Glenn Watson

A single neuron takes a num­ber of inputs, each with their own weight. A activ­a­tion func­tion determ­ines if a neuron is going to be fired or as referred to in the icon above Static non lin­ear char­ac­ter­istic. There can be a num­ber of dif­fer­ent ones. This activ­a­tion func­tion will take as input the sum of the input weights against a bias. It will then determ­ine if the value is in the cor­rect range for a neuron to fire.

There are three com­mon activ­a­tion functions:

• A step func­tion can be used to emit only one of two val­ues if a threshold has been reached.
• A logistic func­tion is sim­ilar how­ever it will inter­pol­ate a smooth range of values.
• A Gaus­sian func­tion allows a gradual ramp­ing up and down around a cent­ral input.

The set of neur­ons get con­nec­ted into a neural net­work. There are some rules when con­nect­ing a neural net­work to make them use­ful is as follows:

• Have an input layer. The input layer takes inputs from the envir­on­ment that you are try­ing to test. Inputs for example for a robot in a maze may be the loc­a­tion in the maze.
• Have an hid­den layer.
• Have an out­put layer. These take data from the hid­den layer and then send out­puts to its limbs to change the envir­o­ment. For example the speed at which to turn the wheels, or the dir­ec­tion to turn.

The rules is that each of the input nodes is con­nec­ted to the nodes in the hid­den layer. Then each of the hid­den layer nodes is con­nec­ted to the out­put nodes. The idea is that when you get inputs com­ing in that the out­puts are sens­ible val­ues based on the situ­ation. That relies how­ever on the hid­den lay­ing hav­ing valid biases/weights etc. This is unlikely to occur when you first setup the net­work. These type of net­works are called feed for­ward and fully con­nec­ted net­work. It turns out you don’t need to leave out con­nec­tions because each neuron has a weight and if it turns out a neuron is com­pletely undesir­able a weight of 0 can be applied.

What you need to do is train the net­work. You take a ton of sample data. You attempt to eval­u­ate the data against a end res­ult. For example a robot suc­cess­fully nav­ig­a­tion a maze. Based on the suc­cess meas­ure we adjust the val­ues of the neuron’s weight and biases.

Out­side of games neural net­works have been used quite extens­ively. They are extremely good at recog­nising pat­terns. For example pick­ing up human hand­writ­ing when sort­ing envel­opes in a post office. Whenever you got a com­plex pat­tern you want to detect then neural net­works can be quite good.

In a game though for example, we could have a neural net­work that would steer a ship in the world so it could probe and land on a planet. We could provide inputs such as the ships head­ing and the plan­ets dir­ec­tion. The out­puts would be the dir­ec­tion we should take with the ship. We could test the neural net­work based on if we were head­ing in the cor­rect dir­ec­tion towards the planet.

Another example for a game could be for a sword fight­ing game. An NPC could get from the human player the human player sword pos­i­tion, human player shield pos­i­tion, human player crouch pos­i­tion, human player move­ment dir­ec­tion. The NPC responses can then determ­ine the height of the shield, and the way in which the NPC swings the sword.

In games how­ever, Neural net­works aren’t com­mon. One of the major issues is the amount of time it takes to train the net­work, when a finite state machine can achieve sim­ilar res­ults with much less time involve­ment. Some games how­ever have used neural net­work data but they would of already provided the pre-trained data and then switch­ing off the learn­ing beha­viour and then using a fixed con­trol­ler. The second approach is still to pre­train the game but leave the AI learn­ing turned on which will adapt to the player. The dis­ad­vant­age of this approach is from the QA point of view where the AI may become unpredictable.

One advant­age of a neural net­work for games is you can have emo­tions come into play inside the neural net­work. For example the NPC gets scared and will run away if its gets hit badly, or gets angry if only mildly hit. Inputs can be also NPC state of other NPC’s to cre­ate group behaviour.

Neural net­works is a top down approach to Arti­fi­cial Intel­li­gence. Top down mean­ing we think about how humans would achieve result.

In the brain there is no neuron that stores the inform­a­tion that our grand­mother is 80 for example. This is not how our brains work. There was a big debate in com­puter sci­ence should it be a sym­bolic approach or should it be the way our brain works where our brain changes into dif­fer­ent states based on what we know and don’t know. So every­one has dif­fer­ent con­nec­tions between neuron, dif­fer­ent num­ber of neur­ons and dif­fer­ent strengths. Should we build­ing a sys­tem that is sym­bolic or non-symbolic. A neural net­work is non-symbolic and there are no states as there would be in some­thing like a finite state machine. There are no vari­ables to describ­ing know­ing a par­tic­u­lar bit of knowledge.

We have in our brains about 100 bil­lion neur­ons. Each neuron has inputs to the cell body, these inputs are called Den­rites and they con­duct weak elec­trical sig­nals. The beha­viour of the cell as the inputs come into it, the cell mon­it­ors if the sig­nals com­ing in are meet­ing some strength threshold or not and if they do then one sig­nal through a axon will send out an out­put. These are con­nec­ted to one another in a com­plic­ated net­work and they are con­nec­ted in a syn­pase and from here they enter the den­rites of the next cell.

As well as going to receiv­ing inputs and out­puts from other neur­ons, some of them have spe­cial inputs and out­puts. The inputs con­nect up to our sens­ory organs and they said sig­nals about whats hap­pen­ing in our envir­on­ment (for example, optic nerve etc). The out­puts would be for con­trolling muscles etc which allows us to move for example.

The thing about human brains is they aren’t fixed at birth, you don’t get given one fixed brain where you know everything. As you take in input from around the envir­on­ment your brain chem­istry changes. The paths that get fired often are rein­forced and become more likely in the future to fire. Espe­cially if the same input con­di­tions are met. Some­times you are less likely to use some path­ways and these are less likely to be fired in the future. This can be a phys­ical thing, for example con­tinu­ously hit­ting a ten­nis ball, if you prac­tice hit­ting a ten­nis ball your brain chem­istry actu­ally changes and the path­ways for hit­ting the ten­nis ball get rein­force and are more likely to fire in the future.

Dif­fer­ent sec­tions of the brain are respons­ible for con­trolling dif­fer­ent parts of the body and are respons­ible for pro­cessing the dif­fer­ent inputs from the dif­fer­ent sens­ory inputs from the body. There are some­times times when a neuron fires it will slow down or decrease the like­li­hood that other neur­ons will fire and thats its job. It’s a inhib­it­ory neuron and it will pre­vent some­thing from fir­ing. Some­times the reverse is true, if a par­tic­u­lar set of neur­ons fire and that can encour­age oth­ers to fire. For example the neur­ons that are respons­ible for detect­ing heat, if you touch a hot object, the inhib­itor neuron will fire to pre­vent the neur­ons from fir­ing that caused you to touch the hot object. In reverse, those neur­ons may be con­nec­ted to the neur­ons that make your arm move away from the object and cause them fire more frequently.

Brains how­ever are not par­tic­u­larly fast com­pared to mod­ern day CPU’s. In terms of indi­vidual neur­ons, but there 100 bil­lion neur­ons work­ing in par­al­lel, they don’t have to wait for the next one to fire they can all work in uni­son. Most applic­a­tions on a mod­ern com­puter will only run on one thread, and have to wait around for one pro­cess to com­plete before mov­ing to the next one. This allows us todo things eas­ily that com­puters have alot of trouble with.

Artifical neural net­works are not a pure sim­u­la­tion of how the brain works. It’s based off work done by McCul­loch, Pits work in the 1940s and it was roughly based around the under­stand­ing at the time of how the human mind worked. Hebb only 6 years later found a way to make the neural net­work start to learn by chan­ging weights so you didn’t have to hard code the neural net­work cor­rectly in the first place. So ini­tially you can build the net­work with ran­dom val­ues and weights and give it data to allow it to improve at tasks over time. In the 1950s people actu­ally star­ted to attempt to make com­puters that were wired up based on the concept of a neural network.

Next art­icle will be more about actual imple­ment­a­tion of a Neural net­work in terms of com­puter science.

All mod­ern com­puters are based off a model that Alan Tur­ing cre­ated dur­ing the 1940s. He developed a the­or­et­ical machine that could run any algorithm and called it A uni­ver­sal tur­ing machine.  A uni­ver­sal tur­ing machine is not pos­sible with todays tech­no­logy because it requires a infin­ite amount of stor­age space. It’s import­ant to have a com­mon model for com­put­ing so we can under­stand the lim­it­a­tions of our pos­sible algorithms. Mod­ern PCs are imple­men­ted as a con­crete ver­sion of Tur­ing machines with lim­it­a­tions. So you can in the­ory run the same algorithm on any com­puter that is based on a Tur­ing machine and the meth­ods should func­tion in the­ory. So it shouldn’t mat­ter if its a mod­ern com­puter or a 1980s Macin­tosh, in the­ory they should all be able to run the same algorithms. Some PCs may be quick, some slow, you can in the­ory run the same pro­grams. There are some things how­ever that can’t be com­puted on any Tur­ing machine, if the prob­lem isn’t pos­sible to com­pute on one Tur­ing based machine its not pos­sible on any Tur­ing based machine. This includes the Halt­ing Prob­lem men­tioned in a earlier post. So one of things we want to know, is it pos­sible to use a Tur­ing based machine to com­pute intel­li­gence and is it pos­sible to write an algorithm that will give us rational thought or rational actions. So in the­ory it may be pos­sible to achieve these things but our com­puters are so slow that its just not feas­ible for such a com­plex algorithm.

One of the first ideas for com­ing up with Artifical Intel­li­gence was in 1943, which was a Arti­fi­cial Neural Net­work by Pitts and McCul­loch. It was a the­ory of com­pu­ta­tion back when it was inven­ted, i.e. they didn’t really have the pro­cessing power back then to actu­ally run the neural net­work. It was based off the human bio­logy of neur­ons fir­ing in our brains and based pro­pos­i­tional logic. It’s a attempt to rep­lic­ate how humans think. The idea is that a Neuron has a num­ber of con­nect­ing inputs. The inputs are numer­ical val­ues which can be mod­i­fied. The idea is that you can rein­force val­ues that are achiev­ing the cor­rect res­ult, and pen­al­ise (not allow­ing to fire again) inputs that are achiev­ing sub­op­timal res­ults. A very simple example of a neural net­work is a space invaders game, for the ship to avoid bombs. The out­put of the neuron can be if the ship should move left or right, if the shop is col­lid­ing with many bombs then we pun­ish the neural net­work and adjust the input weights accordingly. Minksy and Edmonds built the first neural net­work com­puter in 1951. The com­puter was actu­ally wired up like a neural network.

Back in 1956 there was a sum­mer research work­shop in Dart­mouth, research­ers such as McCarthy, Min­sky, Shan­non, Rochester, More, Samuel, Solomonoff, Sel­fridge, Newell and Simon atten­ded. These guys shaped the route of AI for the next 20 years and bey­ond and it was the first time that a research field in Arti­fi­cial Intel­li­gence was suggested.

Simon and Newell at the time of the work­shop had soft­ware which was a reas­on­ing pro­gram, Logic The­or­ist, that was cap­able of “think­ing non-numerically”. That is a way of manip­u­lat­ing ideas. Their suc­cess lead to a Hypo­thesis called the “Phys­ical Sys­tem Sym­bol Hypo­thesis”. This Hypo­thesis stated that A phys­ical sym­bol sys­tem has neces­sary and suf­fi­cient means for gen­eral intel­li­gent action. That is if you build a machine, like a com­puter or a series of buck­ets setup in the right way, and you set it up so it can manip­u­late sym­bols, that is enough so you rep­lic­ate gen­eral intel­li­gent action. Manip­u­la­tion of sym­bols is quite a dif­fer­ent way of achiev­ing res­ults to Neural net­works. Neural net­works don’t store sym­bols only numer­ical val­ues. So for example in a space invaders game, there is noth­ing in a neural net­work that would rep­res­ent one of the enemies on the screen or rock­ets on the screen. All a neural net­work has is numbers/weights and neur­ons, neural net­works are a non-symbolic approach to Arti­fi­cial Intel­li­gence. Simon and Newell approach was very dif­fer­ent because there would of been sym­bols for the vari­ous objects in the environment.

In the 50s and 60s research con­tin­ued and they were able to build pro­grams that could prove math­em­at­ical the­or­ems that most under­gradu­ate stu­dents would of found extremely dif­fi­cult. A learn­ing pro­gram was con­struc­ted by Samuel to play draughts which dis­proved the hypo­thesis that com­puters could only do what they were told. It didn’t just play the same game over again, it could actu­ally gain exper­i­ence by play­ing the game. This proved that com­puters can even­tu­ally do more than their ori­ginal pro­gram­ming. One of the early com­ments about com­puters was that they couldn’t ori­gin­ate any new ideas, they couldn’t do any­thing they weren’t told todo. If you imple­ment a learn­ing algorithm how­ever they can gain exper­i­ence and they can act dif­fer­ently over time.

Lisp which is a very pop­u­lar lan­guage for AI which was inven­ted by McCarthy. Its a func­tional lan­guage which is con­sidered one of the core lan­guages used in AI development.

Min­sky super­vised a bunch of stu­dents deal­ing with micro­worlds. To have a robot to nav­ig­ate in the real world is a really big chal­lenge. To con­dense the prob­lem down into micro­world allows com­puter sci­ent­ists to focus on smal­ler prob­lems and build­ing upwards. They were very con­strained, only a few objects in them, only a few type of oper­a­tions avail­able and only a few type of responses that were suf­fi­cient to describe the world. Terry Winograd’s PhD thesis in 1971 focused on hav­ing a robot per­form tasks based on nat­ural lan­guage (how humans would ask a task to be done). For example “Pick up a big read block”, and also being able to use the same nat­ural lan­guage to be able to query about the world, for example “How many blocks are not in the world?”.

Early com­puter sci­ent­ists had dif­fer­ent pre­dic­tions when com­puters will have intelligence?

Alan Tur­ing pre­dicted in 1950 that we’d have intel­li­gent com­puters by the year 2000 cap­able of passing the Tur­ing test..

Her­bert Simon (1959) thought it’d be within 10 years that a com­puter would be able to beat a human chess cham­pion and a sig­ni­fic­ant math­em­at­ical the­orem would be proven by a computer.

Ray Kur­z­weil (2008) said that in 2029 that both hard­ware and soft­ware can achieve human level intel­li­gence with a broad sup­ple­ness of human intel­li­gence includ­ing our emo­tional intelligence.

As I men­tioned in the last post, AI has a lot of found­ing in philosophy.

Aris­totle attemp­ted back around 384–322 BCE to come up with a form­al­ised set of laws for reas­on­ing. If you can make a set of rules based on how todo a argu­ment you can then use those rules to con­struct a machine to argue. Without rules its not pos­sible to con­struct a autonom­ous agent that can act ration­ally, or think like a human. Aris­totle also had a algorithm called regres­sion plan­ning that assisted in mak­ing know­ledge into action. For example, you regresse back­wards based on your goal. “I need cov­er­ing. I need a cloak”. So the first goal is that you need cov­er­ing, there­fore you regresse back­wards in your think­ing to “I need a cloak”.  This tech­nique doesn’t focus on the end res­ult, it focuses on the means of achiev­ing that result.

Around 1232–1315 AC a man named Ramon Lull con­struc­ted a machine based on a earlier Arabic device that could take two ideas and suc­cess­fully cre­ate a new idea. The machine would have dif­fer­ent arabic sym­bols that would rep­res­ents ideas, then based on cer­tain rules it was able to com­bine other sym­bols to con­struct new ideas.

Leonardo da Vinci designed one of the world’s first mech­an­ical cal­cu­lator in around 1452–1519. Shick­ard and Pas­cal both built (rather than just designed) cal­cu­lat­ing machines. that could per­form math­em­at­ical computation.

It wasn’t though until Hobbes in around 1588–1679 linked numer­ical com­pu­ta­tion to reas­on­ing. He pro­posed there was a link between numer­ical com­pu­ta­tion and reas­on­ing. Just as we today we are think­ing we can sim­u­late intel­li­gence on the com­puter now-a-days. Leib­niz took the concept even fur­ther and between 1646–1716 con­struc­ted a simple mech­an­ical concept manip­u­la­tion device. Remem­ber these mech­an­ical devices are with gears and cogs without any electronics.

There were two tra­di­tional views on the brain. Descartes who lived between 1596–1650 was fam­ous for the say­ing “I think there­fore I am”. His idea called Dual­ism was that there were two type of things for our brain, which is that the brain is mat­ter while the mind being non-physical. If you accept this idea you have to come up with some way for a non-physical mind can alter a phys­ical body but there is no clear indic­a­tion that this can occur. The altern­at­ive view is called “Mater­i­al­ism” and that what we think has to be related to some­thing phys­ical such as neur­ons fir­ing etc. Descartes believed that anim­als didn’t have a mind so there­fore were like clock­work machines.

Fran­cis Bacon (1561–1626) and John Locke(1632–1704) were try­ing to work out where know­ledge came from. Their approach was that all know­ledge has to come from the sensors. If you put a child in a black box with no stim­uli then they will have no know­ledge. David Hume (1711–1776) pro­posed through induc­tion that humans can work out a com­mon set of asso­ci­ation with objects. For example a light switch will all work in sim­u­lar ways and there­fore we have a under­stand­ing of how to turn them on and off. If we fol­lowed the approach that only know­ledge comes from the sensors we would have to have explore each light switch to determ­ine how it worked. So for AI agent the concept of induc­tion is crit­ical because we can’t pro­gram ever situ­ation the AI agent will encounter. We have to give it cer­tain com­mon rules about its environment.

Then came the notion  from Mor­itz Schlock (1882–1936) that we can make the­or­ies about our world, and cre­ate our own rules about the envir­on­ment. For example if you exhale viol­ently your likely to be breath­less. Even though you haven’t exper­i­enced this before. You can then use your sensors to con­firm or deny these the­or­ies. So we as humans are build­ing up our own internal rep­res­ent­a­tion of the world or model. So this is very com­mon way to build an AI agent, for example the AI agent can start build­ing a map intern­ally of the room as it moves around the room.

Charles Bag­gage (1791–1871) was the first to con­struct some­thing sim­ilar to a mod­ern CPU. Charles was quite the mad genius for his time. He had him­self lowered into a vol­cano he could observe mini erup­tions in pro­gress. His dif­fer­ence engine was able to cal­cu­late dif­fer­ent math­em­at­ical computations.

Thats it for this post, the next post moves away from Philo­sophy and more towards the elec­tronic com­puter age.

This art­icle is about what Arti­fi­cial Intel­li­gence (AI) is, and some brief his­tory points about it.

To find out what Arti­fi­cial Intel­li­gence is, you need to start out to identi­fy­ing what Intel­li­gence is. Intel­li­gence is the abil­ity to acquire and apply know­ledge and skills. And life is the con­di­tion that dis­tin­guishes anim­als and plants from inor­ganic mat­ter, includ­ing the capa­city for growth, repro­duc­tion, func­tional activ­ity and con­tinual change  pre­ced­ing death.  The defin­i­tion of life how­ever is hard to define.

Nor­mally at the moment we attrib­ute intel­li­gence to liv­ing things. We don’t attrib­ute for example intel­li­gence to rocks or the sky etc, although in ancient his­tory this what they might have done. This is the ori­gin of alot of religions.

There are four dif­fer­ent aims when build­ing a artifical intel­li­gence. These are:

• Thinks like a human.
• Acts like a human.
• Thinks ration­ally.
• Acts ration­ally.

When com­puters were first intro­duced around about 60 years ago, people were mis­con­ceived into think­ing they were a arti­fi­cial brain. The reason for this is because they could things like cal­cu­late tra­ject­or­ies of shells much quicker than a human being. The real­ity is there is no intel­li­gence behind these actions and the com­puter is just being a fancy calculator.

One approach is to make the com­puter act as a human brain does. Humans con­tain soft bod­ies with cells and chem­ic­als that per­form the intel­li­gence. We can’t achieve exactly this in the com­pu­ter­ised world but we could sim­u­late a human brain and body. The other main approach though is not to worry about how the machine thinks, but get the out­come to be the same as the human. For example, one per­son may use a sheet and paper to work out a proper, while a com­puter algorithm uses com­pu­ta­tional the­ory to solve the prob­lem. The solu­tion is the same in both cases but the way the solu­tion is solved is done in dra­mat­ic­ally dif­fer­ent ways.

So in AI one ques­tion is are we con­cerned so much in how humans think, or are we con­cerned about rep­lic­at­ing the res­ults of the out­come of the human thought pro­cess. Both ways of think­ing about the prob­lem are equally note worthy.

When people star­ted out with AI it was pre­sumed in approx­im­ately 50 years they could get a com­puter to be more intel­li­gent than a com­puter. Gen­er­ally AI is focused around a fam­ous quote “The study of how to make com­puters do things at which, at the moment, people are bet­ter”. So when they star­ted research­ing AI one of the tasks they decided to rep­lic­ate was chess. They choose chess because in humans it requires a great deal of intel­li­gence to play chess. How­ever, they found com­puters can use a approach which requires no intel­li­gence, which is a brute force tech­nique that tries every move pos­sible. The com­puter can make these cal­cu­la­tions much faster than a human. So its not the fact that the com­puter is smarter than the human, its just that the com­puter is smarter at cal­cu­lat­ing the moves. So you can’t expect a human to try all the dif­fer­ent com­bin­a­tions because there are poten­tially bil­lions but com­puters have actu­ally done these type of com­pu­ta­tions. The inter­est­ing thing though is that Chess was determ­ined as the great determ­in­a­tion if a com­puter was intel­li­gent, but after the com­puter star­ted beat­ing humans at chess com­puter sci­ent­ists redesign their mean­ing of AI because it was obvi­ous the com­puter was not intel­li­gent in the human sense since the human is still able to per­form many tasks bet­ter than the computer.

The hard­est things todo in AI are some of the things earlier com­puter sci­ent­ists thought would be easy todo. For example, hav­ing a com­puter fly an air­plane with the skill of a com­mon house­hold fly is impossible. At the moment its thought that cock­roaches have been accur­ately sim­u­lated as a intel­li­gence by the com­puter, but small mam­mals like cats and dogs have not been.

So how do you know if some­thing is act­ing like a human? Turns out its extremely hard to tell. Alan Tur­ing came up with a test for determ­in­ing if a com­puter could act like a human. He came up with the idea of two black boxes, each with a ter­minal on top. One had a human inside the black box respond­ing to quer­ies from a user out­side the boxes, the other with a com­puter inside but the out­side user has no idea which is which. So the idea is the user will type in a ques­tion at both ter­min­als. If the user can’t tell which one is the com­puter and which one is the human than the com­puter has said to have passed the Tur­ing test.

To pass the Tur­ing test there are a num­ber of things a com­puter must be able todo. It must be able to pro­cess nat­ural lan­guage, it has to have some way of rep­res­ent­ing know­ledge so it can respond with a respons­ible answer, it needs to be able to reason about things for example to have a argu­ment with it, and it has to be able to learn over time. For example with the learn­ing if you tell the com­puter your name and then later ask the com­puter your own name then it might be a good sign that it is not a human.

A autonom­ous agent is a agent inside the com­puter that is able to per­ceive itself in some way about its envir­on­ment and it’s able to adapt to changes in its envir­on­ment. A autonom­ous agent is usu­ally able to set goals. For example a agent may have the goal to explore its envir­on­ment, to defend itself etc.

Another import­ant con­sid­er­a­tion in AI, is to ask the ques­tion is it pos­sible to have the com­puter rep­res­ent a intel­li­gence in some way. For example in com­puter sci­ence there are prob­lems which have been proven not to solv­able. For example the halt­ing prob­lem, where you can detect if a pro­gram is going to halt, or keep loop­ing. At one point though they received mil­lion of dol­lars to attempt to solve the prob­lem even though it had been proven it couldn’t be done. You need to make sure there is no proof that a prob­lem is unsolv­able so you don’t go research­ing an entire branch of know­ledge without jus­ti­fic­a­tion. At the moment though there is no known such proof at the moment for AI.

At the end of the day, we are just a chem­ical reac­tion with neur­ons fir­ing. How does that relate to intel­li­gence, this one of the aven­ues where alot of AI research is going into (and med­ical for that mat­ter). How is it the thoughts in our mind allow our body to manip­u­late our environment?

Next post I will talk about the his­tory of AI.