless links that use extremely narrow directional beams. This provides us with a new opportunity to exploit spatial reuse in order to scale network throughput. Exploiting such spatial reuse, however, requires aligning the beams of all nodes in a network. Aligning the beams is a difficult process which is complicated by indoor multipath, which can create interference, as well as by the inefficiency of carrier sense at detecting interference in directional links. This paper presents BounceNet, the first many-to-many millimeter wave beam alignment protocol that can exploit dense spatial reuse to allow many links to operate in parallel in a confined space and scale the wireless throughput with the number of clients. Results from three millimeter wave testbeds show that BounceNet can scale the throughput with the number of clients to deliver a total network data rate of more than 39 Gbps for 10 clients, which is up to 6.6x higher than current 802.11 mmWave standards.
I started my career as a Software engineer working on software to manage telecoms networks.
I then worked in IT for over ten years, providing planning and implementation of systems both in the field and from an office. To better support my Clients, I studied and passed the six exams leading to an MCSE in Windows Server.
I have experience working with clients from the public and private sector, including SMEs, Investment Banks, Schools, and local Government Organizations such as Councils. Over time my work became more security-focused with projects that involved the design and configuration of cloud infrastructure, anti-Spam solutions, VPNs, Firewalls, and ensuring that companies with card payments systems were compliant with PCI DSS.
I have successfully obtained qualifications in the field of security, including the CISSP from ISC2, and I undertook the MSc in Information Security, achieving a Distinction from Royal Holloway.
I have accepted an invitation to be part of the CDT at RHUL, where I am working towards a PhD in Cyber Security. My current research interests include enhancing the security available to Small Businesses, Network Security, Malware Sandboxes, Machine Learning and Reverse Turing tests.