Towards AI/ML-Driven Network Traffic Engineering

Internet is one of the largest scale distributed systems made up of multiple networks that is used to digitally connect billions of users. Traffic Engineering (TE) is a core problem in networking, which is responsible for routing packets across networks to provide best user experience while ensuring secure, stable, well-utilized and cost-efficient network. The time-varying graph nature of the network along with unexpected topology and traffic changes makes TE in large operational networks challenging. This paper first provides a holistic sketch of different viewpoints taken by researchers and practitioners to formulate and solve the TE problem. We first cover the systems view where researchers have defined the problem and used creative heuristic protocols to manage large networks. We then focus on the theoretical formulations from optimization and control theory to provide optimal and stable networks. These formulations provide clear definitions and provable properties for designing TE. We devise a taxonomy of existing studies on how such theoretical problems are being solved today. Finally, we present the AI/ML and heuristic perspective for near-optimal TE. Owing to the dynamism and large scale of the networks, especially planet-scale cloud networks, these theoretical methods need real-time data and greedy approximations to solve TE. Recent progress in AI/ML provides encouraging tools here, where time-series and graph based AI/ML models can be used to detect/predict the network state, and control algorithms such as Reinforcement Learning (RL) be powerful tools to solve TE. We survey these approaches and propose promising directions towards AI/ML in network control and TE.

 Marina Thottan is Principal Research Scientist at AWS Amazon. Prior to joining Amazon, she was Vice President of Network and Security research at Bell Labs. Marina has contributed to a wide variety of research areas, including Content Distribution, Routing protocols, Data over Optical networks, High Speed Router Design, Network Management, Anomaly Detection and Smart Grid Networks. Most recently she has been leading research work on Cloud network engineering and operations.
Marina received a Ph.D. in Electrical and Computer Engineering from Rensselaer in 2000. She has published over 70 papers in scientific journals, book chapters and conferences and holds several patents in the areas of network management, interactive network applications, routing algorithms, data analytics and network architectures. She is co-author of the book “Communication Networks for Smart Grids: Making Smart Grids Real” and has also Co-edited a book on “Algorithms for Next Generation Networks”.
Marina is a Bell Labs Fellow and an IEEE Fellow.

CLarge Language Models in Telecommunications: Is the Network Finally “Intelligent”?

In the 1990’s the concepts of an “Intelligent Network” started to take shape in telecommunication networks.  The intelligence envisioned then was, of course, constrained to algorithmic intelligence, i.e., the algorithms drove the intelligence.  Thirty years later, the  nexus of intelligence has shifted from algorithms to data as evidenced by the rise of Artificial Intelligence in the form of Large Language Models (LLMs).  In this talk, I will examine the role of LLMs in telecommunication networks to discover synergistic opportunities and challenges that arise in making the network more “clever”, if not completely “intelligent”.

Vijay K. Gurbani is the Chief Data Scientist at Vail Systems, Inc. where he established and manages the AI/ML team across the company doing innovative work in the use of AI in Natural Language Processing (NLP) and the systems area.  Vijay is also a research associate professor at the Illinois Institute of Technology and a research fellow at the University of Luxembourg.

Prior to Vail Systems, Vijay spent 21 years at Bell Laboratories where he was involved in the increasing use of machine learning algorithms and techniques to make sense of the data generated by 4G and 5G networks. At Bell Labs, Vijay’s work explored multimedia signaling protocols, especially Session Initiation Protocol (SIP) and the security and privacy aspects of multimedia protocols; the results of these efforts was a general-purpose SIP transaction layer library used to create SIP user agents, proxies, and registrars. This library was subsequently used as the basis for the Lucent Common SIP Stack (CSS), which is currently used in service provider networks of national and international companies and powers their Voice-over-IP solutions.

Vijay has authored or co-authored over 70 papers in peer reviewed journals, conferences and workshops, five books, 19 Internet Engineering Task Force (IETF) RFCs, and been granted 9 patents by the US Patent Office. 

CIVILIZING AI: Examining Emerging Capabilities and Mitigating Potential Risk of Foundation Models

While large language models (LLMs) and generative AI models such as GPT(s), LLaMa, Mistral, MidJourney, Stable Diffusion, and many others offer extraordinary benefits, they also pose substantial risks. These include the potential for intended misuse and unintended vulnerabilities, such as hallucinations and adversarial attacks. AI has reached a level where distinguishing AI-generated content, whether in text, images, or videos, has become notably challenging, a phenomenon we term “eloquence,” aka “usability” characteristics. Conversely, AI models’ troubling proliferation of hallucinations and adversarial attacks raises credibility concerns, referred to as “adversity” characteristics. In this presentation, I will explain three measures we recently introduced: i) AI Detectability Index (ADI), ii) the Hallucination Vulnerability Index (HVI), iii) the Adversarial Attack Vulnerability Index (AAVI). “CIVILIZING AI” embodies a delicate balance between the machine’s eloquence (ADI) and its propensity for adversarial behaviors (HVI and AAVI), to uphold constitutional principles. “CIVILIZING AI” is a broad and ambitious goal that demands progress on various fronts. I often describe it using an impossible pentangle, where each angle represents a specific challenge requiring research attention. These challenges include reducing model size, extending context window, improving watermarking techniques, mitigating hallucinations and adversarial attacks, optimizing alignment/policy techniques, enhancing human interpretability, and model editing. I have been actively working towards civilizing AI systems encompass all the aspects mentioned above and aim to continue this quest in the coming years. The goal is to create AI systems that adhere to evolving AI policies.

Dr. Das is a Research Associate Professor at The Artificial Intelligence Institute (AIISC) at the University of South Carolina, USA. Previously, he played a key role in founding Wipro Labs in Bangalore, India, where he was instrumental in its establishment from the ground up. He continues to be affiliated with Wipro Labs as an Advisory Scientist. Additionally, Dr. Das holds an adjunct faculty position at IIT Patna and has previously worked for Samsung Research, India. Dr. Das’s research interests lie at the intersection of three broad areas: human language, cognition/mind, and artificial intelligence, with a particular focus on Natural Language Processing (NLP). Over the past two decades, he has made significant contributions to the field of language technologies, authoring over 120 papers that cover a wide range of topics. In recent years, his research has expanded to include social computing and multimodal AI, addressing complex challenges such as multimodal misinformation and disinformation. His present focus is on the domain of CIVILIZING AI / CONSTITUTIONAL AI, which involves taking proactive measures to mitigate hallucinations and manage other associated risks in AI systems.