Derin Öğrenme.pdf

INTRODUCTION to DEEP LEARNING kgDeep inspired learning is a AI subset of machine learning by the neural networks in the human brain mania Learning 2How the GPUs work why do they chosenfor these type of Deep Learning application has already been used in scientific field such as by It computer vision...

INTRODUCTION to DEEP LEARNING kgDeep inspired learning is a AI subset of machine learning by the neural networks in the human brain mania Learning 2How the GPUs work why do they chosenfor these type of Deep Learning application has already been used in scientific field such as by It computer vision Natural Language Processing speech Recognition and many others proved suitable to solve well defined logical problems such as playing by AI cheese LEARNING MACHINE G The to make by A objective of machine learning is to capture regularity in data general predictions machine learning Rules system is trained rather classicalprogramming Data byTypes of Supervised seni machine learning Answer Data Answers than explicitly Machinelearning programmed rules algorithms Learning supervised learning Unsupervised Learning Reinforcement learning Supervised ByThe Learning machine is thought by example as The operator provides the ML algorithm with a known dataset that includes desired inputs and outputs and the algorithm must find a method to determin how to arrive at those inputs and outputs bg The algorithm identifies patterns indata learns from observations andmakes predictions kg The algorithm this process performance makes predictions and is corrected by the operator and continues until the algorithm achieves a high level of accuracy Unsupervised Kg The learning ML algorithm studies data to identify ay There is answer key or human operator to provide instruction no be Instead the machine determines available data by In an unsupervised kgThe algorithm tries to learning process the the and address that data accordingly structure byThis might patterns mÂĞĞʰĞ 7Ü organise that data in some interpret by analysing large data sets way to describe its grouping the data into clusters or arranging it in mean that looks more correlations and relationships organised a way Semi supervised Learning kg Semi supervised learning labelled and unlabeled bg Labelled algorithm Ey By data is supervised learning data but instead uses both that has meaningful tags so that the the data whilst unlabeled data lacks that information essentially information can understand using this is similar to combination MLA can learn to label unlabeled data Reinforcement Learning by RL focuses with a on set of regimented learning process where actions parameters and end values a MLA is provided Bs By defining the rules the MLA then tries to explore different options and possibilities monitoring and evaluating each result to determine which one is optimal as Reinforcement learning teaches the machine trial and error learns from post experiences and begins to adopt its approach in response to the situation to achieve the best possible result Sylt Fromdata is an algorithm MACHINE Classification is LEARNING k 1 Object Recognition cancer detection speech processing experienceEE Transcription Find structure inText Machine Translation Parsing analyze a sentence into its parts and describe their syntaticroles Experience IEE.IE imE of clusters unknown patterns Supervised known pattern want to classify missingvalues tion data performance iYn Input Chadcrafted AdaBoost ftp.FEI.FI Feature Extraction and incorrect Tyra 0 1 loss However all this needs to beper formed on data that the algorithm was not trained on to give a good estimate the right Output classification FE.gg Input image Examples Effff outputs A label or identifies that that choosing performance index cases 0cT aEE an I.EE EE TrodionetML output error rate as can be quite difficult N of of correct output us Can also Output O EE.E.E examples Learning EE L easy Measure accuracy proportion Machine learning Vs Deep Learning Input E Unsupervised Want to find Anomaly detection identifying data points in data that don't fit the normal patterns A computer program is said to learn from experience E with respect to some class of tasks Tana performance measure P.lt _ISMEEEI 5f.tt feIIizn Regression customer satisfaction stock prediction epidemiology Generators Prediction Horning in som Traditional Feature Extraction Ss Features are unique properties in the image that are used to classify its objects bg Extracting to be fed and consolidating features to the classifier from thousands of images in one feature vector By characteristic of a good featurefollow Identifiable Easily tracked and compared Consistent across different scales lighting conditions and viewing angles Still visible in noisy images or when only part of an object is visible AUTOMATIC FEATURE asInput images features layer EXTRACTION BY DEEP LEARNING pass through by layer the layers of a neural networks so it can learn Classification Example CLASSIFICATION APPLICATION WORK FLOWN DEEP LEARNINGN Deep learning lets us express difficult representations as is approximate a function t which maps input parameters m theta by Goal ByNetworks kgHidden given that often we have simpler representations to a fig next The chain is category y given denoted layers since can not directly see output due to layer depth what do the representations Learned by a Deep Learning Algorithm Look Liken by Five A decades taxonomy of Research in of ML ML provided concepts classification generative models clustering kernels linear embeddings etc A sound statistical formalization Baysein estimation PAC A clear picture of fundamental issues bias variance dilemmaVCdimension generalization bounds A good understanding of optimizationissues An efficient large scale algorithms kgFrom a practical perspective deep learning the need for a deep mathematical graps makes the design of largelearning architectures a systemsoftware developmenttask allows to leverage modern hardware clusters of GPUs does not plateau when usingmore data makes large trained networks a commodity lessens ML PROBLEM TARGET be Classification LL EE anEEE ii KsRegression 4 random variables on we wantto estimate DYJYradfm.si 2 P XR ELY Effpeonpz ML and DL X r STEPS Preparing dataset Test set Building Model Model Architecture Design Selectingand configuring the training Algorithm Performing the training Normally takestime EY.fi jttothmeeaEET MEe hm isdoing a goodJob Sükür DEEP HUMAN LEARNING LEARNİNG Buildingthe trainingenvironment pythonMATLAB Preparing dataset En Performing the training Normally takestime of the trained model is a way to measurethatthe good job APPLICATIONS Deepfakes NDL EEEE.ec on Text Books Lecturenotes Final tests Selectingand configuring the training Algorithm algorithm isdoing a Preparing curriculumandboo Periodic tests Building Model Model Architecture Design Evaluation Selecting school ChatGPT Lip reading Image to Image Translation Selecting students Selecting the teacher Going into the educational year Evaluation of students Awayto measurewhether thestudent is doinggoods TOP OPEN SOURCE DL TOOLS DL with Python as there areseveral Tses Lies hiddenlayers depth in deep learning in comparison g Tensorflow to traditional MI networks _Nueval networks nodes represent operations while edges stand for multidimensional data arrays tensors flowing between them Lisused for NLP Most effective for image classification summarization speech recognition translation and more b 2 Microsoft interfaces with C C python MATLABand is Psmachines speed modularity and expressions through massive volumes of data in very little time summed industry focused commercially supported distributed 5 Deeplearning deep learning framework finemost Speed off IYI 4 Keras ET neural networks F CNN and RAN to track especially relevantfor convolution can skim cognitive Toolkit MCTK sufis both for scale 3 Caffe MI speech and text basedata advantages CNN and Python EEEEs EE.EE DATA by A REPRESENTATION tensor is a TENSOR generalized matrix along Teveral discrete dimensions A OD tensor is EGA AD tensor is leg a scalar a vector Sound sample Be A 2D tensor is a matrix leg table of numerical values indexed Compounded data can be seen as a vector of identically sized matrix e g multi channel image o video Be A GD tensor can be seen matrix of identically sized matrices or a sequence of 3D tensors e g sequence as a of video or so on multichannelvideos structures can represent more diverse βdata types data Manipulating this constrained through structure allows us to use CPUs and GPUs at near peak perfermance a greyscale image AGA 3D tensor And a finite Tensors are used to encode The signal process The internal state The parameters of models ARTIFICIAL NEURAL NETWORK assynapse is accountable for receiving the external environment sensory organs for sending motor instructions to our muscles and for performing other activities Tuttum A neuron also receive input from can the other neurons through a branchlike structure called den drite bg These inputs are strengthened or weakened that is they are weighted according to their importance and then they are summed together in the cell body called the sofa kafrom the cell body these summed inputs are processed and move through the axons and are sent to the other neurons How the artificial neuronswork byInputs 12 3 to predict output y These inputs are multiplied W If1 I add a XzWz off strengthening the inputs by their weights wi.ws wzand are 3W I Sum value called bias b 2 W summed together xzowztxz.ws b bias w I I Sum b