Perhaps no technological shift has as much potential to transform our perception of computing power as machine learning and artificial intelligence. 

Historically, computers have been dutiful followers of the instructions provided to them, executing them at ever-increasing speeds. With machine learning (ML), however, computers can refine results by improving their compute processes automatically. They do this by using models to analyze massive amounts of data, 'learning' more every time they get a 'correct' answer.

ML and AI promise to address vexing challenges that have eluded traditional computing approaches. Today, for example, AI-based systems can be used to block out photo backgrounds or noise on calls, track items picked up at a store to avoid checkout lines, and translate languages in real time, even American Sign Language. AI can also identify anomalies and vulnerabilities in source code far more quickly than human reviewers. Tomorrow, these systems will guide vehicles as they become increasingly autonomous, and they'll accelerate drug discovery by running statistical models and treatment scenarios faster than any human study. 

However, as with many technological advances, there are risks and the potential for abuse. From the reality of policy discourse to the fantasy of sci-fi, worries have ranged from AI-based processes displacing humans in many kinds of jobs to fictional AI-based systems that can run amok. There are also concerns about bias being coded into models inadvertently. 

Intel believes strongly in the potential for machine learning to help humanity in a wide range of tasks that span our personal and professional lives. One way that these advances can provide assistance is by analyzing and acting on the extensive data sets generated by our intentional and reactive activities and movements. These include such domains as assistive technologies, education, and manufacturing.

Adaptive communication

Carrying on work that began as a collaboration with the late Dr. Stephen Hawking in the field of assistive technology, Intel continues to develop the Assistive Context-Aware Toolkit (ACAT). ACAT is free, open-source software kit that developers can use to build natural communication solutions for individuals with severe verbal and motor disabilities. ACAT leverages the input of sensors that can detect facial gestures, eye movements, and brain signals, putting them all together to output computer interface commands and verbal language.

Lama Nachman, Intel Fellow and Director of its Human and AI System Research Lab, explained: "We are a multi-disciplinary team that spans ethnography, design, and AI, and we are focused on creating technologies that can understand people, assist them in everyday life, and amplify their potential. With our ethnographic research, we try to understand unmet needs, and develop technologies that can satisfy these needs through sensing and AI. Take the case of people with disabilities. We are creating open-source platforms that can take any minimal signal that the user is capable of providing and turn that into an ability to fully control their PCs, unlocking their ability to engage with the world."  

Intel ACAT works with a range of sensors that detect eye or muscle movements, and even brain signals, translating them into computer inputs.

Blending the best of human and machine capabilities, ACAT incorporates feedback from live conversation partners to tune the experience, and it optimizes large communication data models to provide meaningful advances, including:

• Learning and adapting with minimal data and supervision
• Adding emotional inflection to speech generated by text
  
• Personalizing responses to the user's content and communication style
  

In a demo shared at Intel Innovation this year, ACAT was used together with language modeling to understand the context of conversation and anticipate a user's intent, suggesting keywords and phrases to facilitate communication. This enables individuals to communicate more naturally and quickly, vs. typing one word at a time with their adaptive apparatus.

Personalized education

As educators adapt their teaching techniques to achieve the best outcomes with the knowledge that all students learn differently, machine learning can provide new levels of insight into student reactions and performance via adaptive learning. As with ACAT, interpreting the output of sensors is key to uncovering insights. In this case, the system analyzes cues like facial expression and head pose captured via cameras.

Lama Nachman, Intel Fellow and Director of its Human and AI System Research Lab

The solution 'understands' certain poses as conveying states such as satisfaction, boredom, confusion, and engagement. Based on this knowledge, it correlates educational material to on-task consistency and attention. These insights are presented in a teacher's dashboard, helping identify the average level of engagement or what parts of a lesson may have confused particular students or the class in general.

Adaptive learning has also been applied to a projection-based learning environment for younger students called Kid Space. "We are creating multimodal conversational agents that act as peer learners to help students engage and practice their learning in the early childhood context," Nachman said. "Since students need to learn in the physical environment, manipulate objects, learn with their whole bodies at this age, these AI systems will use multi-modal sensing to engage these children and guide them through the tasks, responding to not only their utterances, but their gestures and activities in the real environment."  

AI agents can process multimodal input -- based on factors such as students' facial expression or gaze, tone of voice, language, and gestures -- to drive technology-assisted learning while minimizing screen time. This has led to promising results, including:

• More than 99 percent of students stayed on-task during the learning sessions
• For 95 percent of the time, students demonstrated positive emotions, such as excitement and satisfaction
  
• More than half of the students were physically engaged (walking, jumping, and so on)
  

Safer manufacturing

Processing multimodal data generated by human movements and the environment can also increase productivity in a range of work settings, such as in manufacturing. Enhancing robots with greater AI capabilities can improve throughput, reduce cost, and increase safety. For example, computer vision-powered cameras inform systems that can predict human intent and offer ways to improve task performance, scan factory workflows in order to minimize dangerous processes, and provide metrics that can pinpoint bottlenecks or developing problems.

Nachman expanded on the thinking behind these developments:  "In manufacturing settings, many of the tasks people perform follow a procedure or a specification. The AI system will start by ingesting that spec, and then learn from watching people perform these tasks to recognize the actions. Then it can assist people as they perform these tasks by providing them relevant info or flagging when errors occur. Even more interestingly, since it's watching many different people perform these tasks, it can facilitate learning across people, such as bringing expert knowledge to a novice, or even accounting for errors. On the opposite side, as tasks change, users can help the AI system learn changes and adapt to new tasks and environments. It is a true human/AI collaboration."

Reducing bias

As we develop computers that can act on their own conclusions, and particularly as the implications of those decisions become more significant in our daily life, we must take measures to ensure that AI serves us fairly and equitably. Some algorithms, for example, may reflect the biases and deficiencies in the data sets used to train them. Intel is committed to ensuring that AI preserves the privacy of individuals and to sharing the results of its research transparently. To help minimize developer biases, Intel is building frameworks for rapid creation of new models and optimizing hardware and software to run models more efficiently. Both will accelerate iteration and help developers detect and remove biased inputs more quickly. 

While care must be taken as we progress down the path of AI, the destination is one in which computing can enrich human lives in unimagined ways. For more information on how Intel's AI leadership can serve us tomorrow, and how it can help your business today, visit https://www.intel.com/ai.