The Algorithms, Analytics, Augmented Intelligence & Devices research area consists of several teams working across network protocols, algorithms, complex systems, and real-time analytics.
Mathematics, Algorithms & Fundamentals
In the last decade, computers have come to assist humans in new cognitive tasks that were previously outside computer automation, such as identifying faces and voices and making sense of the written language. Key to these new functionality of computers is reusable algorithms that are derived from sound formal and mathematical principles, as was the case with the initial computer revolution of the 20th century said to be on a par with Moore’s Law; e.g., see SIAM’s list of top 10 algorithms of the 20th century as examples of the former. We already see the emergence of two such fundamental algorithms in the first two decades of the 21st century: PageRank and Convolutional Neural Networks, with the latter still in need of clear boundaries for its applicability.
The vision of the Math & Algorithms Team is to develop novel reusable algorithms in machine learning for the information and communication industry. Our mission is to apply these algorithms in the context of Bell Labs’ large projects aimed at solving future problems facing our Business Divisions.
- Information Theory in machine learning (math of communication) – aims to leverage methodology, techniques and tools from information theory for machine learning, with examples such as lower and upper bounds for ANN levels and layers, feature compression, regularization and selection.
- Graph algorithms for data analysis (math of networks) – aims to leverage the well-recognized graph structure in many data sets for extracting information and knowledge, with examples such as clustering and community detection in graphs with special structures in real data (hyperbolic graphs), and layer extraction in optical coherence tomography.
- Automated operations research & management (math of systems) – aims to use ML methodology to better solve traditional resource allocation problems in setting where real-time data is available, such as use of reinforcement learning for traffic control, inventory control and fulfilment management.
- Fast Forward Neural Networks
- Neural Networks with Adjustable Quantization
- Digital Parameter Reduction via L1-Regularization
- Text Summarization
- Ticket Classification
- Optical Coherence Tomography Analytics
- Deep localization
- Predictable Service Delivery via Reinforcement Learning
- 5G Slicing
Networked Systems & Automation Research
With the ongoing transition from a human-centric network to a machine-centric network filled with sensors, bots, robots, drones and smart objects where every object is network attached, we face unprecedented challenges in scaling, securing, managing and optimizing networks. Since there is no bound to the number of machines, or the service requirements of individual machine classes, tomorrow’s networks will need to hyper-scale to support billions of nodes while supporting ultra-dynamic and diverse workloads from the vast pool of machines ranging from dumb nodes with minimal requirements to sophisticated endpoints with stringent, real-time demands.
The group’s research focuses on two broad themes. One theme focuses on solving the network automation and optimization challenges associated with hyper-scaled, low latency, machine-oriented networks. This includes advanced routing algorithms for SD-WANs, MPLS, Segment Routing optimizations, modular data planes and data plane acceleration for SDNs, Optimizing networks for IoT and Industrial automation, performance characterization of new data-center switch architectures.
The second theme is focused on network augmentation leading to networks that learn. Networks will need to automate their operations far beyond the current state-of-the-art to support the envisaged level of scale efficiently. Networks will need automated techniques to configure and onboard devices, secure the network from the vast attack surface posed by the large number of connected devices, optimize the network to carry more revenue bearing traffic while providing an enhanced Quality-of-Experience (QoE) to endpoints, and predict future demand, network hotspots and equipment outages. For hyper-scale automation, the network needs to self-adapt to different requirements automatically with virtually no human intervention.
- Data-driven optimization of networks: Solving the network automation and optimization challenges associated with hyper-scaled, low latency, machine-oriented networks.
- Application-customized dynamic data planes: Enabling high performance, agile networking for the cyber-physical world.
- Networks that learn: Networks that learn to self-adapt automatically to changing requirements with virtually no human intervention.
- Adaptive Data Center Resource Allocation using Reinforcement Learning
- Machine Learning based Microservice Fingerprinting and Classification
- Barricade: Traffic fingerprinting for network attack detection
- Flexible and Programmable Data Plane: Virtual Network Switch and IPSEC Acceleration with Intelligent Offload
AI is thought to be the technology that will expand human limits, solve human needs and advance our species in unprecedented ways. AI, today, can be thought as a sort of conventional interpreter on the data we feed it.
The current generation of AI technology focuses on extracting meaning from data and reasoning with this information as though it held unconditioned universal truth. There are three fundamental flaws in this approach: 1) datasets are not complete, neither are they unique or static, thus they produce only approximations of reality and cannot model a universal truth; 2) data contain a reflection of producer’s model of the reality and without considering this bias in the AI pipeline, the resulting information and therefore derived meaning inaccurately captures the original intent; the receiver's state of mind needs to be considered during the generation of the model so that the receiver will perceive the right intended meaning. And, 3) the context makes the same data have different meanings to different individuals.
Bell Labs’ Analytics team aims to redefine the way we approach the meaning association to data by designing novel algorithms that are able to understand and explain the dynamicity and diversity of our complex world, the incompleteness and biases in the data and finally the need to interpret the personalized and contextualized factors of the consumer of the resulting knowledge. This research has profound impact in those fields where AI has the potential to influence our decisions in unprecedented ways and we owe humanity the thoughtful engineering of AI systems that inspire deep thinking, diversity of perspectives, empathic mental alignment and convergence of mutual understanding as much as our, in person, communications.
- Social analytics: The team works on the idea that digital data allows for the study of people and society in unprecedented ways. As part of the emerging research area of “Computational Social Science”, the team answers questions typical of the social sciences using computational methods to produce long-term research (in very ‘Bell Labs’ style).
- Analytics for autonomous systems: True machine intelligence needs to focus on causally relevant aspects of the ever-growing information flow in order to understand the current situation, extract the meaning of the observations, and learn how to take the most appropriate decisions. To enable the emergence of the future self-actuated systems the team elaborates reinforcement learning based control mechanisms that integrate with causal inference disciplines.
- Analytics for verbal and non-verbal communication: Devising novel analytics for a personalized meaning computation involving both the verbal and the non-verbal dimensions of the human communication.
- Machine learning for NP-Hard problems: this research investigate how machine learning can be used to detect and exploit structure in practically relevant instances of computationally difficult problems.
- The Theory of Meaning: Devising a novel computational model to capture the multidimensional aspects of the human communication from verbal to non-verbal for an accurate personalised meaning association, extraction and computation.
- Root Cause Analytics (RCA): We elaborate automated intelligence able to interpret the flow of measurements and alarms generated by a large monitored system in order to bring explanations in terms of underlying root causes and causal propagations.
- Market Monitor: Interactive news consumption framework which enable reader to quickly cover the knowledge with fast learning capabilities of the reader’s current interest and readjustment accordingly.
- Neuron Marking: A novel technique to mark neurons that resists a large variety of attempts to copy and modify a neural net
- Machine Understanding of Emotions: New AI systems that track and explain emotional changes of individuals based on operationalization of social science theories.
Deep Modeling of Complex Systems
Modeling complex phenomena has gained a leap in performance as the availability of big data sets and new computing power have made it possible to surpass the feature engineering and rule-based programming paradigms.
It has now become possible to tackle problems, in which we do not have the knowledge of the first principles to help us. Mathematical models can be created even if these would be too complicated for humans to understand. Deep learning allows to extract value from complex physical systems and big data. We do research on methods how to efficiently train and use deep learning models.
- Expand human knowledge through machine knowledge: As we are now able to create models for extremely complex processes, we can use this to gain deep insights of our own behavior, for example, we can now find and quantify hidden biases in peoples’ decision making and study its integrity.
- Explore structure in data using generative models: Using generative methods, we can investigate how much data is really needed to reproduce the originals and how to create new artificial datasets that have the same characteristics as the original.
- Generating synthetic data for ML models: Synthetic data from accurate simulations can be used to push the limits of what can be measured - we can scale the training beyond what’s possible today with experimental means.