Catching a Ball: How Our Brains Make Quick Decisions

[Ivan Seeking]

Perhaps we have a few members who can shed some light on this.

It has always amazed me that humans [if not dogs and other animals] can do what they do. For example, consider a baseball out-fielder catching a high fly ball. I would assume that we are drawing mental lines of flight that we constantly adjust based on the latest visual cues from the ball. But even if so, I know that a pro ball player must begin to position himself for not only the line of flight, but also for the trajectory of the ball at the moment that it leaves the bat. So do we coordinate the sound and angle to do this this - we estimate how far it will travel by the sound, and the line of flight by the angle that the ball leaves the bat?

And what of more difficult tasks? While talking about my days as a hellion on wheels, I was thinking about how we judge speed and can anticipate how fast we might try to negotiate a turn. Even if we have never seen the turn, and even if we don't know our actual speed, with practice we can somehow estimate this all in a second or less, and make the correct decision in a moment while coming into a turn, then constantly adjusting for the motion according to a variety of subtle clues - the sound of the engine, the forces acting against he steering wheel, the accelerations felt, the relative angle of the car wrt the surroundings and how much that angle changes over time, the properties of the road being slick, sticky, hard or soft, wet or dry, sludge or rock, etc -that we somehow interpret with amazing speed and then make life or death decisions with relative ease.

How much do we understand about how the brain manages these sorts of tasks, and how do we do it?

 

[DocToxyn]

Even if we have never seen the turn, and even if we don't know our actual speed, with practice we can somehow estimate this all in a second or less, and make the correct decision in a moment while coming into a turn, then constantly adjusting for the motion according to a variety of subtle clues - the sound of the engine, the forces acting against he steering wheel, the accelerations felt, the relative angle of the car wrt the surroundings and how much that angle changes over time, the properties of the road being slick, sticky, hard or soft, wet or dry, sludge or rock, etc -that we somehow interpret with amazing speed and then make life or death decisions with relative ease.

I think you hit the biggest influence right there - practice - and the experience you gain from it. I watch my son when we play catch and in the beginning he was horrible, didn't even put up his hands in time (it was pretty comical ). Anway, he has gotten better and his other skills have progressed as well. Obviously there is a component of increased strength/coordination as he gets older, but much of that coordination has to do with body control and repetition. Doing something so often until it becomes "unconscious"..."mind of no mind"...these are terms/phrases you hear athletes use and I think experience weighs heavily in this. As far as actual brain systems activated, I'm sure its a complicated symphony between memory, sensory, and motor systems. I don't have anything specific right now, a few searches tommorrow should give some more.

 

[chroot]

Strange as it sounds, artificial neural networks can be taught to perform "tasks," even though no one can look at the resulting, trained neural network and point at "how" the network does it. The learning process results in a weight matrix -- each number representing the strength of a connection between two neurons. The weight matrix certainly encodes all of the network's functionality, but no single mechanism, or pattern, can really be singled out from the entire weight matrix. The learned ability is distributed throughout the entire network. The network's functionality, in a phrase, is "emergent behavior."

If you take potshots at the network, cutting connections between neurons at random, you'll find that the network will continue to do its job fairly well until a great many such cuts are made; thus, common networks incidentally include a lot of redundancy distributed throughout.

I'd say the baseball player's brain is wired up in a similar fashion: some input "vector," a list of all the cues the player is currently sensing, is mapped to some output vector, a list of all the actions the player's muscles must take. The actual path from input to output is not a "computation," in the common algorithmic sense, but a filtering of the input through layers and layers of interconnected neurons. The computational machinery exists in the form of a large weight matrix multiplied by the input vector, resulting in an output vector.

- Warren

 

[Ivan Seeking]

Do you think that we could we make a computer catch a fly ball as in the example - as does a pro ball player - or is this still beyond, or far beyond modern computing speed, technique, etc? What are the greatest limiting factors; speed for one I would assume?

 

[chroot]

I'm pretty confident that the computational power of a standard PC would be more than adequate to perform the calculations of where a ball is going to land, given constant visual stimulus in usable form.

The problem with "making a computer do it," as I see it, is a matter of the sophistication of sensory apparatus. It's very difficult -- if not impossible with today's technology -- to build a video-camera system that has the visual acuity, the dynamic range, the capacity of incredibly high angular-velocity movements, and so on of the human visual system. If a computer is expected to do what a ball-player does, e.g. track a tiny ball over a ballistic trajectory occupying a large solid angle as viewed by the player, constantly altered by wind, ball spin, etc., even with the ball passing through the glare of the Sun, you'd need to give the computer a sensory system with sophistication equal to that of the ball-player. Such sophistication does not yet exist, that I know of.

- Warren

 

[zoobyshoe]

Yeah, sensing the ball and the environment is the first thing that has to be accomplished and the outfielder is using every sense except taste and smell to catch a ball. What's left are: touch, balance (the sense organs of acceleration in the inner ear), proprioception (sense of body position), sight, and hearing.

Just because he doesn't need taste and smell for the task at hand doesn't mean they shut off, of course, and they are actually in use to the extent that they're being monitored for anything alarming or dangerous. A sudden whiff of the smell of dog poop could easily interfere with the task.

Balance/acceleration is probably the easiest sense to mimic mechanically with various kinds of gyroscopic devices. They've had sophisticated systems that do this for guided missles for years now. The other senses are all a lot more complex and consist of many sub-senses that, as far as I know, no one has intelligently tried to mimic (but I don't make any effort to follow AI type stuff, so maybe there are people doing it.)

 

[CosminaPrisma]

I have a psychophysics book at home that might help me shed some light on this. Maybe I'll look at it the next time I go home and get back to this thread.

 

[gerben]

Do you think that we could we make a computer catch a fly ball as in the example - as does a pro ball player - or is this still beyond, or far beyond modern computing speed, technique, etc? What are the greatest limiting factors; speed for one I would assume?

The computing speed would probably not be the major problem. Nervous systems are very slow compared to machines with copper wires. The fastest reaction to any change in a visual signal occurs, at least, more than a tenth of a second after the signal. This means that the visual information that is used to steer the muscles that produce the response corresponds to the measured outside situation more than 100 ms before the response (a computer could do much better in this respect, i.e. base its response on more current information).

This inherent slowness of neural systems implies that animals must rely heavily on predictions of what the ‘current state of the outside world’ (i.e. its state at the time they intent to act) could be, otherwise responses to moving objects would always miss the intended goal. Such predictions are presumably a function of the measured sensory information (from more than 100 ms before the response) and of previous information about similar situations. This previous information could be information about ball-flight acquired during previously seen ball trajectories; or perhaps the physics of object motion has acquired a fixed spot in our brains in the course of evolution. Whatever is the case, evaluating this function obviously requires computing power, but any modern computer could easily deliver the required amount and apparently the nervous systems of many modern animals can do so too.

However, the hypothesis that a computer could probably do the necessary computations to determine where the ball would fall does not make the basic problem less interesting. Although in many instances we are able to catch a fly ball, determining where the ball will land AND how to stimulate the muscles (or motors) in order to get there at the right time is still an unsolved theoretical problem in both neural science and in robot control, in almost any case.

 

[mandril]

I saw a robotic arm that does exactly that: catches a flying ball. It uses Neural Networks and Genetic Algorithms to create a very smooth movent, cathing the ball almos as we do it. unfoutunetly I don't have the reference, but am sure that if you look on the web you can find it