Prior Postings in the Series
- Exploring Consciousness
- Who is “I”
- Talking to Myself, Who is Listening?
- The Epitome of Consciousness
Toward Self-Aware Robots
What would it even mean that a robot could be self-aware?
In my last several posts I have been exploring consciousness as a process produced by the brain. I have been examining the phenomenon of awareness as a basis for what we experience as consciousness. But I have also pointed out that consciousness is not just something experienced by humans; animals at different levels in the phylogenetic tree experience some forms of consciousness as well. And those closest to us in evolutionary terms even experience self consciousness. In this posting I want to go way out on a limb and declare I think I know how to produce self consciousness in a robot!
For many years now I have been exploring how it might be possible to emulate natural intelligence in an electronic computer. My approach is very different from that being taken by the classical artificial intelligence (AI) crowd where the thinking is along the lines of “finding the right algorithm for intelligence.” This thinking seems to come from the belief that our brains work like a computer and it is just a matter of finding the program that produces the same intelligence. Nor is my approach the same as the newer versions of machine learning based on statistical approaches, for example a Dynamic Bayesian Network, though there are similarities and some overlap.
My approach starts from the premise that evolution is a powerful “designer” in the sense that it has had so much time to fiddle with various ways to make animals smarter and the architecture of brains in real animals represents an optimal design. It also recognizes that biological designs tend to be conservative and that nothing in such a system is wasteful or unnecessary to the large-scale purpose. Brains must work the way they do because evolution discovered the designs that work best to produce overall fitness. Thus I adopted the approach of discerning how brains work and attempting to emulate those capabilities by simulating the components at just the right scale of size and dynamics. In my Jan. 26, 2011 post: Brain Complexity at Multiple Scales I review the complexity of brains in a hierarchy of subsystems. I started with the synapse, the communications interface between sending and receiving neurons, and worked up to larger scale structures that constitute the human brain.
I got my PhD in computer science, mostly based on my development of a unique learning mechanism that does emulate real life synapses from invertebrates to higher vertebrates. Many of my research papers can be found at: Adaptive Agents Lab, Publications. This research also was a basis of being granted tenure and promotion to associate rank at Western Washington University and carried over with my move to the University of Washington Tacoma.
That research hit a roadblock (or several) in terms of how far I could extend the emulation approach to larger more complex brains. The hardware and software of the day were simply inadequate and UWT was, at that time, an upper division transfer institution without an established graduate program (I would later help design the latter). Finding students who were interested in the kind of research I wanted to pursue was essentially impossible. That is about the time I started investigating energy matters and came across biophysical economics and made progress in my long-standing interest in the brain basis for what I conjectured was a massive lack of wisdom in human beings, who called themselves Homo sapiens or “man the wise”, but failed to demonstrate much wisdom in the way they were managing their affairs. So I put the adaptive agent work on the back burner.
As I came to realize I was also facing a snag in that my approach was way ahead of its time. The prevailing popular approach involved artificial neural networks (ANNs) and fuzzy logic (and control). The former received wide scale popularity and no small amount of hype about how it emulated the way the brain worked. Everyone at the time (late 1990s) expected breakthroughs in, for example, the so-called “learning” algorithms that supposedly captured the essence of synaptic plasticity, but were in fact very wide of the mark. Unfortunately for me the various journals that published articles in that field were dubious of my papers because they didn't look anything like ANN papers. And the problems that I claimed were solved by my approach were not even on the radar screens of researchers in the field (more about this below). When I started getting reviewer comments back that said some unkind things about my work I realized it was time to find a new line of research. So I got into the modeling of biophysical economics using principles from systems science and refining my working papers on sapience. All of which got me interested in writing a general systems science textbook because I couldn't find anything of the sort when I looked. There were tons of systems science sub-field books on cybernetics, complexity, network theory, etc. but nothing that really presented the panoply of subjects and showed how they interrelated with one another.
So I put down my research in adaptive agents and left it for another day. That day finally arrived in the form of recent significant improvements in sensor and computer technologies. Much else had changed as well. We have a very active graduate program now, and I have several students working with me on a new robot. We have an undergraduate engineering program and a number of students in that are working on the physical implementation of the robot platform, reusing the old bodies of my earlier robots, but with new and improved sensing. As importantly the realization has sprung up among a number of researchers that neuromimic processing was not just being able to demonstrate a change in a synaptic weight. Today many researchers are actively partnering with neuroscientists to discover the deeper aspects of live neural network processing. In part this is bolstered by the new neuroimaging techniques that allow non-intrusive observation of brain functions in living beings (including the human kind). There are now several massive-scale research programs to discover and develop more true to life neuromimic processing, for building artificial brains. So far I have not come across any work that rivals my own with respect to what I hold is the key to understanding all else in neural processing. In my description of synaptic processing in the above mentioned post, I claim that the key is the multi-time domain aspects of synaptic plasticity that is hard to simulate using the current computational methods. My Adaptrode mechanism, however, does simulate this aspect and allows my neuronal networks to much more closely emulate living ones.
Now with new hardware/software/sensors and related technology, and with the increasing understanding among other researchers of the importance of the brain emulation pathway, I am gearing up to tackle this problem once again.
The Adaptrode solves several long-standing problems in simulating real synaptic dynamics. Many of these problems were recognized by neuroscientists and behavioral psychologists but were largely overlooked by machine learning approaches until those approaches hit their own walls in emulating natural intelligence. The core of the problems resolved in the Adaptrode is dealing with non-stationarity in real-world environments. Put succinctly, things change. More to the point the statistical properties of causal relations between systems are known to change under a wide variety of conditions. The Adaptrode mechanism (I prefer to call it that since I am working on a hardware implementation as well) can adapt to changing relations while still maintaining long-term traces of prior relations. Other forms of machine learning that rely heavily on single-weighted nodes, regardless of their underlying algorithms cannot handle what is known as the catastrophic interference problem, where learning new relations tends to destroy previously held ones. Biological memory is known to exhibit what is known as “savings” or the ability to maintain memory traces that seem to have been forgotten but are really just in a low power dispositional form. If the old relations return, the animal doesn't have to start from scratch to re-learn what it had once known. This was explained in a 2000 SPIE conference session on Biologically Inspired Robotics, in Boston MA. Adapting Robot Behavior to a Non-stationary Environment: A Deeper Biologically Inspired Model of Neural Processing.
How to Make a Robot Self-Aware
The typical view of agent intelligence is what we might call the throughput model. External sensors (eyes, ears, etc.) pick up signals from the world. The brain processes those signals, comparing the current situation with memories of prior situations and, given mission motivations, makes decisions on what actions to take. The commands are passed back to the motor control centers and the agent behaves, presumably appropriately. What is missing from this scheme, and I had been missing it as well in my previous work, is what Antonio Damasio has brilliantly identified as the requisite sense of self, or as he called it, a “proto-self image”. He contends that even the most primitive brains have this capacity and that it is absolutely prerequisite to any kind of truly intelligent adaptive behavior. In the first post in this series, Exploring Consciousness I hinted at what it takes to produce a self image in a brain. Starting with a way to distinguish external sensory inputs between those caused by the self (e.g. touching one's self produces a sensation of being touched) and those caused by other forces/agents by comparing proprioceptive inputs (e.g. muscle tension) to the external stimulus to detect a correlation. In figure 2 I also showed another map called interoception which produces images of what the internal milieu of the body “feels” like. The viscera are also sensed and this information is part of the overall body management process. For example the current state of blood pH is sensed and sent to the brain where commands to produce behavior in other organs or even in external movement to correct any errors are computed.
We do not routinely think about interoception with robots. Proprioception is, of course, essential. For example, we use quadrature encoders to measure motor rotation and that is fed back to the brain to give the robot a sense of how fast it is moving and turning. But interoception has been generally limited to something like the battery charge to signal the need to recharge the battery. This is far from the rich information stream that the viscera produce telling the brain what the continuously updated state of the body is. And, therefore, it is hardly possible to create a map of internal state when we don't really bother to measure it.
Damasio's whole argument about self-consciousness is built up from the idea of a body-state map that continuously monitors the body's needs and produces the core motivations for behavior, e.g. hunger, thirst, etc. In my prior work I simulated some of this algorithmically without thinking it was important to have a real map of internal states. Now my thinking has changed. In fact I am now planning to incorporate many more internal sensors that measure the “health” of the robot's internals. For example I plan to deploy an array of temperature and humidity sensors throughout the robot's body interior. These will monitor its internal conditions. We can then use external heat sources, for example, to change the temperature profile inside the robot. But at the same time we will be incorporating a number of power-consuming processors that will, in their own rights, generate heat. We can use the internal temperature profile to regulate power consumption by selectively down modulating some computational functions when it gets too hot in their vicinity.
For now there is no real “chemistry” taking place inside the robots that would stand in for physiological processes needing to be regulated. However, we are toying around with a few possible ideas on how to incorporate something like this in the future. For example there are now odor detectors that are used in robots that look for things like gas leaks or explosives by their odorant signatures. So the idea of a robot having physiology isn't so absurd. If, in the future, robots not only patrol their environments but “produce” substances such as antidotes to toxins detected then an internal chemical plant would have to be monitored and controlled.
My group is just starting to consider what additional internal state information might be possible and even necessary in order to produce an interoception map. We have figured out the exteroception (external senses) and proprioceptive maps and now we want to add this new interoceptive map to round out the perspectives needed for an agent to have a sense of self.
This is the first step toward producing a machine that is not just aware of its environment and is not just able to learn in a non-stationary one, but toward a machine that is simultaneously aware of its own self and uses all of this information to control behavior for its own best advantage. Figure 3 in the Exploring Consciousness post shows basically what we are trying to build now. By Damasio's model, the next step would be to develop the cerebral neocortex and have its ability to construct models of the world and self (Figure 2 in the Who is I post). That is roughly about the level of intelligence and consciousness as in early mammals. That is our goal. To produce an agent with roughly the behavior competencies of a mammal like a mouse.
I'll try to keep you up to date on our progress. I am putting together a video of the current robot version that uses the updated hardware and is at the stage of wandering and avoiding obstacles (see some of my work from the Adaptive Agents Lab page). Meanwhile if anyone wants to make a grand contribution to the advancement of science, I am seeking funding!!!
. Of course I do not mean design in the sense of that created by an intelligent designer. I use the term more broadly as the organization and structure of a system that suits a purpose (e.g. staying alive) or performs a function that is in some sense useful to other systems. When I talk about human engineering or architecture, or for that matter birds' nests, I will use the phrase, “intentional design.” Even so, I have argued, on occasion, that even intentional design is a result of a universal evolutionary principle.