Number Crunching Machines

[vidam]

I just got that robots are essentially number crunching machines in software.

So if humans are not number crunching machines then how is it we are so far superior to robots in dynamic environments?

[JesseWelling]

I think of humans more as self modifying FPGA's. I think they key part is the self modification...

[vidam]

I don't know enough about FPGA's to know the difference. But I am interested in finding out more about self modifying FPGA's? If they can be applied to robotics such that robots could be a little smarter than fax machines then this is the field for robotics...

[EDITED]

Before today, I was one of the few people willing to accept the true definition of today's modern robot. I was so ignorant to believe that robots are close to being able to replace humans in basic functions.

I realize now that deep down inside I always knew this subconsciously that this would never happen. It's like the medicine man coming into town selling you sugar pills that promise XYZ.

Well today it all changed for me. I had a meeting with a very smart person today about the current state of robot cognition. After the meeting, I stopped ignoring what I always knew and it just clicked in my brain. Robotics as we know it today are really just machines that are basically functional for specific tasks much like a fax machine is functional for sending faxes. We're so far away from the real meaning of a robot that it isn't even realizable yet.

Well maybe for once I can focus on the issues that will help robots become what I had imagined they should be.

Sorry for the rant! I'm just going though major transition in my thinking.

[EDITED]

[szhang]

Before today, I was one of the few people willing to accept the true definition of today's modern robot. I was so ignorant to believe that robots are close to being able to replace humans in basic functions.

What are these "basic functions"?  Robot CAN and have replace humans in many situations.

I realize now that deep down inside I always knew this subconsciously that this would never happen. It's like the medicine man coming into town selling you sugar pills that promise XYZ.

There is nothing stopping it from happening. As the cliche goes, "Never say never"

Well today it all changed for me. I had a meeting with a very smart person today about the current state of robot cognition. After the meeting, I stopped ignoring what I always knew and it just clicked in my brain. Robotics as we know it today are really just machines that are basically functional for specific tasks much like a fax machine is functional for sending faxes. We're so far away from the real meaning of a robot that it isn't even realizable yet.

It is dangerous to draw too much out of one conversation.  Sure, robots are machines, but they will always be machine because that is the definition of machine.

AI research still has a long way to go but I am optimistic it will get there.

Think of a computer.  Can you pin a specific function?  It can't learn yet, but it has gone a long way from simple fax-machines.

Well maybe for once I can focus on the issues that will help robots become what I had imagined they should be.

What did you imagine they should be?

For a look at comtemporary robotics, check out the Darpa Urban Challenge.

[vidam]

Before today, I was one of the few people willing to accept the true definition of today's modern robot. I was so ignorant to believe that robots are close to being able to replace humans in basic functions.

What are these "basic functions"?  Robot CAN and have replace humans in many situations.

I realize now that deep down inside I always knew this subconsciously that this would never happen. It's like the medicine man coming into town selling you sugar pills that promise XYZ.

There is nothing stopping it from happening. As the cliche goes, "Never say never"

Well today it all changed for me. I had a meeting with a very smart person today about the current state of robot cognition. After the meeting, I stopped ignoring what I always knew and it just clicked in my brain. Robotics as we know it today are really just machines that are basically functional for specific tasks much like a fax machine is functional for sending faxes. We're so far away from the real meaning of a robot that it isn't even realizable yet.

It is dangerous to draw too much out of one conversation.  Sure, robots are machines, but they will always be machine because that is the definition of machine.

AI research still has a long way to go but I am optimistic it will get there.

Think of a computer.  Can you pin a specific function?  It can't learn yet, but it has gone a long way from simple fax-machines.

Well maybe for once I can focus on the issues that will help robots become what I had imagined they should be.

What did you imagine they should be?

For a look at comtemporary robotics, check out the Darpa Urban Challenge.

Robots cannot handle context. Something humans can do very well. For example, if I mounted a bedsheet between 2 orange cones a robot will think it is a wall. But a human know it is just a bedsheet.


Actually I did research the Darpa Urban Challenge as well as the Grand Challenge before I made my comments on this board. At the urban grand challenge they had to put up concrete blocks around the edges of the sidewalks so that the robots could not go off the roads. They had to do special things like that to make a very controlled situation. One robot at the Urban Grand Challenge actually destroyed a building by going off the road and running through it.

So we have a long way to go still. I'm learning to be realistic about what robots can and cannot do. I feel upset that people fooled me when I was a newbie in 1998. I and most other people don't like to chase a pipe dream and waste their life. Once we get down to reality then we can figure out where the real research lies. It's not making robots that function in controlled environments like fax machines. Although that will be the kind of robot of our modern world that has any practical application.

[EDIT]

If you know of good examples that prove robots can understand context the way humans do, please post your replies here. I'd be happy to learn about how they did it. Thank you!

[EDIT]

[hgordon]

"Never" is a long time, but I would agree that there is a very large gap between the computational ability of robot vs the cognitive ability of humans to adapt to new situations and environments.

However, that does not mean that it is not a worthwhile challenge to try to bridge that gap.

With the right sensors, a robot probably could tell the difference between a wall and a bed sheet as well as a nematode worm, but probably not as well as a mouse.  10 year from now, perhaps mouse-level capabilities might be exceeded.

This is a big challenge, and the techniques that are currently in favor will probably be rejected in coming years while new techniques are developed or old ones are revisited.  That's the process.

[szhang]

There are teams that failed the urban challenge, and there are teams that broke new grounds.  It was a controlled situation so that the bad teams won't kill the bystanders.  The CMU robot, Stanford robot, and Vriginia tech robot both performed perfectly.  Probably even better than a person could have done.  And don't tell me that people don't frequently crash into buildings and other cars too.  Concrete divider on freeways serve exactly the same purpose as the concrete blocks near sidewalks: to prevent accidents.

Robots cannot handle context. Something humans can do very well. For example, if I mounted a bedsheet between 2 orange cones a robot will think it is a wall. But a human know it is just a bedsheet.

Some techniques such as neural networks can identify objects fairly well even with context.  That is not the biggest problem in AI, which is rather learning and flexibility.  In addition, cognitive studies have shown that humans aren't good with context until significantly late into their development cycle (~4-5 years old).

People fooled you?  Please! You fooled yourself!  Any attempt at research into AI and you'll know that reaching human cognitive abilities is still a long way off.  I doubt you wasted any of your life (on this "pipe dream" at least)  because you never tried.

Why are you so quick to dismiss artificial intelligence as a pipe dream now?  Because it seems harder than you thought?  If we gave up everytime something is hard, none of our achievements such as computers, planes, or electricity would exist.

Trying to reach human intelligence immediately is foolish.  It is like trying to get to the moon before you learned how to work metal.  Science is a process.  Robotics is still a fairly new field with many exciting things going on right now, if you'll just look for them.

[Admin]

Remember, robots only do what they are programmed to do, and what the hardware can handle from laws of physics.

AI must be programmed in by 'intelligent design'.

(sorry, but I felt that a perfect word to describe the creation of robots)

[vidam]

WOW szhang you are really taking this whole discussion out of my original context. (no pun intended)

I'm not advocating giving up on robotics or automation. There is a reason why the word "Artificial Intelligence" is a dirty word at some famous institutions. They spent millions of dollars on AI and without much return on the investment.

I've spent years studing AI..neural networks, evolutionary computation are among them. I know the failing of most of these algorithms. They are slow to learn and cannot do well in dynamic environments. Robots don't handle change very well either.

Enough said. Think what you want. I'm merely pointing out we can't have progress without realizing the limitations.

[vidam]

I'm still waiting for people to answer this question. Thus far I've been getting hit back with a lot of fluff.



If you know of good examples that prove robots can understand context the way humans do, please post your replies here. I'd be happy to learn about how they did it. Thank you!

[hgordon]

Not a direct answer to your question, but probably the best work I've seen recently in building robots that are truly adaptive is discussed in "How the Body Shapes the Way We Think" by Pfeifer and Bongard.  The term applied to their work is "embodied intelligence", and while it's still neural nets and genetic algorithms at the core of their techniques, some of the results are pretty impressive (e.g. robots that adapt to loss of a leg).  The contention made by researchers in this field is that computation without sensory feedback, i.e. disembodied reasoning, is fundamentally limited and won't solve the challenges you describe.
 

[vidam]

Thanks hgordon. I will buy a copy of this book. It shows that people have begun to think about these issues and find alternative solutions.

While I'm here I thought I would post another place where they are working along the same line of thought:

http://www.numenta.com/about-numenta/numenta-technology.php

[hgordon]

I have looked at their work and read Hawkins book, and it looks like they may have a new twist on neural networks that adds a temporal component, but I think they will still run into the fundamental limitations that other "disembodied intelligence" approaches have faced.

Actually, some other reading you might find interesting is "Where Mathematics Comes From" by Lakoff and Munez.  It approaches embodied intelligence from the human side, basically developing the idea that mathematics concepts are very much developed from the way the brain is wired combined with the way that our sensors interact with the world.

[superchiku]

WELL, people are really getting influenced by terminator,matrix etc etc etc

There is always a strict margin between reality and fantasy .

The reality is that humans contains thousands and thousands of neurons and receptor brain cells which have the ability to differentiate light ,heat,vision and most important emotions , even a baby can realise when it is danger or that it is going to fall off a table if it crawls any further... can a robot do that unless it is pre-programmed???

The reality is no matter how much we try , we can never play God we can never make what he has done , we can never imitate nature and save ourselves from facing the consequences ..

this gap should be kept always..

[vidam]

WELL, people are really getting influenced by terminator,matrix etc etc etc

There is always a strict margin between reality and fantasy .

The reality is that humans contains thousands and thousands of neurons and receptor brain cells which have the ability to differentiate light ,heat,vision and most important emotions , even a baby can realise when it is danger or that it is going to fall off a table if it crawls any further... can a robot do that unless it is pre-programmed???

The reality is no matter how much we try , we can never play God we can never make what he has done , we can never imitate nature and save ourselves from facing the consequences ..

this gap should be kept always..

Just for the record, I was never a fan of terminator. I hated the Terminator movies.

I was influenced by a movie of Asimo in 1998 at a seminar Carnegie Mellon. When I saw that android walking, it was like an asteroid hit my planet.

Keep in mind, they never mentioned that the robot was just demonstrating a pre-programmed walk and if you walked up to it and tapped it on it's shoulder it would fall over.

I was also given a tour of the Robotics Institute and I saw the navlab an autonomous vehicle that traveled across the US driverless.

what I wasn't told is why we don't see Android in the streets or autonomous vehicles driving us to work.

Therein is where I failed to ask the question.

years later, x number of robotics courses and AI courses later, I'm sitting in some guys office trying to come up with a project proposal that uses Robots for military applications and why we can't just start using them tomorrow.

What are the problems why we can't use them tomorrow. Darpa Grand Challenge and Urban grande challenge were deemed a huge success. But why are we still holding robots in the wings.

That brings me to this post and why I posted it here today.

[szhang]

WOW szhang you are really taking this whole discussion out of my original context. (no pun intended)

I'm not advocating giving up on robotics or automation. There is a reason why the word "Artificial Intelligence" is a dirty word at some famous institutions. They spent millions of dollars on AI and without much return on the investment.

I've spent years studing AI..neural networks, evolutionary computation are among them. I know the failing of most of these algorithms. They are slow to learn and cannot do well in dynamic environments. Robots don't handle change very well either.

Enough said. Think what you want. I'm merely pointing out we can't have progress without realizing the limitations.

I find it hard to understant that someone who really studied AI extensive just realized robots are number crunching machines, or that robots cognition is nowhere near human capabilities.

I come from a "famous institution" for robotics and AI (Carnegie Mellon), and thus far I have yet to notice that "aritificial intelligence" is a bad word.  Nobody denied current research is limited, the point is we shouldn't let the current limitations deterr research.  Besides, current AI research HAS been focused on the "issues" such as learning adaptability to new environments.  So I fail to understand what "real issues" you are planning to focus on.

The problem is I don't quite get what you're advocating.  Maybe I'm taking you out of context, but in one sentense you dismiss AI research as a "pipe dream" and "waste of time".  In another you deny that you are advocate giving up research.

Don't confuse me with someone who thinks machines will achieve human intelligence in 10 years.  I don't.  All I'm saying is do not let your beliefs swing from one extreme to another.  It is counter-productive.

EDIT:

Technology always starts with baby steps under controlled conditions.  I agree many demos are almost "staged".  The Honda Asimo for example, is a great demonstration of biped robots, but if it steps on a pencil it will fall over.  That is what alot of the research nowdays are in, to make robots more robust, so it can survive outside the lab.  The robotics research I'm currently on for example, is a robotics leg that utilizes dynamic balance that allow the robot to lost balance for a fraction of a second, much like humans do.

[hgordon]

years later, x number of robotics courses and AI courses later, I'm sitting in some guys office trying to come up with a project proposal that uses Robots for military applications and why we can't just start using them tomorrow.

What are the problems why we can't use them tomorrow. Darpa Grand Challenge and Urban grande challenge were deemed a huge success. But why are we still holding robots in the wings.

Remember that the DARPA challenges were held in isolated locations under controlled conditions, and were not conducted in public locations for good reason.

At the same time, consider that most commercial airplanes are flown almost completely by autopilot with little pilot interaction.  That probably wasn't the case 50 years ago - it took a long time for the bugs to get removed from the system and for the FAA to certify safe operation, but at this point, the onboard computers have the possibility of replacing all pilot functions, including takeoff and landing.

[vidam]

I dismiss the current standards in AI as a pipe dream to getting where we need to be.

Maybe that is where we are crossing our wires here.

[szhang]

I dismiss the current standards in AI as a pipe dream to getting where we need to be.

Now that is more specific, though your new "standards" are very similar to the direction of current research already.

I've met lots of people and I have to say it is the people I've met from CMU that seem to have the hardest time accepting any notion that they might not have the answers to all the problems.

Ah, yes, the ultimate tool in debate: ad hominem.

[vidam]

years later, x number of robotics courses and AI courses later, I'm sitting in some guys office trying to come up with a project proposal that uses Robots for military applications and why we can't just start using them tomorrow.

What are the problems why we can't use them tomorrow. Darpa Grand Challenge and Urban grande challenge were deemed a huge success. But why are we still holding robots in the wings.

Remember that the DARPA challenges were held in isolated locations under controlled conditions, and were not conducted in public locations for good reason.

At the same time, consider that most commercial airplanes are flown almost completely by autopilot with little pilot interaction.  That probably wasn't the case 50 years ago - it took a long time for the bugs to get removed from the system and for the FAA to certify safe operation, but at this point, the onboard computers have the possibility of replacing all pilot functions, including takeoff and landing.

The autopilot take off and landing is a perfect example of this. The human takes over the landing part I think? The airplane doesn't land. At what point does the airplane signal to the human to take over the landing. How did we gain trust in airplanes to fly us from point A to point B. This is a good paradigm to study and perhaps model for getting robot-human relationship to trust each other collaboratively in a real-world fast-changing environments.

[szhang]

Currently many planes can land automatically.  All the pilots have to do is flip a couple of switches.

[vidam]

Just looked up the definition ad hominem... http://en.wikipedia.org/wiki/Ad_hominem

I had no idea they had a dictionary word to describe it.

[paulstreats]

Heres my humble opinion on the original question:

So if humans are not number crunching machines then how is it we are so far superior to robots in dynamic environments?

I think that in many cases biological brains(such as human) are in fact number crunching machines.

Think about when you are crossing a busy road you monitor traffic coming from both directions and then judge when to cross, how fast etc... You can do this long before being taught vector movements at school yet the math equivalent is actually happening in your brain. Albeit without you realising it.

The reason is that in the advanced biological brain there are 2 seperate structures, there is 1 subconcious/subliminal structure and 1 concious structure. In most cases the subconcious actually handles most of our basic functions and performs the calculations, whereas the concious is the area that we think with. The two areas are defined as such:

a)The concious cannot access the subconcious, it can only send requests such as movement requests.
b)The subconcious actually programs the conscious area and sets boundaries, rules and behaviours
c)The concious area is what is used to make decisions
d)The subconcious can override the concious decisions only when primary instincts are being compromised (such as if you command your body to fall over, the subconcious interrupts and forces you into a state of balance)
e)the subconcious records everything that you see, hear, touch, taste, smell.
f)the conscious only records the information that it sees as important to the situation. (meaning that during the day you see millions upon millions of points of objects, yet the conscious area can only recall matbe a few hundred when asked. Wheras the subconcious records everything!)
g)there are more, but i think this is the core of it.

finally:

computers / robots are number crunching machines, when they are programmed in the lower levels. When you get to higher level programming (look at java). You can actually make an entire program without putting any mathematics into your code. The number crunching has already been done in the lower level, and so all you are left with is manipulating objects this is similar to a brain allowing a subconscious area to do all of the work while the actual front end just does the higher level thinking.

Also, i mentioned that biological brains are number crunchers. You have to open your mind up a bit to accept this, since it doesnt have a base number system to work with, this is taught to the conscious area later in development. More likely is that it calculates in fractions or ratios, so that it can decide between this is bigger or this is further away etc.. without placing an actual number to the size or distance. I think that it works as a big mass of altering ratios. When you look at something and think: that can of coke is 1 metre away, what actually happens is that your subconscious returns its ratio of the distance compared to previously known ratios and then your conscious brain takes this ratio and applies a number to it (that was taught in school) that is compared to numbers associated to previously known assigned number/ratio pairs. Until you are taught a measurement system such as metric / imperial, you cannot say that something has a distance, yet you actually know that can of coke is further away than the book on the table due to ratio assignment.

Sorry for the long post, and im in no means a brain surgeon this is just the understanding that i have gained over researching the subject myself

[vidam]

I recall an interesting research area called "Active Learning". Active learning would require a human being in the loop of a robot perception system to make decisions where the robot cannot react quickly in a dynamic outdoor, context sensitive environment. It's a human-robot collaboration that might help robots get to the field more quickly. But then you must worry about network delays, noisy connections, and so forth.

A quick google search turned up someone's thesis at CMU. http://www.ri.cmu.edu/pubs/pub_5529.html

[vidam]

Heres my humble opinion on the original question:

So if humans are not number crunching machines then how is it we are so far superior to robots in dynamic environments?

I think that in many cases biological brains(such as human) are in fact number crunching machines.

Think about when you are crossing a busy road you monitor traffic coming from both directions and then judge when to cross, how fast etc... You can do this long before being taught vector movements at school yet the math equivalent is actually happening in your brain. Albeit without you realising it.

The reason is that in the advanced biological brain there are 2 seperate structures, there is 1 subconcious/subliminal structure and 1 concious structure. In most cases the subconcious actually handles most of our basic functions and performs the calculations, whereas the concious is the area that we think with. The two areas are defined as such:

a)The concious cannot access the subconcious, it can only send requests such as movement requests.
b)The subconcious actually programs the conscious area and sets boundaries, rules and behaviours
c)The concious area is what is used to make decisions
d)The subconcious can override the concious decisions only when primary instincts are being compromised (such as if you command your body to fall over, the subconcious interrupts and forces you into a state of balance)
e)the subconcious records everything that you see, hear, touch, taste, smell.
f)the conscious only records the information that it sees as important to the situation. (meaning that during the day you see millions upon millions of points of objects, yet the conscious area can only recall matbe a few hundred when asked. Wheras the subconcious records everything!)
g)there are more, but i think this is the core of it.

finally:

computers / robots are number crunching machines, when they are programmed in the lower levels. When you get to higher level programming (look at java). You can actually make an entire program without putting any mathematics into your code. The number crunching has already been done in the lower level, and so all you are left with is manipulating objects this is similar to a brain allowing a subconscious area to do all of the work while the actual front end just does the higher level thinking.

Also, i mentioned that biological brains are number crunchers. You have to open your mind up a bit to accept this, since it doesnt have a base number system to work with, this is taught to the conscious area later in development. More likely is that it calculates in fractions or ratios, so that it can decide between this is bigger or this is further away etc.. without placing an actual number to the size or distance. I think that it works as a big mass of altering ratios. When you look at something and think: that can of coke is 1 metre away, what actually happens is that your subconscious returns its ratio of the distance compared to previously known ratios and then your conscious brain takes this ratio and applies a number to it (that was taught in school) that is compared to numbers associated to previously known assigned number/ratio pairs. Until you are taught a measurement system such as metric / imperial, you cannot say that something has a distance, yet you actually know that can of coke is further away than the book on the table due to ratio assignment.

Sorry for the long post, and im in no means a brain surgeon this is just the understanding that i have gained over researching the subject myself

thanks for the effort. How about a baby, a puppy which are born with innate instinct such as nursing for milk from the mother. That wasn't a number crunching behavior was it now?

[paulstreats]

Not exactly, and dont think that im arguing here or anything its a subject that im trying to work out myself, but the puppy must be able to measure the distance of its mother to its own jaw, or if it doesnt it eventually bases its initial distance judgement on that. Also judge how much force to bite down with before suckling. The baby is born without any ratio threshholds etc so it must learn these. When it comes to basic instincts such as knowing its supposed to get milk from its mother than thats an area outside of my research scope for the time being

The whole starting life part is difficult because there is nothing to act as a comparator, which is probably why it takes a few attempts at suckling before the baby/puppy actuall gets it right.

I was putting some thought into the whole comparator thing last month, in that if a baby experiences some slight painful event(such as being scratched accidentally), then it doesnt have anything to compare that with and it will feel the most painful thing imaginable. Once the baby experiances something more painful than a scratch then the comparason threshhold gets heightened and a scratch doesnt seem that bad. So basically the threshhold must have some value, it might not be something that we think of as a conventional number but it is still definable as a number when compared to the previous state/value I think that it just operates as an undefined ratio where a new experience can increase the ratio..

[vidam]

I was actually thinking/proposing thresholds too. I thought if the robot knows it's own thresholds/constraints then it can signal the human to intervene when it needs help making a decision such as a decision to fire after determining if this is an enemy vs our ally.

[szhang]

I think that in many cases biological brains(such as human) are in fact number crunching machines.

Think about when you are crossing a busy road you monitor traffic coming from both directions and then judge when to cross, how fast etc... You can do this long before being taught vector movements at school yet the math equivalent is actually happening in your brain. Albeit without you realising it.

Actually, some research suggest that our brain take shortcuts instead of crunch numbers in these circumstances.  We are horribly bad (and slow) at crunching numbers compared to computers.

For example, imagine you're trying to catch a baseball, do you actually compute the trajectory? No, instead you take shortcuts (sometimes without recongnizing it) like move to a position where the falling basebal seems to be going straight towards you.  This applies to animals as well.

[paulstreats]

For example, imagine you're trying to catch a baseball, do you actually compute the trajectory? No, instead you take shortcuts (sometimes without recongnizing it) like move to a position where the falling basebal seems to be going straight towards you.  This applies to animals as well.

And so having to calculate where the ball is going to land in order to move there.....
And then once there recalculate new trajectory

One of my family is into bike racing etc.. And when he hits a jump, i can easily see where his bike is going to land before it actually does, this must be calculating and visualising trajectory - and no i dont move myself in front of the bike without realising

  There must be some system that performs mathematics without actually involving numbers and standard mathematical systems. After all you can invent your own number system and maths systems, they just wouldnt follow convention

Anyway thats it from me on this subject for now, its too much of a wide umbrella to tackle all at once

[szhang]

No, you're moving around UNTIL the ball stops moving.  You don't take one look at the ball and know exactly where you want to go.  That is why you keep staring at the ball while running.

can you really predict exactly where the bike land. or are you just using some general knowledge to approxiamately predict where he lands?  You can't solve complicated quadratic equantions without knowing any of the variables.  All you're doing is making an educated guess (based on experience).  That is the kind of "shortcut" i'm talking about.

EDIT:

It is easier to think these actions as controlled by a fuzzy logic PID controller in the brain.  You look at the error, if it seems BIG, move fast, if not, move slower, if I'm already moving very fast, move slower.  No complicated math necessary.