Cloud AI

Many customers today get stuck on the very first step in the ML process because of the lack of large/ high quality enough data set to be able to apply specific AI models (e.g., deep learning) with decent accuracy level. Research areas such as Transfer Learning, Data Synthesis can help reduce the amount of data needed or create a bigger data set using other means. Data Enrichment and Augmentation can automate the process to improve data quality. Novel learning architectures for active learning using combinations of semi-supervised learning, reinforcement learning, and generative models can provide transformational opportunities for faster more efficient learning paradigms. Automated self-learning paradigm exploits augmented reality in a synthetic world setting by combining simulated training with real data in order to substantially generalize to the diverse real-world settings.

Today much of the ML model development and fine-tuning still require a substantial amount of manual work. But data and ML experts are rare and highly demanded resources. This makes a lot of potential AI customers shy away even if they have plenty of raw data. Research areas such as AutoML and Learning to Learn (L2L) that aim to automate data processing and ML model development/customization will offer AI solutions to a broader set of customers without significant data/ML expertise investment requirement. Reinforcement learning that enables machine to “automate” learning directly from sensory inputs via experimentation will remove the need for knowledge in specific ML models.

The goal of this research area is to remove dependencies on network/ computing needs for Enterprise AI customers. Currently Cloud AI solutions mostly focus on markets/ business environments that has great connectivities. Research areas such as Federated Learning can help create great value in both commercial (e.g., in the field medical device) and social perspective (e.g., developing country with limited internet bandwidth) to enable AI solutions in environments with limited computing and network connectivities.

Large Scale Cloud ML enables analysis and processing of continuous, high volume, open ended data streams to enable customers to make the most out of their data asset. Online retraining closes the ML feedback loop for continuous "in-line" training during customer business operations and improves performance in real-world applications. Lifelong Learning allows continuous learning and accumulation of knowledge from past experience, and uses it to help future learning.

Human-in-the-loop learning will further improve AI model accuracy by embedding human input in the AI model training and deployment cycles. Research in Human-Machine interaction and collaboration provides a tighter and smoother integration of human/ machine collaborations that are more interactive, effective, real-time and mobile. Explainable/Interpretable AI systems that explains its decisions and actions to human users enables users to understand and appropriately trust AI partners in human-in-the-loop systems. Developing fundamental cognitive ML tools for social-cognitive systems and socially-aware IoT products (e.g. digital assistants) that merges cognitive with social abilities (e.g. social norms) helps AI agents to navigate the social world and interact more effectively with humans and other autonomous agents. Like other basic AI cognitive and problem-solving abilities, social cognition is associated with the integrity of interrelated neural systems for accurate perception and interpretation of the behaviors of other autonomous agents, and the effective emotional and behavioral response to those behaviors. Value alignments between AI agents and human provides safety and appropriate trust for optimized autonomous operation.

Generate deeper understanding and insights from data collected from various environments, or from learned models to understand the implied events, behavior, and intent.

Today most of the AI-based dialog is based on simple rule-based mapping from topics to responses, which limits the flexibility of the conversation and leads to confusion or mistakes when the dialog diverges from the well-documented topics or includes deeper-level implicit intentions. AI technologies such as knowledge map, domain specific translation, tonal analysis can enable more “natural” conversations that takes into account the specific context, roles, goals and emotions, and provide a better experience in person-person, person-enterprise, enterprise-enterprise communications. Such technology can enable solutions such as Goal-driven deep dialog, automated booking based on personal preferences in travel (e.g., vacation) and retail (e.g., gifts) industry, Role-play conversations such as automated consulting (e.g., remote medical advice) and Q&A (e.g., shopping assistance), and more accurate translation with domain knowledge for travel, legal, finance, healthcare and more.

The proliferation of data from IoT devices creates new transformational opportunities for AI-based applications, which requires effective balance of on-device and cloud-based intelligence (e.g. cloud robotics, smart cities, etc.). AI agents require to communicate at the appropriate levels of abstractions and autonomous operation.

Provide an agile mechanism for researchers and innovators to introduce and test their research output in order to quickly assess performance and acquire user feedback by leveraging cloud infrastructure and platform tools.

AI is helping many enterprise businesses to get closer to their customers through intelligent interactions between humans and machines. Research in deeper dialogue systems creates a transformational opportunity for businesses to learn from direct interactions with their customers, and to generate insights that guide critical business processes.

Despite many recent success, classic ML algorithms have several limitations that make them difficult to use in many real-life enterprise applications. For example, convergence guarantee of existing RL algorithms is brittle, data requirement for ML algorithms is high, and deep networks are sensitive to noisy labels that are often present in training data for real-world applications. Our research focus is to create reliable, high quality and easy-to-use algorithmic tools, including deep learning and reinforcement learning, that are robust and data-efficient for cloud customers.

Our mission in AI+ is to advance the state of the art in AI that is necessary to improve many critical services in our society driven by enterprise needs. Key focus areas include Healthcare and Education. In healthcare, our focus is on long-term research for creating assistive tools that can help medical professionals make more informed critical decisions. For example, one of our ongoing projects include Thoracic Disease Identification and Localization, which addresses the label-scarcity issue in disease localization. Other projects include synthesizing tumors & tubes in X-rays/CT scans, and domain transfer across image capture parameter with GANs. The AI technology and tools under development in these projects will reduce the need for large volumes of labeled training data and variations that are typically necessary for personalized healthcare.