Top 10 Predictions for Artificial Intelligence in 2019

Available at What's Next in Tech: Teradata's Experts Weigh in on 2019 Predictions

As we wrap up 2018, our Emerging Practices team reflects on our experiences in working with our Top 500 Global customers to drive machine and deep learning capabilities. In looking forward to 2019, we share our perspective on where AI trends will be going for the upcoming year -

• Escalating Financial Crime - Synthetic Identify Fraud will continue to be a significant concern for a majority of companies involved in electronic payment. Many will have experimented using machine learning and several will have demonstrated improved performance over rules-based systems. Many of those companies will also have recognized the need to integrate information across different lines of business (e.g., HELOC, small business credit card, consumer credit) to enable early detection of fraud patterns using deep learning to better detect complex signals. Virtually all of companies will be challenged to deploy their new models due to operational and compliance issues dealing unless they can solve their model lifecycle and model risk management strategies. 
• Auto Labeling – For supervised learning, large sets of human annotated data is needed to train a deep learning model that performs a particular task. A fundamental challenge that the Enterprise faces today in their AI journey is the creation of customized high-quality human annotated data. This process is slow, repetitive, may involve subject matter experts, and at times need to be redone. For enterprises, this is a significant upfront investment with a big risk and big costs. In 2019, we will see a trend towards AI powered tools that assist humans in the creation of high-quality annotated data through auto-labeling techniques. AI involvement at early stages of the journey will reduce cost, risk and help create efficiency, these will play a big role in fueling AI adoption at enterprises.
•  Reinventing Retail - Brick-And-mortar retail businesses are turning their attention to AI to significantly improve customer experience, profitability and remain competitive. In 2019, we will see emergence of new data sources (surveillance cameras, on-the-shelf-cameras, robots) and AI models for inventory management, better customer retail experiences, targeted marketing, and adding new capabilities such as self-checkout. The key challenge, however, is to develop and scale AI operations to thousands of retail-stores that differ in planograms, camera models, and network infrastructure capabilities. 
• Robust context-aware models – Today, machine learning models are trained with very narrow tasks in mind, with increasing specialization. The need for more generalized inference will lead to more models that perform joint estimation of disparate output types, replacing chains of specialized models that each perform a single task. For example, instead of creating separate models for detection, object tracking, motion forecasting, and motion planning, executed in sequence, a single model is created that performs all these tasks jointly. These models can benefit from internally reusing computations and sharing high-level features. In 2019, we will also see new system architectures that provide "context" to individual AI models, by arranging these models hierarchically and/or connecting them at different temporal or spatial scales.
•  Data Minimization - Data valuation strategies will become increasing important. More data is better, right? Not always. Organizations are realizing that it's time to be more selective when it comes to data and more is not always better. In the future, we will see extra effort spent into discovering what is the true value of data to draft data minimization policies especially with the progress we see in making AI/Deep learning work with fewer data.
•  Reinforcement Learning - To date there have been very few examples of applying reinforcement learning to enterprise problems like recommending the best offer to a customer or supply chain optimization. These examples, amongst others, can be complex problems for machine learning models because they may have multiple potential factors to optimize and involve a series of events leading to a decision or business outcome. Both of these characteristics make them well suited for reinforcement learning and the opportunity to drastically outperform current approaches. 2019 will have breakthroughs in the Enterprise for RL, with a focus on using off policy learning to overcome the challenge of not having a real-world environment to train the model like you would a robot in a lab.
•  Industrial Inspection - Current smart camera solutions for manufacturing provide a generic software tool kit for product quality inspection that are intricate to tailor to specific requirements laid out by manufacturers. These black box smart camera software solutions limit manufacturers ability to combine features derived from image analytics with operational data that can improve early detection of quality issues. Furthermore, the software tool kits do not use state-of-the art analytical techniques. Consequently, such generic solutions do not cater to the wide variability in requirements resulting in low performances in detecting quality issues. In 2019, Manufacturers will seek a plug-and-play deployment of customized AI models that can provide high value at low cost/risk. Manufactures will need an Analytic Ops framework for ongoing monitoring, retraining and redeployment of models to continuously improve, which is again outside the purview of smart camera providers.
•  Edge AI - With more tools being built to make deep learning models smaller and more energy efficient without sacrificing model performance, edge computing for AI will be more adapted and will improve human/AI interaction. This will lead to more robust applications in retail and manufacturing to drive better understanding of what is happening within the physical proximity of their enterprise.
•  Model Risk Management - As the number of models deployed by enterprises grows, the need to manage model safety and model stability becomes increasingly clear. In Financial Services, Model Risk Management (MRM) is a well-known discipline to ensure the validity of models used in key processes such as credit underwriting. In 2019, other industries will realize the need to take a similar approach in order to realize the full value of their analytical models and data while managing risk exposure to their organization. Beyond model management, MRM provides data pipeline lineage, model governance, clear workflow for promoting models, reproducibility, stress testing, regulatory compliance, model performance monitoring, outlier detection and data monitoring. 
•  Humanized Digital Assistants - We have seen in 2018 big breakthroughs when it comes to AI assistants, synthesized human-like voice, and improved personalized dialog agents for the consumer. AI assistants will become more integrated and pervasive in many businesses with new opportunities for enterprise engagement in creative ways. Next-gen Business Intelligence will incorporate voice-based interfaces to support executive dashboarding and what-if scenarios.