1. Sparse Coding: A Deep Learning using Unlabeled Data for High - Level Representation Dr.G.M.Nasira R. Vidya R. P. Jaia Priyankka

2. Sparse coding is an algorithm mainly used for unsupervised learning for unlabeled data This paper tries to find High – Level Representation from unlabeled data which clearly gives way to Deep learning Unlabeled data is easy to acquire when compared with labeled data because it does not follow class labels Sparse…

3. The main problem with sparse coding is it uses Quadratic loss function and Gaussian noise mode. So, its performs is very poor when binary or integer value or other Non-Gaussian type data is applied. Thus first we propose an algorithm for solving the L1 - regularized convex optimization algorithm for the problem to allow High - Level Representation of unlabeled data. Through this we derive a optimal solution for describing an approach to Deep learning algorithm by using sparse code.

4. Contact us 1 2 3 4 Ground Rule Neural Network Classifier Sparse Coding L 1 Regularized BLUE PRINT

5. Supervised… Supervised learning can basically be described as learning by example. In this task, a system learns to recognize certain trends by comparing a sample data set against known and labeled examples. In essence, the platform is told to extract features and patterns from the known examples and apply them when examining new, unlabeled data. For instance, let’s say you wanted to teach a system to recognize George Clooney’s face. Using supervised learning, you would provide a data set containing many images. You’d label some of these images as George. From there, the idea is that the system’s supervised learning algorithm would adapt to the data and pick out the right faces.

6. Unsupervised… Here, the goal is to find hidden patterns and structures, which is accomplished by mining and reducing the complexity of data in a multitude of ways. Essentially, unsupervised learning is used to discover trends in data sets when we don’t exactly know what to look for. Sticking with George Clooney, rather than tracking his face let’s say we wanted to track his usage across platforms instead. Multi screen tracking is a problem that the entire data analytics industry is currently struggling to solve. Short of attaching unique IDs to users, it can’t be done with complete accuracy. However, an unsupervised learning system could potentially accomplish this feat with great precision.

8. Contact us

10. Supervised Learning Model

11. Contact us

12. Classification Deep Learning problem arises, when we have limited labeled data for classification task and also large amount of unlabeled data available which is very mildly related to the task.

13. Contact us  The main concept here in Supervised learning is we give our inputs that is to be observed. Only after processing we know whether the input observations are right or wrong. if it is right then it is output. If it is wrong then another set of input under observation will be feed inside. Hence we cannot name it as output. Both the inputs and outputs are only observations.  In unsupervised learning we do not feed input we feed the latent variable (which is half result from the input) and in the result we receive observations, which is again feed as latent variable. This process continuous until we receive output. Process Supervised Learning Unsupervised Learning Observations (inputs) Latent Variable

14. Contact us For more Clarification Latent Variables Observations (output) Observations (inputs) Consider this diagram where is this latent variable comes from.

15. Contact us Starting from Ground Rule: To complete it in simple sentences. we are following non - Gaussian method which is implemented with light modification from Gaussian and Regularized Algorithm. Our rules states that we can use any number of inputs for any number of observations. It means 5 input for first observation,7 inputs for second observation, 3 inputs for 3 observation and so on. There is no limit or minimum number of inputs stated for observations. Next we can even implement our observation in positive form or negative form. Even it can be turned neutral for verification. Rules

16. Contact us Sparse Coding Gaussian:

17. Contact us Sparse Coding Gaussian:

18. L 1 - Regularized Convex Optimization Algorithm

19. Contact us

22. Classifier…  Our goal is to use Sparse coding and to find high - Level representation of unlabeled data in deep learning. The task is to find the best classifier for the implementation. This approach is mainly used for Sparse algorithm which is designed for deep learning.  Neural Network Classifier Consists of Neurons i.e Units. The units are arranged in a layer which converts the input vector into output. Each and every unit first takes an input and applies some functions on it and then it passes the input to next level. Networks are said to be feed - forward, which means it feeds the first input's output to the next level which is input to the next level. It does not gives the feedback to the previous level so it is called feeding forward. weightings is also applied to the each units signals which passes from one level to another. These weighting actually help or tune the particular problem to adapt the neural network.

23. Conclusion & Discussion … In this paper we presented a general method to use Sparse coding for unsupervised learning with the help of unlabeled data mainly for High - Level Representation. Since Sparse coding does not support unlabeled data or non-Gaussian data we derived an algorithm for the optimization usage which came out as L1- regularized convex optimization algorithm. With the help of this algorithm we derived an Deep Learning algorithm for Higher - Level Representation. Overall our result suggest that L1- regularized convex optimization algorithm can help Higher - Level Representation of documents form unlabeled data and this knowledge is useful in classification problem. We believe this model could be applied generally to other problems, where large amount of unlabeled data are available.

24. Future Work… The model represented in the paper consists of two algorithm, which indirectly or directly related the Gaussian Sparse coding with Deep Learning. We have only implemented the model to the images, whereas we like to extend it to Text, Audio and Video related data. It really Seems to be challenging yet this extension when fully related and applied to all the types of multimedia content, It can be applied in Robotics and mainly in their Perception Task.

25. Thank You… Queries…