now let us get into our next chapter deep learning so basically deep learning uh is a branch based on machine learning so it is like kind of a subset of machine learning so probably what you can tell is it is a one of the main subsets of machine learning again machine learning is subset of AI okay so that's how uh deep Le learning will come under it okay so basically I would like to discuss about something known as neural networks so as uh human brains are actually having completely something known as a neuron okay and there is always interconnection between uh the neurons with respect to all the neighboring or you can tell layers of neurons okay so these are completely interconnected so with this uh ideology itself we are trying to create artificial intelligence so that's why with respect to the studies we are going to have our architecture in the similar way okay so after uh the ml we the Deep learning to existence okay so the Deep learning is kind of having the neural networks that imitates the human brains and so deep learning uh will do okay so it will be imitating same as human brains so in deep learning nothing is program explicitly similar to how we have seen in every uh machine learning algorithms so basically it is a machine learning class which makes numerous nonlinear processing units so as to per perform features extracting as well as Transformations okay so it needs to extract features as well as do transformation of that okay so the output of each preceding layer okay so as I was telling uh there are kind of many layers in our brains itself so similarly neural networks also we will be having multiple layers okay so uh I would like to break it down into three main categories right now okay so it might be more which we'll be seeing later so basically uh the layer which takes the inputs are called as input layers and the layers which gives the output are called as output layers and every each and every other layer in between the input layer and output layer are known as hidden layers okay so here they have like kind of fully connected so from one layer to another layer every uh neuron is interconnected okay so might not be for the the third uh layer okay so first layer and third layer might not be connected but uh first layer and second layer every consecutive layers will be completely interconnected okay so that is something know as fully connected Network which will be again uh discussing more a bit later so that is what exactly uh neural network is made up of and this type of uh learning uh we categorize it as deep learning okay so deep learning models are capable enough to focus on accurate features themselves by requiring a little guidance from the the programmer okay so we have to give a little guidance of what exactly is going to happen again it's kind of similar to machine learning algorithms itself we'll have output layers again if it is like kind of classification problem we can give uh output layers as two different classes where it needs to end up okay so either uh in one neuron or you can tell in uh here the term which we use for neuron is actually node okay so it should end up in one node or it should end up in other node okay so these are actually somewhat similar to decision trees okay so decision trees based on uh one or two one decision it will go to one or two branches okay so left Branch right Branch but here it has so many layers which are hidden layers input layers like in between input layers as well as the output layers so all these are interconnected So based on its assumptions it will start jumping hopping around in different layers okay and then it will finally reach the path to the output layer where it is expected okay and also it has very high accuracy too so deep uh learnings uh sorry let me uh add another statement so uh as I was telling we the accuracy or the uh deep learning models have enough Focus for accurate features by themselves by requiring a little guidance by the programmer and also are very helpful in solving out the problems of dimensionality okay so whenever you have multiple Dimensions you have more Dimensions just give more input layers that's all need not to worry at all about the in inside layers okay so again uh we have to do some kind of pre-processing steps similar to pre-processing steps here too since it is again a sub branch of machine learning itself so here what we can tell is uh we are making a machine learning algorithm towards a generalized case okay so in the previous uh chapter we had seen differences between machine learning and uh artificial intelligence right so artificial intelligence works on more kind of a generic reasoning as wide range of Scopes whereas machine learning performs for a specific task okay so here we are trying to make uh machine learning model into more uh generic okay or you can tell more uh more number of specific task it can handle okay so that's how we have done and also uh with the help of whole lot of uh uh kind of layers and all it will help it will be boosted much more towards accuracy okay so deep learning algorithms uh are used over here especially when they have huge number of inputs and outputs okay so we'll discuss about algorithms uh in the next video so again uh since again the Deep learning has evolved by the machine learning itself so since it is a subset of artificial intelligence so the fundamental idea behind the Deep learning is nothing but the artificial intelligence itself okay so basically artificial intelligence to mimic the human behavior Behavior okay so the idea of deep learning is to build such algorithm that mimics the human brain okay so that is an agenda of deep learning so as mentioned before deep learning is implemented with the help of neural network the idea of motivation behind this neural network is our biological neurons so which is nothing but our complete brain cell itself so how they are connected everything so that is the uh idea of motivation over here with respect to the neural networks or the Deep learning which is implemented with the help of neural networks coming to the definition of deep learning so deep learning is collection of statistical techniques of machine learning for learning features hierarchies that are actually based on artificial neural networks okay so here we are addressing neural networks as artificial neural networks because it is human-made okay so that is how uh deep learning is defined as and basically deep learning is implemented with the help of deep networks okay so here as I was telling there are hidden layers okay so these hidden layers forms a networks so that's why deep uh networks uh is nothing but neural networks with multiple hidden layers okay so that's how deep networks and deep learning came into existing okay so the the term uh the idea behind the term has come up in okay so here I actually don't have an image I wish I had so basically uh I can fetch one of the images okay so I can think uh this is uh visible okay so here we have whole lot of uh humans okay so here we are trying to do face recognition Okay so face features so you can see these are my input layers for each individual uh person we can have an uh input layer so basically different kind of images so then you have uh kind of few of the pre-processing which are connected okay so here again you can you will be like we'll be seeing uh in detail about these steps later on okay and then you will be having hidden layers okay so you have hidden layers hidden layers then it will come and come to a conclusion okay so here output is two so maybe we can have a conclusion whether a face is recognized as a uh I think the faces over here are kind of uh with respect to a male or a female okay so something like that so final output is that so there is whole lot of feature faces uh features which are applied over here in the hidden layers okay so uh like here in this example like we actually provide raw data for our first layer okay so here we actually completely provideed raw data itself and then it is actually uh creating whole lot of pre-processing steps so these Pro pre-processing steps for this example would be like uh basis of colors uh Luminosity Etc okay so then we have the first layer second layer which can actually work on maybe uh in detail about faces features face features okay so I think that's what they have mentioned over here with respect to face features so here face features can be like Eyes Nose Lips Etc and then it can go up to the third layer where it actually determines uh the correct face itself okay so here uh then it will be sent to Output layers and even in output layers it will be having a few layers interconnected also so that's why here with respect to Output lers we have few interconnects okay and finally we'll be ending up so usually uh input layer might be only single layer also and output layer might be a single layer also everything in between will be hidden layers okay so generally this uh like is not for every single case Okay so that is what I wanted to tell so that is about a quick introduction about uh deep learning uh now let us see about few architectures types of architectures present in deep learning okay so deep learning uh it is a neural network that incorporates complexity of certain level which means several number of hidden layers are uh ENC composed between input layers and output layers okay so all these hidden layers are in between uh encompassed between the input and output layers so they are highly uh proficient on model and processing of nonlinear associations Okay so very uh very efficient okay so basically the outcome is very profitable so uh that is that is why we use the term over here as proficient over here for uh models and also for processing nonlinear associations okay so we had already seen like linear regression then we seen logistic regression and all so those are kind of dealing with linearities and all okay so the Assumption over there for those algorithms is like it is linear properties the independent variables and dependent variables has kind of some kind of linear properties okay so they are actually proportional to each other okay so here nonlinear properties also it can easily work okay so that is about deep uh neural networks which is again uh the same thing what we had seen earlier okay of the image also that is also a deep neural network okay so next we have uh deep belief networks okay so a deep belief network is a class of deep neural network that comprises multi-layers belief networks okay so what exactly here uh happens is like kind of with the help of Contra contrastive Divergence algorithm so it is one of the algorithm in short it is known as there is no acronym for this so not D CD no no so you can just uh recall it as contrastive Divergence algorithm so a layer of features is Leed from learned from perceptible units perceptible units so with those layers it'll be learning the features and next uh the formally trained features are treated as visible units okay so all the trained features are kind of treated as visible units which performs learnings of features okay and then lastly when the layers are kind of final hidden layer like finally when the learning is uh complete final hidden layers are accomplished uh then the whole uh deep belief network is uh created okay so that is kind of the steps what are involved so basically it will take one of the algorithm then a layer uh layer of features is learned from its Preble units and then uh it'll form trained features treated as like kind of visible unit which can be uh performed uh for learning of the features and finally it will be after uh when the learning is uh done the final hidden layers is accomplished okay so that is how the uh deep belief networks work and next we have uh recurrent uh neural networks so in short they are also called as RNN so the these are also this type of architecture also is widely used so basically what exactly happens here is it permits parallel as well as sequential competition okay so that is the major advantage and uh since it is like kind of parallel even we actually do multitasking okay so we can be uh like having vocal voice uh we can be speaking in the same time we can be moving our hands legs uh also doing just gestures with them and also bling ining eyes and automatically our systems are working like kind of respiratory system which is continuously breathing in and out okay so all these kind of kind of uh features which are happening in parallel okay so that's why uh recurrent neural networks permits actually parallel as well as sequential Computing also so uh it is like exactly similar to a human brain but with a large feedback network of connected neurons okay so here it will have a whole large feedback Network connected so basically we will be having feedback elements also which will help a lot with respect to the learning aspects and final uh prediction and also with respect to how it can start acting as a actual human brain or actual intelligence okay so since they are actually capable enough to uh remin all the imperative things related to the input they received so they can be more precise about it okay so that's how like the feedback Network everything goes in so that's how they'll be learning so these are the few uh deep learning architectures next we have actually types of deep learning architectures okay so uh these would be give uh much more uh understanding with respect to deep uh learning networks so first we have feed forward new uh neural network so if feed forward neural network is none other than an artificial neural network it's so same as a neural so what exactly is happening is we are feeding towards uh forward okay so same concept over here it is actually going towards forward itself from input to Hidden layer hidden layer to Output layer like multiple in layers then output layers okay so cycle is actually forward okay so what I mean is it ensures that the nodes does not form a cycle okay so not kind of a loop back okay so it is same as artificial neural networks so it is kind of neural network all the uh precept are organized within layers such that the input layer takes the input and the output layers generates the output okay so the hidden since the hidden layers does not link with the human uh world or the outside world it is named after that itself like kind of hidden layers okay it is not visible to us but definitely we have the control of creating all the hidden layers how many it should be we can we'll be deciding those things also okay so each of the uh preon contains in one single layer uh is associated with each node in the subsequent layer okay so each and every node as I was telling before each and every this one will be actually connected so here uh the image is not showing it is connected till the end so each single node over here will be connected to the end node also okay so every node on the next layer okay so that's how it will be the next PR uptron so I just I think I just went forward okay so let us continue so each perceptron contain in a one single layer is associated with every node in the subsequent layer so it can be concluded that all the nodes are fully connected okay so as I mentioning in the image also it is not completely shown over here as fully netor fully connected but definitely in a deep Learning Network it will be fully connected okay so sometimes we call them as ful in Connected networks too okay so it does not contain any visible or invisible connections between the nodes in the same layers okay so it does not has any connection between the same layers in visible or invisible okay so there are no back Loops uh in this type of feed forward uh Network so as I sing it goes only towards forward and to predict uh like kind of uh sorry to minimize the prediction error uh back propagation algorithm can be used to update the uh weight values okay so basically here we'll have some kind of weight values uh which are given to the input layers okay so we will be having weight values so after getting the output uh layers from here we'll measure up those weight values and also send them okay and also we have something known as bias values also so basically U these are the things which will be uh helping out with respect to the input layers and all so that's why we if at all we need to minimize the prop uh prop prediction error then we'll be doing back propagation algorithms such a that we are just going to modify the weight values and send them okay and application of this would be like kind of uh data compression definitely uh data like uh nowadays data can be compressed like anything and it can it can be extracted back to to its original size okay so like uh feed forward networks help a lot in that and uh pattern recognition in any kind of uh applications so anywhere there is pattern you want to do some recognition you can do and we have computer Visions as we had seen in the image itself so it is trying to categorize it is trying to recognize faces and then we have sonar Target recognitions okay so with respect to sonar waves also it can figure out with respect to Target uh recognizing the Target and then we have speech recognition Okay so basically uh we will be having inputs as our uh words and then like what exactly is our intentions and all can be recognized and uh another application is like handwritten character recognition Okay so uh with respect to fonts and which are present on our devices okay electronic devices it is much easier for the electronic device to understand which kind of uh font it is using or which character it is using based on that font okay so it has predefined data but when it comes to human uh being writing in handwritten so that font uh might not be same and also every single uh digit might change even if you repeat yourself okay so for this reason like for this reason it is actually needs to do whole lot of calculations okay so that calculations will be done by our can be done with the help of feed forward neur networks so those are kind of few examples for our feed forward next we have RNN we have already seen about the type so it is uh it is actually at another variation of feed forward itself so here uh each of the neural network is present in the hidden layer receives an input with a specific delay in time okay so they will have some kind of DeLay So the NN M accesses the preceding info of existing iterations okay so it'll just Fetch with respect to it will access mainly the preceding in informations okay so maybe for an example like uh there is a guessing of a word in a sentence okay so one must have knowledge about words uh that were previously used already okay so with the that only by using words only we have like kind of uh developed the skill of putting them together making them as some sense okay so that's how even our artificial intelligence should develop okay so based on those kind of preceding values only or like kind of in the existing iterations so then only it will be able to figure out what exactly it is like uh figuring out a word in a sentence okay so it is not only the process of inputs but also it shares the length uh as well as weights Crossway times okay so it does not uh let the size of the model to increase with the increase in the input size so even if input size is more the same number of hiden layers would be sufficient if it is already having very high Precision okay so only thing is the internal layers get better and better not wider or in terms of size okay or bigger something like that it Still Remains the Same okay so even us like kind of after a certain age even our human brain stops growing so basically uh but the networks after stop growing also we won't be creating new neurons or anything like that there is concept of reproducing cells and also so I'm not talking about that but number of uh neurons which are interconnected so but our uh knowledge keeps on increasing as we learn new things so same way over here uh recurrent neural networks also like kind of uh learns new things and it does not increase the size okay with this increase in size of inputs uh inputs okay so every know what we consume is also input to us to our human brain so however the only problem with this uh recurrent uh neural network is that it has some kind of slow computational speed okay so it has parallelism but definitely it is slower due to the delay which we discussed in the beginning okay so uh based on that it has slow computation as well as it does not uh contemplate any further inputs for concurrent States okay so it has a problem with uh reminiscing prior information so examples we have whole lot of examples we can tell over here machine translation robot control time series predictions speech recognition again speech synthesis uh Rhythm learning music composition I think all this comes under RNN so here I was telling like slow also like kind of robot control it should be very fast but our systems are very fast it can operate like Millions of operations per second okay so basically Millions can be converted into thousands for every second even if it is calculating thousands of calculations it is actually lot faster than us okay in terms of uh speed giving results and all but uh definitely with respect to wide range of other aspects it is still not uh perfect so we are still at V and general a okay so we we have not reached at strong a even our human brain is much more capable than what we are doing right now so according to studies they actually say we actually use only 15 to 20% capability of our human brains okay so uh using 100% you can understand what capabilities we can have okay so here it is same even neural networks might have much more capability but still the research is going on the uh development of it is going on by every day okay so next we have convulsion neural networks so again uh if I want to talk about convolution neural neural networks I can go on and on for hours so here let us cut it short and keep it very simple so conversion neural networks are a kind of special kind of neural networks mainly used for image classification clustering of images and object recognition so basically any classification problems converion network uh is one of the very uh appropriate uh neural network you can select for and uh deep neural networks enables unsupervised constructs for any kind of hierarchial uh image representing tool okay so to achieve the best results uh deep convolution networks are preferred more than any other uh neural networks okay so always remember if you're going with any kind of classification or even clustering of images okay so the image clusters like you don't have classes predefined okay so you can just tell give me five clusters of images something like that okay so even those things and object recognition all this convolution neur networks are there I think I am already giving you whole lot of applications so identification of faces street signs uh even in medical Fields like tumors image recognition video analysis and also uh I was mentioning about uh natural language processing right so the uh artificial intelligent tries to understand how we humans naturally communicate in the same way it also can try to communicate okay so that can also be achieved by conversion neural networks and also so like kind of gaming uh which is like kind of checkers games chase games all this can be easily built on it okay so next we have uh restricted boltman machine so RBM for short so rbms are not at uh are nothing but another variant of Bolt spin machine so basically here what happens is the neuron neurons present in the input layers and the hidden layers uh encompasses symmetric connections okay so they will have symmetric connections to Amit them so however there will be no internal Association within the respective layers okay so in but in contrast uh restricted bolt spend machines uh do Encompass internal connections inside the hidden layer okay so these restrictions help uh uh bolman machines like uh to train um the model much efficiently okay so these are kind of used in kind of feature learning filtering uh risk detection uh business and economic analysis okay so that's how it goes on again we are not diving deep into this okay so not that significance for this uh course okay so it is little out of scope for this course okay so next we have Auto encoders so again uh an auto encoder neural network is another kind of unsupervised learning algorithm so here the number of hidden layers can be merely very small than the input layers input cells itself okay so the input cells can can be very less so here you are actually doing some kind of process known as encoding and it is done automatically so hence the name is also known as Auto encoder okay so basically uh what you can call the encoder as it is converting inputs into lower Dimensions okay so if it has more Dimensions like kind of three four five 6 10 features it is converting into 2 three four features okay so it is reducing by doing some kind of encoding okay so where one feature itself like in whatever the hidden layer is getting right the neuron so individual nodes uh whatever they are getting it the uh data is already encoded of the multiple features so multiple features is fed in so basically here architecture would be smaller okay so uh but again the number of input cells is equivalent to the number of output cells okay so basically how much how many you need uh those will be there so an auto uh encoder network is trained to display output uh similar to Fed input to force Auto encoders to find common patterns and generalized data okay so this is how the training process goes on so the auto encoders are mainly used for smaller representation of the input uh it helps in reconstruction of original data from the compressed data okay so uh it is the reverse process okay so I was telling uh feed formal networks where we will be using for compression right so for uh uncompressing you can actually use this okay so you can reconstruct original data from compressed data okay so uh algorithm comparatively simple uh as it is only uh necessity that the output indicates uh the to the input okay so it is like uh the output is indicating towards the input so here basically encoder means it converts input into lower dimensions and uh the reverse process would be uh decoder okay so that is the reverse process and application wise classification clustering and feature compression okay so mainly for feature compession and definitely the output layers can be in classification type or clustering type of uh again uh unsupervised super learning okay so again classification problems can be dealt with unsupervised learning also okay so we have uh again I think applications we had discussed with these itself and the higher level applications what we can tell is self-driving car voice control assistance automatic image caption generation okay so basically uh automatically uh caption will be generated based on the image and all okay and also automatic machine uh translation okay uh basically I'm telling about language translation of any language from one to another so we have a few of the advantages and disadvantages which we can discuss about the machine learning sorry deep learning so uh and also limitations we'll discuss I think okay so basically advantages it is lessen uh the need of feature engineering so nowhere you will be programming what exactly is happening inside the hidden layers so they are actually hidden so it completely lessens the need of feature uh engineering so you don't need to engineer the features so it eradicates all the all those cost that are needless it is easily identifies the difficult effect it results in best class performance on problems okay those are the advantages disadvantages it is it needs ample of data so ample amount of data is required so the more the data the better so because it needs to create its own kind of algorithms like how exactly it will communicate with different nodes and all right from one layer to another layer like how the data transmission should be done and all okay and next we it is like kind of uh expensive uh to train it is actually like we need quite lot of resources okay so that's why it's little expensive to train time consumption all those data consumption everything okay and uh final disadvantages it is it does not have strong theoretical groundwork so we have not achieved many things when it comes to theoretical practice of this since again uh we are not engineering much of things which is handled by itself but if at all we understand how exactly it is doing or uh getting uh kind of the the groundwork of the theoretical aspects so then we can actually improvise on those with our knowledge okay so maybe representing in mathematical form then again you might have another alternative for those mathematical form okay so uh the kind of in mathematics we have come come across always a better theorem always a better accurate uh formula which can be imple implemented especially for like kind of example like approximation okay so you have whole lot of approximations like you have uh many approximations like one example is like you can do guess and check okay so if your machine takes 30,000 times for guess and check if you use some approximation like Newton rapson okay so it basically takes some 30 steps that's it so 30 iterations instead of 30,000 of guess and check algorithm so similarly if at all we have lot of theoretical groundwork for uh deep learning so then we can actually improvise uh its way also maybe we can force in few of the inputs uh by actually re-engineering but since there is no uh this one engineering like it is doing a good job okay but we can actually boost there is a possibility of boosting it okay so that is the concept uh that is the ideology of not having the theoretical groundw okay so if we had we could have done uh we had a possibility of doing it better okay uh limitation is only it learns only through observations okay so whatever comes through those input layers only it will be learning right so it cannot learn anything other than those okay so we should make sure that uh for example us we have sense organs okay so these sense organs will send inputs to our brain our brain doesn't has its own kind of visualization without the help of our eyes Okay so our eyes sends all those informations to it okay so that's how it works so here also it doesn't learn it through it learns only through observation and it comprises of biases issues okay those are the two limitations so I think that is a uh in detail explanation of uh deep learning so in the next uh video we'll have few deep learning algorithms okay so very quick introduction about them