What is the difference between artificial neural network and biological neural network?

Answer 1

Your brain is a neural net, which is a class of computing architectures. Neural nets can be made out of neurons along with supporting brain cells, or out of electronics, or with chemistry, or with levers and cogs like an ancient adding machine, it doesn't matter, it's just a type of computational hardware architecture.

For the sake of round numbers, at the tone, your brain contains exactly one trillion living cells. Oops, and declining. You can do the math. Some of these are neurons, and some are glilal cells and other support cells. You might find it interesting that the support cells outnumber the neurons by two or three to one, and that they don't just "support", but they are essential to the computation, the back-propogation network , for you geeks.

These trillion cells do not operate one at a time in sequence. There are cycles and pipelines, but all of the cells are powered up and doing their thing at the same time, all one trillion of them. A million million. There's a lot of you in there.

But I'm just one thing, you cry. I think one thought after another, sequentially. Some kind of secondary consciousness, subconscious, hunches, intuition, esp, sure, any thing is OK as long as I'm only one me.

Wrong. Consciousness is a simple brain trick. Don't get hung up on it, I'll explain it later.

You've got this network of neurons and other brain cells, I can prove it, I have an electron microscope and a very sharp knife. (Pause) I'm going to ignore the distinction among cell types and just go with the one trillion number. These trillion cells are arranged in a small number of layers, maybe seven or thirteen, compared to a trillion it's peanuts. The this flat sheet is folded and refolded, stretched and bent, and crosswired a bit, but the macroscopic, gross, structure is also a secondary part of the story. It is still this, call it ten-layer, net that does the heavy lifting.

Somebody has taken a nice flat neural network, with ten layers of one hundred billion cells each, stomped it, stretched it, crumpled it up, and stuffed in into your head. More or less.

Biology is messy, don't think of this as being crisp and neat. Everything is fuzzy. Reward diffusion. Advanced topic. Simplification follows.

Think of this ten layer mat as having the outside world come in from the top. Part of the top layer is your retina, part is the nerve cells attached to those feelers inside your inner ear that let you hear, part to the olfactory receptors in your nose , and so on. The outside world stimulates the top layer through your senses.

Cells in the second layer each connect to thousands of the cells in the top layer. A cell in this layer might connect to a row of cells in the retina in order to detect an edge, a line, an sharp change in the image. A second layer cell might connect to a set of top layer cells that let it recognize a C sharp note, or something.

The hundred billion cells in the next layer down each connect to thousands of cells in the second layer. They might look at multiple edges to find a patch of uniform color, for example.

And then another layer might put patches together into objects, and another layer might put images from the two eyes into a three dimensional model of some sort.

And then another layer would connect each of its billions of cells to thousands of cells above, including the three D model, and some of these cells would make deductions about the model. What does it mean, what will happen next, what are all the things it connects to, how does Warren Buffet relate to it, am I going to trip over it. All of these connectons, millions of them, being made at once.

(Which is why I get going and spout nonsense and foam at the mouth all the time. I encourage the pesky connections to grow. I pay attention to the kids in the back row of the classroom with the crazy ideas and the ludicrous connections. It makes me laugh, and that encourages them more. If you want to be crazy too, try it. But start demanding a psychiatrist early, there's a waiting list.)

Looking up from the bottom, you might find a cell that controls one fibre of one muscle. It can say go, stop, emergency power, and like that to that one fibre.

Above it is a layer that can send commands to all the fibres in a muscle, such as a quadricep. That is, the bottom layer listens for the signal, the information here is flowing down. Above that is a layer that does a particular action, such as a snap kick, and a layer above that decides who to kick, and so on.

Don't take my breakdown of what's in what layer literally. It's not designed by humans. You can argue about how it came about, but that much is clear, it doesn't have a software architecture that is intended for people to think about. It wasn't built with the intention of having well-defined layers with well-defined functions used consistently in a logical structure. There is no human-comprehensible description of the particular computation any brain does, except that they've done it for cockroach brains, they think. I doubt it.

That's still how it works, an neural net in layers, a few hundred billion nerve cells with thousands of connections each. Cascade computation. It's just that some of the concepts that particular cells recognize have no names in any human language. Some of the relationships you compute are beyond description, so they just come to you as a vague feeling that there is some connection. Some things you know but can't say.

Learning. The damn things do not come prewired, outside of the gross structure. They can't, because you've got the trillion cells and the thousands of connections, and only twenty-five thousand genes in your DNA. (plus support genes and 'junk', don't get me started). Not enough DNA to specify the structure. The blueprints for the Taj Mahal written in crayon on a postit note. Can't be done.

The knowledge is in the connections, the strength of the connections, and how well the neurons like the connections, and how well the support cells like the neurons.

Synapses, neurotransmitters, dendrite growth, weighting functions, pulse modulation, ion wave signal propogation, back-propogation, Google it. That's learning. There's instincts, and emotions, and other stuff. Perception triggers emotion, emotion triggers action, the sets that are hardwired are called instinct. Food triggers hunger, hunger triggers eating. Two instincts, one perception, one emotion, one action. In most cases, other emotions are also considered, and may balance the immediate response.

The emotional computer is relatively built in and only a modest upgrade of the reptile version, except that primates devote an immense portion of general purpose neurons to analyzing their relationships with other primates of the same species. Pecking order, turf, he said, she said, who's doing what with who, and how does everyone feel about it, and which of my relationships would be changed if I did so and so about it.

Some people say that those very computations are why the primates developed such big brains in the first place. Important in a cooperative society. Worth allocating the brain cells.

Now, back to the three D model I'm trying to implant in your brain. Your brain isn't a model railroad, it's a coffee filter. (Pause). The grains of coffee are the brain cells, a small number of layers, the hot water coming in the top is data from the outside world, sights and sounds, and the stuff dripping out the bottom goes to your muscles and glands to affect the outside world, or your body.

Now there are only a few, or a few dozen, but not many layers of grains, and you can only do so much computation with it. Sometimes the coffee comes out too weak. So what do you do? Pour it back into the top and run it through, over and over, until it's dark enough. Can come out bitter. If all you care about is is the color, you can run it through until you reach diminishing returns. You can get some strange brews that way, though.

Some of the gross structure is about that, some of it is just in individual connections that run backward through the filter. Biology is messy, and the layers and connections just growed under some general policies and guidance from the DNA. One way or another, it recirculates.

Instead of sending the command to your leg to execute the snap kick, you loop it around to the input as if you saw someone else do the kick that you intended. Now you can use a dozen layers or two to figure out what's going to happen next. Laws of physics, anantomy, how he might respond, and what you could do about that, etc. Then you put that back in the top.

Near the top of the filter, there is a double input, "needs kicking", and "you'll fall on your butt if you try a snap kick". Before there was just the one input, "needs kicking". This time, as it flows through the filter, it might trigger the spin kick, which might pass muster and get through to the muscles.

Same thing with words. Chop it off short of saying it, consider it, how would you react to hearing it, try again, again, good enough, spit it out.
Same with smells. Lilac, no minty lilac, no minty lilac with bergamon, it's beebalm. Like at Grandma's house twenty years ago, that summer that the ....
Same with maps, circle and arrow diagrams, pictures, melodies, quaternions, and things there are no names for in any human language.

Lots of bulletin boards, for every type of media. Not orderly, messy. Each bulletin board partially visible to many parts of the brain, connected strongly or weakly to many things. Refereed boards. Goal directed referees. "Attention" is the collective action of all of these referees. They know what is likely to yield good results. Judgement and forebrain, advanced lesson.

So that's it, consciousness and awareness, and all that. No big deal, all mammals and birds do it, possibly reptiles too. The internal monolog, and the movie running in your head -- just what's running on the buiietin boards, to give all the cells in your brain something to focus on.

Multiplexing, advanced topic. Same neuron participates in many computations, accoring to need, probably communicated through the support cells. Attention, activitly level on boards. I know this is pretty dry, and not relevant to most people. Thinking is a rare hobby, but I'm almost done, and will go back to thinking with the other head soon.

In neural nets, a whole weighty topic in computer science you could get PhDs for, not that there's anything wrong with that, learning happens through reinforcement. If there is a good outcome from what the net computed, then the connections involved in the computation get stronger. Devils in the details, but the principle is that simple. That's how we learn.

When you learn a new concepts, discover a new relationship, find a new connection, your brain decides if it's new and potentially useful. If so, it sends back a message that says "good trick, make more like it" to the general area of the brain cells that were involved. Biology is messy. And so the brain makes more tricks like it, recruiting nearby cells, extending connections, and maybe you say "still like it, make some more". Some of the new tricks you can describe in words, some you can't, some "you" aren't even aware of. Advanced reading, Roger Sperry. No matter, you can tell if you like them and keep running the cycle until diminishing returns sets in. People vary in how long they will run the cycle.