Difference between revisions of "User:Cnapun"

Latest revision as of 03:27, 11 August 2017

Personal Information

• Case Western Reserve University Class of 2019
• Applied Mathematics and Computer Science Major

Weekly Log

Week 0 (30 May - 2 Jun)

• Met Dr. Povinelli 4 times: to discuss possible project topics, to get oriented with the lab, to decide on a project topic, and to establish weekly milestones
• Obtained MU and MSCS account logins and ID card
• Read most of "Learning Deep Architectures for AI"
• Read section on Sequence Modeling (Ch 10) of The Deep Learning book
• Read various other papers

Week 1 (5 Jun - 9 Jun)

• Attended GasDay camp and learned about what GasDay does
• Attended responsible conduct of research training
• Continued to read papers
• Decided to use TensorFlow and Keras for now
• Got Anaconda, TensorFlow, and Keras installed on a couple lab computers
• Learned how to access customer data

Week 2 (12 Jun - 16 Jun)

• Started playing with TensorFlow, mostly using the GEFCom2014-E dataset to allow me to continue working when not at the lab
• Things I got working in Tensorflow:
  • n-layer Sequence to sequence (seq2seq) model (encoder-deocder architecture)
  • Autoregressive seq2seq model (using slower ops)
  • Multilayer Perceptron
  • 1D ConvNet on weather inputs, in parallel with MLP to process load and time inputs
• Will try ConvNet on all 3 inputs, feeding into RNN, with the seed state being the month, day, and day of week at the initial timestamp
• Will try Clockwork RNN

Week 3 (19 Jun - 23 Jun)

• Reinstalled CUDA
• Started training seq2seq models on GasDay (electric) data
• Got TensorBoard to work
• Trained AR LSTM on GasDay gas data

Week 4 (26 Jun - 30 Jun)

• Continued to experiment with hyperparameters
• Achieved good performance on both gas and electricity datasets
• Prepared a presentation of my research for other REU students (slides)
• Attempted to train a model with objective that was not at all convex; it failed miserably
• Still can't figure out why Dropout makes everything worse

Week 5 (3 Jul - 7 Jul)

• Continued to experiment with different models
• Incorporated both long- and short-term dependencies into an encoder-decoder architecture
• Presented to GasDay information about what I have been doing and neural networks in general (slides).

Week 6 (10 Jul - 14 Jul)

• Did not finish draft of paper by deadline
• Worked on determining the best model
• Reincorporated BatchNorm and DropOut with good results
• Experimented with convolution layers, but they rapidly overfit data

Week 7 (17 Jul - 21 Jul)

• Started creating poster
• Tried to evaluate models on test data, but results were highly subpar
• Spent the rest of the week trying to figure out why errors on test data were inconsistent with val errors
  • It is probably because the model is trained on years 0-14, validated on 14-18, and tested on 18-20
  • 4 years skip between training data and test might cause issues
  • Tried using detrended data, but results did not improve significantly

Week 8 (24 Jul - 28 Jul)

• Finished poster (pdf here)
• Finalized models for gas forecasting. I decided to not use GasDay's electricity data at all.
• Started running models on electric data from GefCom 2014 so that results may be reproduced.
• Finished first draft of paper (which I may try to submit to a conference)
• Presented my work to GasDay

Week 9 (31 Jul - 4 Aug)

• Gave draft to Dr. Povinelli for feedback
• Presented to REU students
• Poster session
• Continued to tune hyperparameters on GefCom data
• Finished Responsible Conduct of Research training

Reading

Here are some of the papers I have read, skimmed, and partially read:

Forecasting with Deep Learning

• "Building Energy Load Forecasting using Deep Neural Networks"
  • Discusses the direct application of seq2seq LSTMs to load forecasting
• "Training Recurrent Networks by Evolino"
  • Describes the use of genetic algorithms to train RNNs
  • Should offer better performance than Echo State Networks (ESNs)
  • No description of how crossover and mutate operations work; I referred to "Neural Network Weight Selection Using Genetic Algorithms" for a more complete explanation

Hybrid Methods

• "Hybrid Neural Networks Over Time Series For Trend Forecasting"
• "Time series forecasting using a hybrid ARIMA and neural network model"
• "Intelligent Hybrid Wavelet Models for Short-Term Load Forecasting"
• "Short-Term Load Forecasting: Similar Day-Based Wavelet Neural Networks"

Review Papers

• "Short-Term Load Forecasting Methods: A Review"
• "An overview and comparative analysis of Recurrent Neural Networks for Short Term Load Forecasting", which gives a good, current overview of methods

LSTMs, Training, and Possible Improvements

• "Recurrent Neural Network Regularization"
• "A Theoretically Grounded Application of Dropout in Recurrent Neural Networks"
• "Multiplicative LSTM for sequence modelling"
• "Phased LSTM: Accelerating Recurrent Network Training for Long or Event-based Sequences"
• "A Clockwork RNN"
• "Semi-supervised Sequence Learning"