Reporting
Please prepare a few slides to present (which will be uploaded on our confluence page)
- Developing a customized optimizer : Sammy
- Developing a customized loss-function : Sammy
- Finding algorithmic error :
Impact from initialization of weights
Limitation in experimental, statistical & systematic error: Aaryan
(How big can be the error of F be to do a decent extraction) - Identifying the optimum # of points of F corresponding to different angles : Pranav
- Identifying kinematic limitations in a decent extraction for a particular formalism and for particular-model (model for CFFs)
- Cross-validation model comparison: Cole
- Comparison between Tensorflow Vs Pytorch : Cole
- Handling different observables such as FUU, FUL, FUT etc.
- Validation approach to identify the error : Surya
Work organization
GitHub page: https://github.com/extraction-tools/ANN
Confluence page: https://confluence.its.virginia.edu/display/twist/ANN+Fitting+Project
Weekly Updates
| Date | Title | Presenter | Slides | Summary / Notes |
|---|---|---|---|---|
| Cole | |||
| Error on F and Consistency Statistics | Aaryan | ||
| limiting the angles of phi | Pranav | ||
| Pranav | |||
| Pranav | |||
| Aaryan | |||
| Pranav | |||
| Aaryan | |||
| Cole | |||
| Sensitivity of (F-BH) (in BKM02) | Aaryan | ||
| Fits using method2 with reduced phi (in BKM10) | Pranav | ||
| Pranav | |||
| Pranav | |||
| Pranav | |||
| Aaryan | |||
| Pranav | |||
| Amanuel | |||
| Cole |
Weekly Updates
Before
[Tuesday][A]
Pranav: Generating pseudo data & optimizing local fits (with Liliet)
Aaryan: Exploring & testing the Method 2 with initialization of weights, and exploring how to bootstrapping effectively
[Tuesday][B]
Surya: Optimizing local fits
Cole: Finding bounds for CFFs using a novel approach
[Friday][A]
Nathan: Finding bounds for CFFs using grid-search
Andrew: Finding bounds for CFFs using Nick's approach & exploring loss-function
[Friday][B]
Sammy: Exploring loss function, implementation of stigmergic loss
Arthur: Fast prototyping with GPUs
| Date | Title | Presenter | Slides | Notes |
|---|---|---|---|---|
| Plan | Ishara | There are 9 working groups reporting updates every other week (please see attached slides) |
Tuesday group
| Date | Title | Presenter | Slides | Summary / Notes |
|---|---|---|---|---|
| Possible approaches: weight averaging | Aaryan & Pranav | ||
| CFF Bounds Local Fit Experimentation | Cole | ||
| Pranav | |||
| Different Loss Function Exploration | Aaryan | Played around with different loss functions including MAE, MSE, and working with the interference values. Also worked with method 2 that does not reinitialize weights after each replica, and saw how predictions for CFFs get better for following replicas. | |
| Optimizing local fits | Aaryan | Exploring methods with seeing when to reset weights and observing differences in averages when weights are reset for each replica and when they are not. | |
| Optimizing local fits | Pranav | ||
| Cole | |||
| Surya | |||
| Different Methods | Aaryan | Tried combining Nick's Method and Pure Bootstrapping Method and saw that averages are similar to correct results while distribution of results are still narrow as pure bootstrapping. | |
| Pranav | Worked on implementing a K-means clustering algorithm in order to narrow down networks at intermediate stages using weight averaging. Two methods: using the networks immediately after weight averaging and training for an additional 100 epochs after weight averaging. Overall, using relatively small numbers of replicas and training epochs, the results were extremely poor for both methods tried. The next step is to experiment with different numbers of replicas and cluster centers to see if better results can be obtained. | ||
| TensorFlow Graph Execution and tf.function | Cole | ||
| Exploring combined-type Method | Aaryan | with/without weight reset and random selection of weights for initial training | |
| PyTorch Vs Tensorflow for BKM02 | Cole | ||
| Consistency/Reproducibility of Experimental Method | Aaryan | ||
| K-means algorithm for intermediate weight averaging | Pranav |
Friday group
| Date | Title | Presenter | Slides | Summary / Notes |
|---|---|---|---|---|
| A model to narrow bounds for ReE | Andrew | Information on tensorflow update is included | |
| Andrew | |||
| Nathan | will be uploaded soon | ||
| Mathematica Implementation | Annabel | ||
| Custom loss function update | Sammy |
Prospective projects
- Developing a customized optimizer : Sammy
- Developing a customized loss-function : Sammy
- Finding algorithmic error :
Impact from initialization of weights - Identifying specific angles (low/high angles critical for good fits)
- Limitation in experimental, statistical & systematic error: Aaryan
(How big can be the error of F be to do a decent extraction) - Identifying the optimum # of points of F corresponding to different angles : Pranav
- Identifying kinematic limitations in a decent extraction for a particular formalism and for particular-model (model for CFFs)
- Cross-validation model comparison: Cole
- Comparison between Tensorflow Vs Pytorch : Cole
- Handling different observables such as FUU, FUL, FUT etc.
- Validation approach to identify the error : Surya
| Chapter# | Topic | Assignment |
|---|---|---|
| 2 | Hyper-parameter optimization for Local Fits | |
| 3 | Hyper-parameter optimization for Global Fits | |
| 4 | Limitation in experimental, statistical systematic error | Aaryan |
| 5 | Identifying the optimum number of points of F corresponding to different angles | Pranav |
| 6 | Validation approach to identify the error | Surya |
| 7 | Boot-Strapping Value | |
| 8 | Cross-validation model comparison | Cole |
| 9 | Developing a customized optimizer and loss-function | Sammy |
| 10 | Comparison between Tensorflow Vs Pytorch | Summy |
| 11 | Finding algorithmic error | |
| 12 | Model Ensemble Study | |
| 13 | Identifying kinematic limitations in a decent extraction for a particular formalism and for particular-model (model for CFFs) | |
| 14 | CFFs extraction through Grid-Search algorithm | Nathan |
| 15 | Handling different observables such as FUU, FUL, FUT etc | |
| 16 | NN Fit only to interference term |