What is synthetic? What does “human-level” mean? Because of this vagueness, different researchers tend to define General Intelligence in the manner that best benefits their own research: biological cognitive architecture, computational intelligence, etc. And without a clear definition and common benchmarks researchers go off in their own directions and the field advances slowly. Perhaps the approach we agree with is that stated by Nils Nilsson who said in effect, forget about building some synthetic human, just build an automated system that does the specific tasks humans currently do for pay and take it from there. Andrew Yang seems to think we are already there!
Most importantly, the ability to generalize experiences to novel situations is what distinguishes AGI from other forms of AI. Many of us are rightly impressed by Alpha Go which defeated the world’s best Go player. And recently Alpha Zero bested humans at multi-player role playing games causing some consternation in the gamer community. But none of these superior skills can be generalized as of yet. As with Alpha Zero or some autonomous car research, the combination of deep reinforcement learning, multi-agent learning, and imitation learning has made significant advances in the use of computers to solve problems. It makes us excited to see this type computing power and it might have some overlap with the work we need to do to make a generalized AGI agent that could, on its own, adapt its operations flexibly to a novel situation, but we have yet to create the additional architectural and dynamical principles that are required to make this type of AGI. Machine learning has helped us to build a variety of narrowly specialized systems but not a system of systems.
If we pull back further, there have been multiple attempts to create a rough architecture of the mind; both the overall working of the human (and other) brain and linguistic systems and perceptive sub-systems. All are beyond the scope of this blog but get in touch with us for an introductory list if you want further reading. Suffice it to say, like Mr. Nilsson, we tend towards the practical rather than the theoretical side of AGI. And to that end we can use a road map to help us focus on what is important. This map is from over a decade ago but is still relevant. In 2009 the AGI Roadmap Workshop (Adams et al., 2012) posited a couple of broad areas we list below. This is important because each of these broad areas of research are being attacked piecemeal by researchers around the world in an attempt to build new systems. Most will fail, but some will achieve something brilliant. And all most receive focus and funding of some sort so we want to reproduce the list here with some examples beside each system. The astute reader will be able to fill in the blanks we leave unfilled.
Perception –The senses a human possesses and is embodied in systems such as computer vision, background noise reduction in voice activated software and soft fruit picking robots.
Actuation – The manipulation of the environment as seen in semi-autonomous warehouse robots.
Memory – Currently being advanced in neural networks to improve prediction in linguistic models.
Learning – Reinforcement learning such as in Alpha Go and “self” teaching chatbots.
Reasoning– Many types of logical deductions which are being introduced into robots such as washing machines operating under a narrow set of choices.
Planning – Social, strategic and physical – what is the optimal route for my car to take or the best social media strategy to achieve my goals?
Motivation – Most difficult to reproduce but generally seen as a sub-set of reinforcement learning.
Attention -Visual and behavioural attention systems. Augmented Reality and Dynamic Creative Optimization are simplistic versions of attention systems that over time will develop perhaps to be highly individualistic and adapted to the viewing environment.
Emotion – Expressing and perceiving emotions are difficult and perhaps too subtle for our current tools to do accurately despite the hype around some systems claiming to “know” what people are thinking. A better area of research is in linguistics and how language expresses emotion. Sentiment analysis is a first step on this path.
Modelling Self/Others -Self Awareness and empathy are so crucial to human identity that progress remains in the realm of theory at this juncture.
Social Interaction – Social relationship are very mammalian in nature and most AI is focused on increasing the number of connections to facilitate social relationships rather than mediating social relationships. This too will remain outside most practical applications for some time.
Communication – Gestural, pictorial and verbal understanding. Perhaps the most exciting area of AI these days. The combination of loose “rules” such as grammar and the dynamic nature of the use of tokens to represent abstract ideas is a tantalizing and rewarding area for NLP and NLU researchers.
Quantitative – Perhaps the easiest of all the AI task as it involves coding the hard rules of maths and statistics to drive model development. Generally the fine tuning of parameters to minimize some objective function.
Building/Creation – This will be the final frontier and perhaps unsurmountable for AGI. Our next blog on ASI will focus on this definition.
The list above is but one example of our attempt to define the attributes of Intelligence. It is not perfect but the science advances a step at a time. And although the advancement towards a fully robotic future is still far, applied AI has yielded amazing advances for workplaces around the world and will continue to be developed in order to relieve us of the most repetitive tasks and allow us to do more of what we love.