William Liu

Colorado Python Data Science Meetup

Meetup

Reddit Extremist Finder

By: ?

Using reddit comments, determine if hate speech.

Selecting Comments and Labeling Nuts or Not Nuts (240 people each)
Clean Comments
Words vs Comments using Term-Frequency Idf/Bag of Words
NMF/LDA for Topic Modeling
Created a ‘Comment Count per Topic Vector’
Went with Gradient Boost, Random Forest, Adaptive Boost Classifiers
Results can be four categories: ‘True Positive’, ‘False Positive’, ‘True Negative’, ‘False Negative’
Plotted against ROC Curve

By: Jennifer Waller, Ph.D.

Use camera to determine what flower.

Pl@ntNet app looks at photo, tells you what plant
Pl@ntNet uses data from Western Europe, not very good for Colorado
First Approach was to scrape Colorado Front Range wildflower images (n=651), create custom convolution neural network (.88 accuracy)
Second Approach to use a fine-tuned pre-trained ResNet50 model; deeper network = more features (~50 layers), avoids underfitting that happens in traditional deep networks. E.g. one layer might be texture, another color. Added additional custom layers
- Model - fully connected layers (flatten, dense w/ 512)
- Image Preprocessing helps, shift/tilt, change saturation
- Results: Accuracy with random guessing (.09), Top 1 model accuracy on validation data (.97), misclassified 4/306
- www.wildflowerfinder.com - built web app with Flask with predicted species and predicted accuracy

By: Dale Taylor

Investigating the Anna Karenina Effect in Yelp Denver reviews.

Are all happy reviews alike? Are all unhappy reviews different?
Took 200,000 reviews
Can we predict a restaurant’s reviews based on number of topics?
Looked at closed restaurants too so there’s no ‘survival bias’
Looked at number of topics
The Gini coefficient is a way to measure income inequality
Results
- Skip waiters (topics with waiters scored lower scores)
- Skip delivery (topics with delivery scored lower scores)
- Be fast casual

By: Tabor C Henderson

Intro to Deep Quality Networks (e.g. DeepMind plays Atari)

Reinforcement Learning - it sucks; an agent does an action in an environment, interprets the state and reward
DeepMind plays Atari, only knows that high score is better
DQN Algorithm is pretty small
Utilized ‘experience replay’ to avoid overfitting and local minima
Neural Nets are a stochastic method of solving problems, meaning they’re approximate
A deterministic algorithm follows specific rules

Can we solve large text based questions with machine learning?

A team of Stanford data scientists solved it
Hardest part is Parsing Rules
NLP Landscape:
- Mostly Solved
  - Spam Detection
- Somewhat Solved
  - Text Parsing
Uses Stanford/Google’s Parsey McParseface (2016)
Translator using Google’s Seq2Seq (what Google Translate is using)
- Source Sequence (e.g Das ist mein Haus.) to Target Sequence (That is my House)
- Models include Neural Net, then a Recurrent Neural Net, then an Encoder/Decoder, then LSTM/GRU, then Bidirectional encoding, and currently Attention (released around 2015)
  - Issue with NN were they can’t do sequences well (e.g. many words back to back)
  - Issue with recurrent NN was reading v writing
  - Issue with Encoder/Decoder is forgets meaning context after about 7 words
  - Issue with LSTM/GRU is can’t look ahead
  - Issue with Bidirectional encoding is single, fixed-length vector (heavy lift for a single vector)
  - Attention works pretty well because it allows putting emphasis on different words
- You can use Seq2Seq even with Source Language as English to Target Language as Python
Some say that NLP future isn’t bright. Neural Nets are powerful, will probably replace
Google uses SyntaxNet, part of a larger suite that includes Parsey McParseface