"No one is harder on a talented person than the person themselves" - Linda Wilkinson ; "Trust your guts and don't follow the herd" ; "Validate direction not destination" ;

April 25, 2021

AI / ML Refresher Notes - Forecasting

Timeseries forecasting notes

Sessions - Timeseries forecasting notes1

Key Notes

  • Things evolve based on a function of time
  • Properties as a function of time
  • Series of data indexed in time order (Year / Month / Seconds)
  • Stationary - Statistical properties like mean, variance are constant over time
  • Non-stationary - things change
  • libraries - statsmodels, pmdarima
  • ARIMA (For Non-stationary it works)
  • Fully Connected Neural Network, RNN

ARIMA - p,d,q

  • AR(p) - Autoregression
  • i(d) - integration
  • MA(q) - Moving Average

SARIMA - seasonal ARIMA

Sessions - Timeseries forecasting notes2

Key Notes

  • Check data stationery or not
  • Split data into months 
  • Additive time series = Base level + Trend + Seasonality + Error
  • Multiplicative time series = Base level X Trend X Seasonality X Error
  • Positive / Negative correlation (Scatterplot also provides similar feature correlation)

Sessions - Timeseries forecasting notes3

Key Notes

  • SARIMA - seasonal dataset
  • Pyramid Auto Arima - pmdarima - pick best models to choose
  • pip install pmdarima
  • pip install statsmodels --upgrade
  • Aggregate data by week / month / dayofweek for insights
  • Up & Down indicates - Trends / Seasonality
  • It is changing not stationary - Upward trend yearly
  • Seasonality - Repetitive every year ?
  • auto_arima model

Sessions - Timeseries forecasting notes4

Key Notes

  • Feed Forward Networks
  • Sequential Layers 
  • Dense - Fully Connected Layer
  • Preprocessing, minmax scaler transformation
  • With every fifth point predict sixth point
  • Timeseries generator in keras can be used

Session - BERT Neural Network - EXPLAINED! 

Key Notes

  • Transformers for Neural Machine Translation
  • LSTM Challenges - sequential words, slow learning, Left to right, right to left context learning
  • Transformer - Processed simultaneously
  • Transformer has encoder / decoder
  • Generate embedding for each word
  • Decoder takes words of translated words and generate translations next word
  • Encoder learns english + context
  • Decoder how english maps to french words
  • Stack just encoders - Bert
  • Stack decoders - GPT architecture
  • BERT for NMT, Question Answering, Sentiment Analysis, Text Summarizing
  • Pretraining - BERT learns language, context. Masked language model, next sentence prediction
  • Finetuning - Q&A / Answers. Replace last layer with output parameters
  • Phase I output is Token embedding, segment (sentence) embedding, position embedding
  • All word vectors have same size, generated simultaneously
  • Training using cross entropy loss
  • BERT base / BERT large models 

Code Exercise - Link

To-do list

Transformer Neural Networks - EXPLAINED! (Attention is all you need)

More Revisions, More Revisions!!!

No comments: