davidvenuto
commented
1 year ago Goals for 6/18-6/25
-
[x] Sift through Bernard's code
-
Particularly look at his code on the MEOFs, and identify the dominant modes. These will give context about the predictor variables in the ML model
-
Find the daily version of Bernard's data, and begin figuring out how to draw these important variables from this data.
-
Look through Bernard's remaining code of relevance
-
[x] Start semi-development of time series model ( Too early not enough understanding)
-
[x] Possibly perform independent MEOF analysis on new data
-
[x] Decide on the proper way to implement model (RNN, CNN, LSTM (most likely LSTM according to Dhruv))
- LSTM appears to be the way to go. Its ability to capture long term patterns and also deal with non-linearity appeals to the goals of our project. May be a little computationally difficult given the size of the data
- Other options that were considered: Vector Autoregression, Gated Recurrent Unit, Transformer Model, Gaussian Processes
- Vector Autoregression: Linear model, so obviously not an ideal choice
- Gated Recurrent Unit: Struggles with long term dependencies
- Transformer Model: Computationally expensive and less interpretable than other models
- Gaussian Processes: Limited scalability to high-dimensional data and requires assumptions about the underlying distribution and kernel functions.
-
[x] Strengthen understanding of LSTM model
- LSTM is a type of recurrent neural network that uses previous inputs to predict the next output (hence recurrent)
- LSTM requires input-output pairs. This is commonly done through the sliding window technique.
- Example: Lets say we have number 1-10, and the window value is 2. The first input would be [1,2], and the corresponding output would be 3, making the input-output pair ([1,2],3). The next input would be [2,3] and the corresponding output would be 4, making the second input-output pair ([2,3],4). This pattern would continue until 10.
- After creating our input-output pairs, we need to convert our training and testing data into numpy arrays and then reshape them to actually use in the LSTM.
- Then we create our model using Keras, and train this model on the training data.
- We can validate our predictions on the test data
- Finally, run model on independent data to determine its precision.
-
[ ] Lay out framework of what is necessary to implement this specific model in a separate git-issue
-
-
[x] Go back to Chen-Yuan (2004) and summarize understandings, and how these ideas will transfer over to your project
-
Essentially, Chen-Yuan describes the identification of the dominant modes (spatial distribution between the variables) between a number of variables. From the leading modes, we can identify what variables are the most "important" from which we can begin creating our own model using ML.
-
Jianna's job will be to come up with a better way of utilizing dimensionality reduction, which will possibly change the modes and thus the variables. The ideal situation is the one in which me and Jianna are able to combine our projects into one.
-
[ ] Potentially integrate over to orca computer, maybe not if LEAP continues serving purposes for now (Not necessary yet)
-
[x] Create Powerpoint for LDEO Research Focus Session
- [x] Edit and provide visuals
- [x] Update throughout the week if necessarry
-
[x] Come back at end of week and review how each of the goals were met, and any complications
-
[x] Write the goals for next week
Update for end of week
-
This week went relatively well. I think the main gap that may come back to bite me if I don't fortify it is my understanding of the previous work. I have a basic understanding of MEOFs, the markov model, etc, but not enough to explain it, as was highlighted in my research focus session. I think the next research focus session and our meeting on Monday may be a bit of a wake up call, as it will force me to 1. Really start knowing what I need to be working on, and to get working on it, 2. To actually understand the basis for my project and 3. To start the implementation of my beginner model.
-
The fact that we are almost halfway through the summer is a bit mind boggling to me, as I feel like I haven't made much progress. I think I need to start being more direct in going to Dhruv and Xiaojun with questions and to really make what I want to happen, actually happen.
-
A lot of the soft goals for this week were met, but development has yet to start. Once again I think is due to a relatively short understanding of the previous material, but I know I can turn that around next week.
This issue will serve as an overview of my goals for this summer, broken down by each week