We are excited to bring Transform 2022 back in-person July 19 and virtually July 20 – August 3. Join AI and data leaders for insightful talks and exciting networking opportunities. Learn more about Transform 2022
Our abilities to invent and use tools are critical to human evolution. Computers as tools have certainly advanced humanity since their inception. As computing technologies advance, human-machine relationships have also been evolving. Initially only computer developers or programmers can operate computers by giving machine (programming) instructions that computers can understand and follow. With the development of graphical user interfaces (GUI), the masses can now operate computers with no code. The human-machine relationships however remain to be operator-machine relationships, during which humans must tell machines precisely what to do.
With the rise of artificial intelligence (AI) — computers with certain human skills — the human-machine relationships may be completely redefined. For example, computers with human visual perceptual skills can augment security personnel to rapidly recognize objects in mountains of surveillance images or computers with human language skills can augment paralegals to summarize large amounts of text documents. However, teaching machines human skills is a complex, time-consuming process, requiring deep expertise and programming skills, not to mention the efforts for collecting, cleaning, and annotating large amounts of training data needed to train machines with desired skills.
Just like the no-code, GUI-driven computer operations, what if humans, the security personnel and paralegals alike, can teach machines human skills with no code? Like in the movie Her, what if we can adopt a turnkey AI assistant with built-in human skills and easily customize it with no code to meet our specific needs? This vision of no-code, reusable AI will certainly elevate our current operator-machine relationships to the supervisor-assistant relationships. Not only will the new relationships enable us humans to be augmented by AI instead of being replaced by it, but the no-code nature will also democratize human augmentation.
1. AI by human skills
Depending on the tasks to be achieved, AI systems are trained to possess different human skills. Figure 1 lists example AI systems by human skills. Certain AI systems use a single type of human skills, such as human visual perception or linguistic skills, to perform a specific task, such as object identification or sentiment analysis. In contrast, more complex AI systems use multiple human skills together to achieve complex tasks. For example, a self-driving car must use multiple human skills, such as human visual perception and decision-making skills, to achieve its driving goals. Likewise, a conversational AI assistant must employ multiple human skills, such as communication skills or certain human soft skills (e.g., active listening), to accomplish its tasks.
2. Multi-level reusable AI
No matter whether an AI system requires a single or multiple human skills to function, creating an AI system from scratch is always difficult and requires much expertise and resources. Just like building a car, instead of building it completely from scratch with raw materials, it would be much easier and quicker if we could quickly customize and piece together pre-built parts and systems, such as the engine, the wheels and the brakes.
While there are many no-code, reusable AI systems, it is most challenging to enable the no-code, reuse of a complex AI system, such as a conversational AI system, because of the technology complexity involved and the requirement of multi-level reuses. Figure 2 shows an example 3-layer architecture in support of a cognitive AI assistant, a new generation of AI assistants with multiple advanced human skills including soft skills.
Reusing general-purpose AI models
As shown in Figure 2, the bottom layer is a set of general-purpose machine learning models that any AI system relies on. For example, data-driven neural (deep) learning models, such as BERT and GPT-3, typically are pre-trained on large amounts of public data like Wikipedia. They can be reused across AI applications to process natural language expressions. General-purpose AI models however are inadequate to power a cognitive AI assistant. For example, general-purpose models trained on Wikipedia typically cannot handle nuanced conversational communications, such as managing a conversation or inferring a user’s needs from a conversation.
Reusing specialty AI engines
To power an AI assistant with human soft skills, specialty AI engines (the middle layer) are needed. For example, the active listening engine shown in Figure 2 enables an AI assistant to understand the focus of attention in a conversation and gives it memory so it can correctly interpret a user’s input including incomplete and ambiguous expressions in context as the examples shown in Figure 3.
Likewise, specialty AI engines like reading between the lines and conversation communication engines power an AI assistant with additional human skills. For example, reading between the lines enables AI assistants to analyze a user’s input during a conversation and automatically infer the user’s unique characteristics (Figure 4). The conversation-specific communication engine enables AI assistants better interpret user expressions during a conversation, such as identifying whether a user input is a question or reflective statement, which warrants different AI responses.
With careful design and implementation, all the specialty AI engines can be made reusable. For example, the active listening conversation engine can be pre-trained with conversation data to detect diverse conversation contexts (e.g., a user is giving an excuse or asking a clarification question) and pre-built with an optimization logic that always tries to balance user experience and task completion when handling user interruptions to guide a conversation.
Reusing whole AI assistants
In addition to reusing individual AI components/skills, the ultimate goal is to reuse a whole AI solution. In the context of building AI assistants, it is to reuse a whole AI assistant based on AI assistant templates with pre-defined workflows and a pertinent knowledge base (the top layer of Figure 2). For example, an AI Recruiting Assistant template includes a set of job interview questions and a knowledge base for answering job-related FAQs. Similarly, an AI Learning Assistant template outlines a workflow, such as checking the learning status of a student and delivering learning instructions or reminders. Such a template can be directly reused to create a turnkey AI assistant or can be quickly customized to suit specific needs as shown below.
3. Reusable AI enabling no-code AI
Since every AI solution typically requires certain customizations, reusable AI enables no-code AI customizations. Below are several examples.
No-code customization of AI assistant templates
Assume that an HR recruiter wishes to create a custom AI Recruiting Assistant based on an existing AI template. Just like using PowerPoint or Excel, the recruiter will use a GUI to customize the interview questions (Figure 5) and job-related FAQs. The no-code customization greatly simplifies the creation of a powerful, end-to-end AI solution especially for non-IT professionals.
Continuing the above example, assuming that the recruiter wants the AI assistant to ask job applicants a question “What do you like the best in your current job?”. If an applicant’s response is something similar to “interacting with customers“, the recruiter wants the AI to ask a follow-up question “Could you give me an example that you enjoyed interacting with your customer?” Since the pre-built AI template does not handle this specific case, the recruiter would need to customize the AI communication. Figure 6 shows how such customization could be done with no coding.
4. No-Code, reusable AI defines supervisor-assistant relationships
No-code, reusable AI enables everyone, including non-IT professionals, to create their own custom AI solutions (assistants). An AI assistant only needs to be instructed what to do (e.g., asking users a set of questions) and then performs the tasks automatically (e.g., how to handle user interruptions). This transforms the traditional operator-machine relationships into supervisor-machine relationships. When humans must program/code a machine to teach the machines, humans act in the role of operators/developers of machines. While humans provide machines with high-level, no-code instructions, such as outlining the tasks and teaching new knowledge, humans now become the supervisors of machines. This new relationship enables humans to do more with machines’ help.
5. Future directions of no-code, reusable AI
No-code, reusable AI democratizes the creation and adoption of powerful AI solutions without requiring scarce AI talents or costly IT resources. Furthermore, no-code, reusable AI elevates the human-machine relationships, enabling everyone to be augmented by machine powers. To make no-code, reusable AI the main paradigm for developing and adopting AI solutions, advances must also be made in several areas.
The first area is to make reusable AI components/systems explainable. To help non-IT personnel reuse pre-trained or pre-built AI components and solutions, it is critical to unbox the “black box” and explain what is inside each component or solution, both pros and cons. The explainable reusable AI not only helps humans better understand and leverage existing AI components/systems and also helps avoid potential AI pitfalls. For example, it would be helpful for an HR recruiter to understand how personal insights are inferred before s/he uses such AI power to infer applicants’ insights.
Automatic AI Debugging
The second area would be the support of automatic AI debugging. As AI solutions become more complex and sophisticated, it is difficult to manually examine potential AI behavior under diverse and complex circumstances. Non-IT users will especially need help in assessing an AI solution (e.g., an AI assistant) and improving it before formally deploying it. Although there is some initial research on profiling AI assistants, much more is needed going forward.
The third area would be ensuring the responsible uses of AI, especially with the democratization of AI. For example, if someone can simply reuse an AI functional unit to elicit sensitive information from users, how and who can protect the users and their sensitive information? In addition to measuring typical AI performance such as accuracy and robustness, new measures and usage guidelines will be needed to ensure the creation and deployment of trustworthy and safe AI solutions.
Michelle Zhou, Ph.D. is a cofounder and CEO of Juji, Inc.