Prateek Nayak
Email: [email protected]
Phone: +91 897 100 5339
| Full Name | Prateek Nayak |
| Institute | 1st Year B. Tech Student Computer Science and Engineering PES University Bengaluru, India |
| [email protected] | |
| Phone | +91 897 100 5339 |
| Github | https://github.com/Kriyszig |
| Timezone | Indian Standard Time (GMT + 530) |
I am Prateek Nayak, a first year undergraduate Computer Science and Engineering student studying in PES University. I have experience in programming with multiple languages such as Python, JavaScript, TypeScript, C, etc. I am passionate about Web Development, Algorithms, Machine Learning and Distributed Computing. I am especially interested in Open Source Software and I actively try to contribute to OSS - anything from fixing simple documentation issue to submitting simple bug fixes and writing unit tests.
I have used TensorFlow in past (mostly in Python) for university projects and was surprised how robust the whole framework is. Most of the cutting-edge components one would need while building a model is already implemented in TensorFlow which allows us to shift our focus on more important optimization tasks rather than reinventing the wheel. TensorFlow.js has especially piqued my interests among all the TensorFlow products as it brings Machine Learning to browser making it more accessible than ever.
When I'm not on my laptop, you can find me reading fantasy novels, listening to podcasts, solving Physics problems, playing badminton or cycling through nature.
Machine Learning has fascinated me, right from the moment I discovered the topic. Every time I wanted to try out something new in the field, TensorFlow has always come to the rescue. TensorFlow has helped not just me, but millions of people - people who are trying something new, people who are carrying out ground-breaking research, people who are trying to make this world a better place. I would like to give back to the community and help people transcend their thoughts to code faster and easier than ever.
TensorFlow.js lags behind its Python counterpart in the sheer number of functionalities it has to offer. By adding more Linear Algebra operations, certain data pre-processing operations will become easier than ever making Machine Learning on browser more lucrative than ever.
TensorFlow.js [tfjs-core]
List of the contributions can be found here
I will be working 6 - 8 hrs per day throughout the duration of the project.
I will have my end semester examination in second and third week of May during which I will be little busy but other than that, I'll be free throughout the summer.
TensorFlow.js is an open source WebGL accelerated Machine Learning library to train and deploy ML models. Compared to the Python counterpart, Tensorflow.js had been around for very little time. Even with less time to work with, Tensorflow.js has matured quite a lot. Given the obvious restrictions of a browser, Tensorflow.js pushes itself to the boundary.
Machine Learning is a field moving at speed of light. While catering to recent demands of features to keep up with the cutting-edge technology, development of less exciting features like linear algebra operations fell behind. In order to be compatible with browser and maintain a small footprint, implementation of features natively is appealing compared to adding a new dependency that will either break the compatibility or increase the footprint of library more than desired.
Linear Algebra operations can be of great help in data pre-processing stage. Implementation of these linear algebra operations in the TensirFlow.js core API can help many developers. It will help bridge the gap between Tensorflow.js and its Python counterpart and will greatly improve tasks like image processing in the browser.
- Implement significant linear algebra operations namely Singular Value Decomposition, Principal Component Analysis and Matrix Inversion in TensorFlow.js core API.
- Document the additions in the official documentation of Tensorflow.js
- Optimize the solution
Implement Matrix Inversion
Gauss-Jordan Elimination would be an ideal way to compute the inverse of matrix with O(n^3) complexity. However, the variation of Method of Four Russians might be a better option because of it's lucrative tight bound of Θ(n^3 / log n). The algorithm is described in this paper by Gregory V. Brad.
Implement Singular Value Decomposition
The PDF from University of Texas has covered most of the algorithms that can be used to find SVD of a matrix. I have looked into some of the implementation of these algorithms and the ones I found promising were:
SVD by Jacobi Rotation(Implementation: Planeshifter/SVD)
The implementation uses Householder Reduction and QR factorization to compute the SVD.SVD by Least Square Method(Implementation: danilosalvati/svd-js)
Implements solution described in this paper by G. H. Golub and C. Reinsch
implementation of Principal Component Analysis I propose an implementation similar to one described by Tomáš Bouda in this blog post. This implementation requires the need of Eigenvectors and Eigenvalues which will give us an additional opportunity to implement the same and expose it in the core API of TensorFlow.js
The algorithms discussed above were a result of limited research. Further discussions with the mentors will help us select the most efficient and optimized algorithm for the tasks described above.
| Duration | Task |
|---|---|
| April 9 | Deadline for submitting Project Proposal |
| April 9 - May 6 |
|
| May 6 - May 27 | Commencement of Community Bonding Period
|
| May 27 - June 19 | Official Coding Period Commences
|
| June 19 - June 24 |
|
| June 24 - June 28 | Phase I Evaluation
|
| June 28 - July 18 | Phase 2
|
| July 18 - July 22 |
|
| July 22 - July 26 | Phase II Evaluation
|
| July 26 - August 16 | Phase III - Final Phase
|
| August 16 - August 19 |
|
| August 19 | Final Evaluation
|
Note
- The timeline above assumes the implementation of the algorithms discussed in Synopsis. The timeline may change if the discussion with mentors yields a better approach to solving the problem however the goals mentioned for each phase will remain mostly unchanged.
- The implementation discussed aren't consistent with the implementation of these operations in TensorFlow Python API. If found necessary, algorithms involved can be changed to make function prototype in TensorFlow.js consistent with TensorFlow Python API.
- Implement popular linear algebra operations - Matrix Inversion, Singular Value Decomposition, and Principle Component Analysis - in the core API of TensorFlow.js
- Implement and expose operations such as Householders Transform which act as helper function to the main operations discussed above.
- Implement more linear algebra operations in the core API of TensorFlow.js
- Get involved with the TF.js community and help in developing features outside Linear Algebra operations.
Thank you for going through the proposal