What is AI
so now we're starting a unit on artificial intelligence I'll be talking a fair bit about artificial intelligence in this course so far particularly in the use of chat GTP in large language models but now we're going to start exploring things in more detail and you'll be creating for your next portfolio item your own chatbot so what is artificial intelligence well it's a whole range of Technologies but essentially they've been focused on replicating human capabilities in particular around human intelligence so we used to see things such as computers being able to play chess or being able to interpret words and language and things of that nature so these things have been mastered quite a number of years ago now and we're slowly improving the range of capabilities we have with computers to be able to replicate what human beings can do now someone reached a point whereby we now accept that artificial intelligence or computers will think and do cognitive tasks in different ways to human beings and that intelligence can be conducted without necessarily doing things exactly the same way as humans think now in terms of um becoming equal to human intelligence which is another area that people are always concerned about in general where the computers have exceeded that um for quite a while now there was a fundamental test known as the Turing test whereby if a computer could have a conversation with someone and another human being had a conversation with that person and that person couldn't tell the difference or couldn't work out which was the computer and which was the human being that they were conversing with then the computer would be considered to have met the artificial intelligence requirement to be considered of equal intelligence with human beings yeah there's still some debate as to whether or not that's occurred but for a vast majority of topics and conversations uh computers can now conduct those at an equal level of human beings and indeed every day millions of people are conversing with computers on phone helplines and support desks without realizing necessarily that they are talking to a computer-generated person other than rather than a human that they're talking to so artificial intelligence is really a collection of techniques and tools that are used to establish some level of intelligence now the three main ones that relate to education are machine learning neural networks and natural language processing now there are other Technologies and techniques as well there's artificial intelligence around vision and image processing being able to detect faces and things of that nature there's quite a bit of work around Robotics and how computers can mimic and utilize the way the humans move or animals move things of that nature so there are a range of other aspects of artificial intelligence Beyond those three they've mentioned but they're the three that we focus on most in education so the first of these is machine learning so this is where computers learn things and they generally do it in a somewhat different way to how humans learn now in the main humans learn through processing experiences so imagine every young child a baby their looking at their parents they're looking at the world around them they're slowly building up an understanding of um others and of themselves and of the interactions with the world around them now that's done primarily because we're autonomous we can move around we can experience the world in general computers can't do that so we have to provide them with information about the world robots are exception to that but in the main we train computers around machine learning where they don't have necessarily sensor information about the world we provide them with information about the world and the way we do that is providing them with lots and lots of information Millions upon millions of bits of information far more than a human being would ever be exposed to in that sort of time frame and from that they're able to learn about various concepts by just processing vast amounts of information so there are three main approaches to doing this one is called supervised learning now this is where we plan and structure a learning process and we have specific data sets and the computer is trained on how to respond to various data now in to a certain degree you did a similar supervised learning process when you created your twine Choose Your Own Adventure game you trained the computer the game to be able to respond in various ways depending upon certain inputs and there would be certain outputs in terms of the descriptions and content of the game as the person played it now that's a very simple example of what's known as a decision tree um so to a very small extent you were training your game to think to respond to various stimulus and have certain outputs as a response to those stimulus now imagine doing the same when instead of just having a couple of dozen decisions you had tens of thousands of decisions or millions of decisions say all the different possible questions someone could ask a doctor and this is where we have things such as expert systems you could train a very comprehensive choose your own Venture type game to answer any question any medical question now the problem with this is it takes a long time to build such a system as you found in creating your simple system if you're going to create one with Millions upon millions of different possible questions that would be a long and liberal risk task now that was certainly has been attempted and there's been a number of very comprehensive expert systems developed including in the medical field but in the main we've moved on to a more efficient way of doing things which is unsupervised learning now with unsupervised learning what we do is we provide the computer with lots and lots of data lots of information and we allow it to look for patterns and to try to come up with an understanding of what all that data means by exploring those patterns so for example we can provide with lots of information um about conversations patients and doctors have and by looking at Millions upon millions of such conversations and the patterns that emerge from that we can then ask questions of such a system as it's being trained and it will respond similar to how patients and doctors have responded to each other from the data and it turns out that that's quite an effective process and by just providing lots and lots of data and having the computer look for those patterns it's able to then create a very effective system now the third process is known as reinforcement learning now this is applied generally to the other two processes but essentially once we have an existing system in place we can improve upon it by having people ask questions and then responding as to whether or not that provided an effective answer or not and if it did we then strengthened the various weightings to respond in a similar way in the future whereas if we had answered it incorrectly or the computer answered it incorrectly we tell it that and then it will be less likely to respond in that way in the future and by having again Millions upon millions of interactions and getting that feedback the system then learns and improves to where it then provides a better process of responding in the way we're attempting so for example with um machine Learning Systems in automatic autonomous vehicles all of the camera data from all of the autonomous vehicles or semi-autonomous vehicles that they currently are are all being fed back into these systems and providing data to reinforce um the training of these autonomous Vehicle Systems so where the vehicle has been able to successfully navigate without the user taking over then that's being the successful interaction and it reinforces that behavior where the drivers had to intervene and stop an accident occurring or something like that then that particular approach would be de-emphasized and so the system would then be less likely to do that in the future and again by having millions upon millions of these interactions the system trains itself and learns to become more efficient and effective in whatever task is being undertaken so human beings do undergo similar training processes but not necessarily within our education systems we might slowly learn how to write essays based upon positive and negative reinforcement behaviorism and we receive some marks and we learn not to do the same thing again but that's sort of only on a few dozen examples now we're very good about learning from those few dozen examples but it's not like we were doing a million essays and then learning the tiny incremental aspects about a computer uses around machine learning to improves itself okay so the next process then is alluding to some of the things I was starting to talk about is using neural networks now in our human brain or in any brain we have a series of neurons which are um send you their at the start of it is like they're like little tadpoles but you have the head which can receive information from other neurons and then you have a whole series of tails that go out to other neurons and that will cause them to trigger or not trigger and all of these form a complex Network now each of these neurons has a certain efficiency as to whether or not it will send information out along it or not and Trigger another neuron to do something called set of neurons and we call that a waiting so if it's highly if it's a very thick neuron for example then it will be more likely to send information along that pathway if it's very thin it's less likely to send information along that pathway now how we learn is that when we're successful in learning then the neuron is thickened and if something is unsuccessful if we don't learn it or we we know we learn that it's wrong or we just don't use it for a long time then the neuron will get thinner and so it'll be less likely to use that particular pathway so essentially that's how the thickening and thinning of neurons um provide weightings that determine whether or not a particular pathway or not is conducted now no one concept is made up of a single neuron generally there are thousands upon thousands of neurons that are involved in any sort of con concept that said however if artificial neural networks we have been able to show that a very small number of neurons can perform very complex tasks indeed we've been able to do similar things with very simple organisms that only have a few neurons in their brains such as fruit flies and things of that nature now one of the examples I'm going to get you to do is using the teachable machine is to create your own neural network your own little brain and you're going to show it a series of pictures of cats and a series of pictures of dogs and you're going to then have a train so that when it sees a picture of a cat or a dog it'll be able to determine whether or not it is a cat or a dog that is seeing now essentially that's how our brain works around all the things that we do we train it to respond in different ways based upon different inputs different stimulus so there are a few things we need to think about in terms of the neurons we have our input layer which might be the different um different cells in our eyes move up quite complexize we've got millions of cells in our eyes but very simple organisms might only have five or six cells in their eyes and depending upon which of those cells is receiving light it depends upon the image that it builds so that's a little bit like how an um the input layer of an neural network works it's receiving some information and there's going to be some output we're either going to run away or we're going to detect it as a threat or something good to eat or something that we like or something that we dislike something that we associate with our parents or something that we associate with the color blue all of these different things can be slowly built up in our neural network which encompasses our whole brain with Millions upon millions of these neurons but what we've discovered is that there is an intermediary set of layers we don't just go from a series of inputs to some outputs we have what is called intermediary or hidden layers and by having a network of those that build up different weights and strengths um that's where the learning actually occurs so you're going to see that when you create your own little brain using the teachable machine and by showing it 10 pictures of a dog 10 pictures of a cat and allowing it to train by establishing the differences between those pictures and the similarities it will then create a neural network that when shown different pictures will make a determination as to whether or not they are a dog or a cat now you could make that determination on anything it could be determining whether or not it's your friend or if it's a stranger depending upon if you've trained it on pictures of your friends and pictures of strangers or it could be pictures of handwriting is it a a picture of a letter A or the letter c or the letter d and by putting in 27 of those and training them you can then have a handwriting recognition system so you could write out messages and the computer would then determine whether it's seeing different letters the same we can have with our voice so we can try to have it determine what sounds and what words we're saying and we can then do that for multiple languages and have translation machines so very simple processes can be used for very complex ideas it's being used very successfully in medicine at the moment for detecting things such as Skin cancers where we can show it millions of pictures of um skin and let it know then which ones had cancer and which ones didn't and by training it and going through this neural network it will then be able to look at a new image of someone's skin and make a prediction as to whether or not it has skin cancer or Not skin cancer or any sort of diagnosis so the third main aspect around these AI tools is natural language processing now humans developed the ability to speak and have language fairly recently in terms of evolutionary processes and most animals don't have this and it allowed us to greatly expand the complexity and capabilities of our brains because we now had a mechanism whereby we could think about things in much more complex ways than if we just had our neurons having to fire through and so forth we now had a language that we could actually associate with different um neurons in our brain and Associate those words and concepts with other words and Concepts and by communicating with one another that then built another whole level of complexity around human thought so what's involved though in natural language processing is actually quite complex of course words and language is actually much more involved than just learning an alphabet or a set of words and what those words mean many of our words can be broken down into smaller aspects which we call tokens and we can rearrange these tokens to make infections and emphasis and a whole range of other aspects around language another aspect which is not included in natural language processing normally is gestures and facial expressions and things of that nature that is now being Incorporated but engaging with visual systems as well as audit natural language processing so natural language processing essentially allowed us to have computers be able to speak and to understand human language which took a long time and was very involved and slowly improved as many people who have used automatic translators have seen the systems have been slowly improving and now they're getting to a human level quality indeed we can now have it so that we can have our language translated we can speak in one language the computer will interpret what the meaning is around what we're saying and represent that out in another language so computers are able to do this and we're now also looking at how computers can use this to help understand um processes themselves and this leads us into large language models which have been the current sort of revolution in artificial intelligence now we've been working on large language models for many years and slowly seeing them improve in fact not that long it's really only been the last 10 years but building up off that idea of gathering in unsupervised learning from huge amounts of data and looking for patterns and using neural networks to be able to incorporate those patterns into a learning process and then using the ideas from natural language processing particular and tokenization and breaking down Concepts into smaller and smaller bits so that the computer can then regenerate those as some sort of conversational language we reached a Tipping Point in 2020 late 2022 early 2023 whereby we now had a level of complexity where it could be generalizable so instead of making a very specific model such as your twine game we can now make a model where we can ask any question or have it um respond in many many different ways this is called generalizability which is what made computers so powerful initially the original computers were very very specific they were mostly designed around um very large-scale sewing machines what were called looms which would weave carpets and and cloth but they could only do one pattern and the invention of the computer what was called the generalizable computer was that instead of just being able to do one pattern we could provide it with instructions to be able to do many different patterns and then we worked out but we could not just create a pattern of weaving we could do it use it to do some mathematical calculations or we could do it to sort some um a list of information and then slowly we could use it to do a word processor or to make a simple computer game and over time we've developed more and more applications for the use of computers because a computer can be used to solve lots and lots of different problems it's generalizable and that's where the large language models sort of reach their Tipping Point so from being able to do very specific things very very well such as detect cancer now we can have it not just detect cancer from lots and lots of images we could say detect which student is likely to fall asleep or to do well on their exam based upon video monitoring of how well they're paying attention in class collecting Millions upon millions of bits of data and processing it through and being able to look for patterns in that data that's what makes large language models so powerful and we'll discuss this in the tutorial and how you've created your own dual network with the thinking machine and some of the implications that has then for what you're going to do with your own chatbot and we'll look at all of this in the tutorial
2023-05-01 01:28
