Advisor: Prof. C Prakash Vyasarayani.
- Developed time series forecast model in Python, using ARIMA and LSTM models, studying sales data for furntiture dataset and risk and stock dividends of Apple and Tata(NSE India) datasets.
- Developed a network model with the LMS algorithm and multi-layers, studying network response for parameter changes. Used softmax and ReLU architectures for my final submission.
- Performed comparative study of models designed with backpropagation, adaline and multi-layer models, BRNN etc. and tabulated the results.
See the final commits in time-series & func. approx folder, for further details and code.
-
Developed models for pattern recognition of digits and extended this to extract from noisy images.
-
Final model submission uses self-normalizing networks (seLU) for occluded, noisy images of aerodynamic component scans.
-
Concept behind the model:
-
In an autoassociative memory the desired output vector is equal to the input vector (i.e., tq = pq )
This model is used to store a set of patterns and then to recall them, even when corrupted patterns are provided as inputs.Sample image scanned above, will have a prototype pattern : p1 = [–1 1 1 1 1 –1 1 –1 –1 –1 –1 1 1 –1 1 –1 ... 1] T
The vector p1 corresponds to digit 0 (6 x 5 grid scan as illustrated below), p2 to digit 1, ....
Using Hebb rule, weight matrix is calculated. Based on Supervised Learning rule, the code was developed.
- The perfomance graph and validation function can be referred further, along with the codes and mnist.mat and other files.
- The code was extended to extract patterns from noisy and occluded image scans. (illustration below)
See the final commits in Supervised Hebbian learning for further details and code.
-
Constructed models for lecture exercisess on perceptrons, supervised learning & backpropagation in Neural Networks coursework.
-
For instance, the introductory session used perceptron models in Matlab. After studying rosenblatt model and previous network models, I designed a code to generate weights and bias for each iteration until convergence is achieved by cycling through the input vectors. ( convergence_perceptron.m )
-
Method:
p1 p2 p3 p4 2 1 -2 -1 2 -2 2 1 Target values 0 1 0 1 Taking initial value as null for weight and bias, perceptron learning rule was applied and weights and bias were generated. W(0) = [0 0]T , b(0)=0.
Weight values: w = [a b]T
-
Results were generated by checking for convergence of the bias and weights, by checking for each vector to verify the learning rule.
- Received a returning internship offer as Data Science Intern for the next summer with the team.
- Primary reference : Neural Network Design by Demuth, H.B., Beale, M.H., De Jess, O. and Hagan, M.T., 2014.
- All sample codes and implementations, other models submitted for project are in the perceptroncodes repo.
- Refer this post to understand backpropagation applied across the LSTM and then the BRNN apporach.
- Blog to refer for Stocks & Time Series: link 1, link 2