At AT&T, AI and machine learning are woven into customer interactions, our software-defined network and next-gen technologies.
AT&T AI Guiding Principles
By People, For People
We incorporate human oversight into AI. With people at the core, AI can enhance the workforce, expand capability and benefit society as a whole.
Accessible and Shared
We support open-source communities whenever appropriate to further access, collaboration, standardization and participation in industry discussion.
Secure and Ethical
We are grounded in ethics, safety, and values at every stage of AI, including our privacy principles and security safeguards.
AI Powered Video Metadata
AT&T creates and delivers the world’s best video entertainment content for our customers. Our research group applies the latest AI technologies to generate highly detailed descriptions of content to power enhanced content experiences and more relevant advertising.
Understanding Customer TV Viewing Preferences
Understanding customer viewing preferences is essential for AT&T’s content, distribution and media businesses. How to appropriately categorize content can be a challenge due to the “organic” approach that many suppliers of content metadata have taken – particularly to the assignment of genre and subgenre classification.
To enhance and systematize content tagging, we use modern Natural Language Processing and deep learning methods to create improved microgenres. This leads to a deeper understanding of viewers’ behaviors and preferences in accordance with our privacy policies – ultimately providing the viewers with more meaningful content.
Machine Learning-Based Media Planning
Making advertising more relevant to consumers is one of the major challenges facing the advertising industry today. The challenge is perhaps greatest for linear TV, which offers far less data for personalization compared to advanced TV services and the digital advertising space. AT&T Labs helps support Xandr's mission to “Make Advertising Matter” by translating detailed DirecTV viewing data in accordance with our privacy policies into focused media buy recommendations that reach an advertiser's target far better than traditional age/gender-based advertising.
The predictive algorithm driving this new service begins with an advertiser's target overlaid onto the DirecTV customer base. Motivated by research on engaged viewing, we create features representing a range of past viewing behaviors including tune-away patterns, program, genre and network preferences, consistency and regularity, frequency and duration. ML modeling at a massive scale produces a TV show or network list ranked by target reach, which in turn determines future media buy recommendations. In addition to offering superior audience targeting with linear TV, we can augment the recommendation with show-specific estimates of ad avoidance, further differentiating Xandr's value to ad planners.
AKG: Automated Keyword Generation for Metadata Enhancement
Large enterprises, such as AT&T, typically manage vast numbers of databases and datasets relating to disparate areas of the business such as finance, networking, and customer care. Metadata about all of this data contains different types of information such as database schema descriptions including column and table names. Searching this metadata for the right table that the user needs, however, is known to be challenging due to multiple reasons.
To address this challenge, we developed a machine learning based approach called Automated Keyword Generation (AKG). Given a corpus of partially annotated noisy data, AKG learns a mapping of abbreviations into the corresponding keywords. A single abbreviation might have multiple distinct meanings in different contexts, so our algorithm learns a smart contextual dictionary that can choose a meaning given the context.
Real-Time Visualization of Large Spatiotemporal Datasets Using Nanocubes®
Due to the ubiquity of GPS-enabled devices, geospatial datasets are increasing in both size and complexity. This creates the need for new techniques for processing and understanding such large datasets. Nanocubes®, which is a highly efficient implementation of the data cube operation developed here at AT&T Labs, helps reduce this complexity by enabling interactive visual exploration of spatial and temporal data all within a web browser running on a laptop. Although widely used within AT&T to analyze mobility, cybersecurity, and network data, Nanocubes has also been released as open source (nanocubes.net) for anyone to use and extend for their own geospatial data analyses.
Acumos AI Day
AT&T Labs Research hosted Acumos AI Day this past November. The event was well-attended and represented nearly a dozen different organizations. Participants came together to learn about the latest in AI and Machine Learning from some of the industry’s leading experts.
Learn more about the economic gains being made by AI now and in the future.
Our work in AI began in 1952 with Claude Shannon's invention of a mechanical mouse that learns how to move through a maze by 'memory'. Over the past 5 decades, we've built on his expertise, researching and developing numerous AI capabilities in areas of language translation and processing, speech recognition, computer speech, image recognition, fleet management, sentiment analysis and more.
Acumos is our open source platform for distributing and chaining different AI tools and microservices, giving developers and businesses unprecedented access to sophisticated capabilities.
Machine Learning for 5G
To bring 5G to life, we need to deploy hundreds of thousands of cell sites. Current manual processes require in-person site visits, but we’re using machine learning to create a “virtual world” that describes its environment – poles, buildings, building materials, foliage – to help operators determine where cell sites can be placed without requiring a site visit. This technology also helps us identify faults in our towers.
Machine Learning for Security
The sophistication and volume of cyber activity has been increasing dramatically. This requires a change in traditional cyber analysis approaches that use static signatures and manual analysis. With machine learning, automated analysis is used to build patterns of normal and abnormal activity based on subtle characteristics that escape the human eye. The increasing wealth of anomalies can be analyzed using machine learning to proactively detect emerging trends before the network is compromised.
Predictive Customer Care
Machine Learning algorithms help us anticipate when customers are experiencing a service issue, and allow us to potentially address the problem before customers reach out.
Content Personalization and Recommendation
We use advanced recommender system algorithms to provide our customers with the most personalized content and experience for our DirecTV and U-verse services.
Drone Video Analytics for Cell Network Maintenance
Maintaining AT&T’s cellular antennas — over 1 million in the US — can be very challenging. We use drones to inspect the antennas and deploy deep learning models to analyze the video feeds and determine if there is something that needs to be addressed — like a bird’s nest, weather damage, or unplugged cables. This not only saves significant manual effort and time, but improves worker safety.
Metadata Extraction for Video Content
Video is expected to take up over 80% of internet traffic by 2020. We understand that video is important to serve our customers better, optimize advertising, and recommend content. We use AI methods to extract metadata from video using facial and voice recognition, text analytics of close captions, and deep learning to find semantic meaning from video scenes.
Towards Identifying Impacted Users in Cellular Services
Shobha Venkataraman, Jia Wang
ACM SIGKDD, 2019
Geofences in the Sky: Herding Drones with Blockchains and 5G
Tamraparni Dasu, Yaron Kanza, Divesh Srivastava
ACM SIGSPATIAL, 2018
Information Market for Web Browsing: Design, Usability and Incremental Adoption
Arash Molavi Kakhki, Vijay Erramilli, Phillipa Gill, Augustin Chaintreau, Balachander Krishnamurthy
ACM SIGMETRICS Performance Evaluation Review, 2018
A Graph Database for a Virtualized Network Infrastructure
Pramod Jamkhedkar, Theodore Johnson, Yaron Kanza, Aman Shaikh, N.K. Shankaranarayanan, Vladislav Shkapenyuk
ACM SIGMOD/PODS International Conference on Management of Data, 2018
Interpretable Graph-Based Semi-Supervised Learning via Flows
Raif Rustamov, James Klosowski
AAAI Conference, 2018
A Stitch in Time - Autonomous Model Management via Reinforcement Learning
Elad Liebman, Eric Zavesky, Peter Stone
International Conference on Autonomous Agents and MultiAgent Systems, 2018
UCBoost: A Boosting Approach to Tame Complexity and Optimality for Stochastic Bandits
Fang Liu, SinongWang, Swapna Buccapatnam, Ness Shroff
Unchain Your Blockchain
Tamraparni Dasu Yaron Kanza; Divesh Srivastava
Symposium on Foundations and Applications of Blockchain, 2018
TV News Story Segmentation Using Deep Neural Network
Zhu Liu, Yuan Wang
IEEE International Conference on Multimedia and Expo (ICME), Industry Track, 2018
Geotagging IP Packets for Location-Aware Software-Defined Networking in the Presence of Virtual Network Functions
Tamraparni Dasu, Yaron Kanza, Divesh Srivastava
ACM SIGSPATIAL, 2017
Autonomous Model Management via Reinforcement Learning: Extended Abstract
Elad Liebman, Eric Zavesky, Peter Stone
International Conference on Autonomous Agents and Multiagent Systems, 2017
AESOP: Automatic Policy Learning for Predicting and Mitigating Network Service Impairments
Supratim Deb, Zihui Ge, Sastry Isukapalli, Sarat Puthenpura, Shobha Venkataraman, He Yan, Jennifer Yates
ACM KDD, 2017
LiveJack: Elastic CDNs-Edge Clouds Integration for Live Content Broadcasts
Shu Shi, Rittwik Jana, Bo Yan, Yong Liu, Haoqin He, Weizhe Yuan, Yang Xu, H. Jonathan Chao
ACM Multimedia, 2017
Deep Hashing: A Joint Approach for Image Signature Learning
Yadong Mu, Zhu Liu
Bridging Heterogeneous Domains with Parallel Transport for Vision and Multimedia Applications
UAI'16 Proceedings of the Thirty-Second Conference on Uncertainty in Artificial Intelligence, 2016
Cristian Borcea, Manoop Talasila, Reza Curtmola
Chapman and Hall/CRC, 2016
MP-DASH: Adaptive Video Streaming Over Preference-Aware Multipath
Bo Han, Feng Qian, Lusheng Ji, Vijay Gopalakrishnan
ACM CoNEXT, 2016, Best paper award
Joint Audio-Visual Processing for Video Copy Detection
Zhu Liu, Eric Zavesky, David Gibbon, Behzad Shahraray
Academic Press Library in Signal Processing, 2014 (book chapter)