Early Experience with Adaptive AI Hardware Shows Current Limitations

876
Enabling adaptive AI DEPOSITPHOTOS ENHANCED BY COGWORLD

By Lars Wood with Juan Miguel de Joya

The Fourth Industrial Revolution, first coined by Klaus Schwab in his 2015 article in Foreign Affairs, represents the revolutionary shift in how we as a society integrate technology into our day-to-day. We are seeing emerging technology breakthroughs in robotics, autonomous vehicles, energy storage, material sciences, and nanotechnology among other fields, transforming existing processes and generating dialogue on their implications.

Artificial Intelligence (AI) sits at the center of this discourse, and is poised to solve fundamental challenges and create value across all sectors. Unfortunately current narrow AI applications have limited scope due to inherent bias, and are subject to catastrophic forgetting where machine learning systems must erase their memories and completely retrain using new data. Another limitation is that fundamental AI algorithms have not radically changed for decades. It’s the inexpensive computational power that has changed. Consequently, the challenges that researchers faced in decades past are back, haunting AI systems deployed today in everything from smart phones to self-driving vehicles.

Recent attempts to move beyond narrow AI applications in industry have struggled to gain traction. ReThink Robotics, a leading startup founded by AI founding MIT researcher Dr. Rodney Brooks to create adaptive collaborative robots for industrial robotics, closed its doors in October 2018 and has since had its IP acquired by HAHN Group. In a retrospective published by The Robot Report, several contributing factors led to the shutdown. ReThink’s reliance on series elastic actuators compromised the precision and repeatability found in typical actuators in favor of safety, which likely led to efforts to compensate on hardware through software.

While the company utilized innovative machine control and machine vision technologies in iterating on their robots, the combination of mechanical motion of firmware at the heart of their products led to a narrow range of issues at varying quality. This made Baxter and Sawyer, ReThink’s flagship industrial robots, ill-suited for adaptive industrial use.

Other companies attempting to build adaptive robots, including Jibo, have met similar troubles. Touted as an interactive social robot with a personality, Jibo launched their eponymous robot in November 2017 with an emphasis on naturalistic human-computer interaction, but entered the market with more limited functionality than cheaper smart assistant speakers. The company has since closed down and transferred ownership of their IP to SQN Venture Partners in November 2018.

This trend is significant because robotics is a field that pushes all aspects of AI. Having automated technology that exhibits precision, repeatability and flexibility is highly useful for robotic assembly and packaging applications, creating quicker, lower cost, safer, and more accurate processes for supply chains. AI is impactful for complex real-time path determination and course correction, which is vital for transportation systems, automated vehicles and their respective manufacturing sectors. In advancing naturalistic human-computer interactions and creating parameters for meaningful engagement with data and technology, we can augment our experiences, perspectives and relationships significantly. This is the promise of the Fourth Industrial Revolution and, by extension, AI. There is no doubt that ReThink Robotics and Jibo have created pioneering work that explores the limits of what we can do with robotics. Consequently, their challenges to build adaptive robotic systems are at the forefront of the limitations of current AI technology.

In 1992 Dr. Hava Siegelmann and Dr. Eduardo Sontag published the Siegelmann-Sontag hypothesis. This mathematical treatise predicted that AI based on Dr. Alan Turing’s computational formalism, which defines the operation and function of all current digital and quantum computing systems, is incapable of implementing adaptive AI. These systems are fundamentally restricted from the self-organizing and problem solving flexibility that is fundamental to all analog brain neuron functionality and operation. The Siegelmann-Sontag thesis suggested a new super Turing analog recurrent neural network paradigm, which surpasses the capabilities of both current digital and future quantum computing machines.

Lars Wood is CEO And Co-Founder of Qai.ai LLC Super Turing Machines; Juan Miguel de Joya is a consultant expert on AI and Emerging Technologies at the United Nations International Telecommunications Union. 

Read the source article in Forbes.