Mark Skilton    Copyright 2019  ©

  • White Twitter Icon
  • White Facebook Icon
  • White LinkedIn Icon
  • White YouTube Icon
  • White LinkedIn Icon

How to test if robots have consciousness

 

I know IIT argues that for a system to be conscious, it must integrate information in a way that makes the whole contain more than the sum of the parts, and it is tested by measuring ‘phi’.

Can this be applied to machines?

 

The theories in this field are still very young, most have been invented in the last 5 years.  IIT is a leading theory created by Guidio Tonoini and developed and debated by many others that measures the degree of interconnectedness of an information network (the value called phi).  In the human brain case the 120 billion neurons and 100 trillion interconnections or high magnitudes.  But other species from insects to microscopic size have much small brains but are still considered “alive” even if this is a much reduced “phi” of semi-consciousness  or unconscious level of awareness. Plants are considered “alive”  but interact with the physical environment differently so the point is that measuring “awareness” and “living” as being conscious  of their environment have may different types “consciousness”.   Arguments against IIT and phi s a measure of consciousness are that it treats it as a functional network feature when the types of quality of experience called “quali” may have other  features of existence that just measuring the physical network connections is only part of the level of being “self-aware”.  

 

Yes IIT and phi can measure in the case of robots and AI programs the level of network complexity of interconnections. In the case of voice recognition or complex simulations of weather forecasting or stock market algorithms it can define complexity. Looking at  walking robotics at Google these have complex feedback sensor actuator systems that are “aware” of their physical environment. But in both cases they are a extremely limited subset of living experience. To that extent phi must be very low  and self-awareness is a long way off to bring all of living experience into a connected robot or android.  

 

Do you have any examples of IIT being tested on machines?

 

At a basic level 20 to 30 % of manual tasks could be replaced by a level of automation between sensors and actuators and to a  connected “Brain” hub of automatic AI  machine learning today. In research I have done at PA Consulting across several industry  case studies in banking, manufacturing, health care and retail using a similar IIT phi to measure the level of effective information that can be automated from a network of collected data.  But again this is a level low Phi because of the experience is a highly focused in task specific automation.   

 

Most academic research has centred on philosophical and theoretical analysis of simulations, the recent work by Max Tegmark as a quick guide to measure thickness of neuron network connections with for example a MRI scanners.  But this is for biological systems and not machine based.   But this again very difficult  to measure in defining the quality of experience , how to know which neurons create consciousness is vastly complex in a living network as a human cerebral cortex with 25 billion neurons and trillions of interconnections

 

Deep mind and other neural network research projects are measuring computing complexity similar to IIT but this is not the level of consciousness but very specific “super human” computing power of gaming and simulation. From Watson to complex simulations this is measuring speed and data computing power only.

 

In future, with quantum computing for example, do you think that computers might test positive for consciousness using IIT?

 

Not at present , Max Tegmark argued that quantum effects don’t work at ”room temperature”.  But in the future is the number of qubits are scaled to thousands or more then the value of phi could be very large  but you can see “it depends” on the external sensors and quality of experience such a quantum computer could use to become truly ”aware” or “self-aware”.  

 

What do you think about the work that Max Tegmark, from MIT, has done about approximating phi?

 

Recent work by Max Tegmark as a quick guide to measure thickness of neuron network connections with for example a MRI scanners.  But this is for biological systems and not machine based But this again very difficult  to measure in defining the quality of experience , how to know which neurons create consciousness is vastly complex in a living network as a human cerebral cortex with 25 billion neurons and trillions of interconnections.