AI Taking Action – Emergence of Decision Making and Generative Capabilities

My previous article explored the evolution of AI capabilities from a recognition-based competence to context-rich understanding. In alignment with this development, we can also expect that the nature of High Performance Computing (HPC) analytics will transform into higher value, earning our trust in making decisions and charting a preferable course of action. A ladder of increasingly-competent capabilities may give some clues as to the timeline and evolution “rungs” in both HPC and cloud-based analytics capabilities.

The most basic level is that of descriptive analytics.  Deep Learning (DL) is exceptionally effective in tasks involving classification, pattern recognition, and detection of anomalies. This level of capability requires two key elements: a) distinct categories, tags, or attributes, and b) the effective mapping of inputs to the set of attributes/tags.

The ladder rung at the next level is that of diagnostic analytics, where recognized states or patterns can be associated with causes or circumstances. This level builds on the descriptive level with additional requirements including: a) causality models tying mechanisms of operations and failure to expected results, and b) a mapping of highly-likely contributing factors (potential causes, circumstances) to a characterized result. Even though DL doesn’t address the mechanics of causality, it can do a very effective work in identifying correlation between potential causes and final results. With sufficient data and examples, this refined correlation can provide effective diagnostics.

These first two levels are based on extracting patterns from large amounts of data, and have found wide application today. The next three levels and are more transformative, going beyond pure data analysis toward producing new insights, anticipating and influencing future situations.

Still higher on the ladder is the level of predictive analytics, which is already provided, to an extent, by advanced HPC solutions. Here the objective shifts from “what is” to “what can be” in a projection of likely future states. In a climate analysis system, for example, this manifests as the hurricane path projection models a likely path for next phase(s) of the storm. This level of analytics builds on strong descriptive and diagnostic capabilities to precisely characterize the current state. Additionally, predictive analytics require new competencies including:

•  A mapping of current states and affecting factors to a set of likely future states.
• Selection of the most likely future state among many possible future states based on all available information.

For example, this NASA project is applying AI to predict an asteroids’ future trajectory, and to assess whether or not it will collide with Earth

The level more advanced than predictive is prescriptive analytics, which produces customized recommendations or instructions based on the insights derived from the data. Prime examples for this are recommendation systems (such as systems producing personalized recommendations for books, movies or purchases), as well as the “proposed routes” feature of every navigation system. This level builds on top of predictive analytics with added capabilities including:

• Associating value (desirability, benefit) levels with various results, thereby adding the perspective of particular beneficiaries – moving toward a subjective view of the recipient of value. Any particular recommendation or end result might have significantly different value depending on the recipient.
• Selecting a recommendation from a set of possibilities based on a) materialization likelihood, b) the level of benefit that can be attained if successful, and c) the risk or undesired penalty caused if perused unsuccessfully. Amazon, Google, Alibaba and many others already deploy in-application prescriptive systems, and they are working on rapidly increasing their effectiveness and user-specific fit. Statistical capabilities, along with adaptive Machine Learning, are required to improve creation and sorting of possible paths to particular beneficiaries.

The last and top level on our ladder can be found with generative analytics. At this level of value and sophistication, the system will employ various levels of autonomy to create sequences of meaningful actions that will interact and change the outcome. It can do this while pursuing optimal value for the designated beneficiaries. This level must be based on strong prescriptive capabilities derived from the level below, where assessing and grading possible courses of action in relation to their value / risks / likelihood is key. In continuation of the previous examples on climate and storm analysis, an autonomous generative system might guide the traffic lights to maximize the flow away from projected risk areas. Note that in the transition from prescriptive to generative systems, the human is taken out of the loop and is no longer the accountable decision maker. Therefore, generative systems pose the following added characteristics:

• More complete view of the potential consequences of the various choices: while in a recommender system, the deciding human can factor in additional aspects that are not comprehended by the system, a system that makes decisions and generates outcomes needs both complete and relevant information to make valid decisions.
• Added confidence in the validity of the recommendations: as systems are allowed to make the transition from making recommendations to exercising concrete meaningful steps, they need to be trustworthy of making such choices.
• Model of addressing failure: all systems are fallible. For a variety of reasons, it is certain that automated systems will fail and incur undesired consequences. Such systems need to minimize the occurrence and cost of failures, to entirely avoid irreparable damage, and to have an accountability model for addressing missed results. The cost of failure obviously varies widely between cases of automated book purchasing, smart city traffic light grid hiccups, or automatic pilot on a plane.

In addition to more autonomous decision making and command sequences, generative AI is also enabling the creation of intelligible – even sometimes creative – natural communications such as images or verbiage.  The most exciting innovation in this domain is Generative Adversarial Networks (GANs), which already demonstrate remarkable results in generating synthesized images and are expanding to other output forms. For example, at the 2018 Sundance Film Festival, filmmakers used generative ML systems to create deeper, more meaningful human-like animations.

In the coming 3 – 5 years, there will likely be a massive shift up this ladder, with prescriptive/recommendation systems taking a key role in many industry sectors, and multiple generative solutions deployed with increasing degree of empowerment and autonomous control over high impact systems. AI will play a central role in this transition by adding strength and accuracy to the lower rungs of the ladder, and crucial learning and target-based customization to the upper levels. The introduction of powerful prescriptive and generative systems will have very large implications to many fields – from science and medicine, to financial systems and smart spaces, as well as policy making.

Thanks for your interest,

Gadi

You might find it interesting to review the previous major transition described in this series:

Key Transitions: Broadening adoption of Deep Learning in HPC

Toward truly intelligent AI: From ‘Recognition’ to ‘Understanding’