Machine Learning Under the Hood: The Technical Tips, Tricks, and Pitfalls

Por: Coursera . en: , ,

  • MODULE 1 - The Foundational Underpinnings of Machine Learning
    • In what way is bigger data more dangerous? How do we avoid being fooled by random noise and ensure scientific discoveries are trustworthy? This module covers the fundamental ways in which machine learning works – and doesn't work. First, we'll cover three prevalent, heartbreaking pitfalls: overfitting, p-hacking, and presuming causation when we have only ascertained correlation. Then we'll establish the foundational principles behind the design of machine learning methods.
  • MODULE 2 - Standard, Go-To Machine Learning Methods
    • This module covers four standard machine learning methods: decision trees, Naive Bayes, linear regression, and logistic regression. We'll show you how they work, checking their predictive performance over example datasets and visualizing their decision boundaries as a way to compare and contrast their capabilities. You'll also see how to evaluate these models in terms of lift and profit, and why improving model probability estimates is so important.
  • MODULE 3 - Advanced Methods, Comparing Methods, & Modeling Software
    • When should you turn to deep learning, the leading advanced machine learning method, and when is its complexity overkill? And is there a way to advance model capability and performance that's elegant and simple, without involving the complexity of neural networks? In this module, we'll cover more advanced modeling methods, including neural networks, deep learning, and ensemble models. Then we'll compare and contrast the full range of modeling methods, and we'll overview the many machine learning software tool options you have at your disposal. We'll then turn to a special, advanced method called uplift modeling (aka persuasion modeling), which goes beyond predicting an outcome to actually predicting the influence that a decision would have on that outcome. We'll explore the marketing applications of uplift modeling and see success stories from the likes of US Bank and President Obama's 2012 reelection campaign.
  • MODULE 4 – Pitfalls, Bias, and Conclusions
    • Crime-predicting models cannot on their own realize racial equity. It turns out that models that are racially equitable in one sense are not in another. This is often referred to as machine bias. This quandary also applies for other kinds of consequential decisions driven by predictive models, including loan approvals, insurance pricing, HR decisions, and medical triage. This module dives deep into understanding the machine bias conundrum and what recourses could be considered in response to it. We'll also ramp up on a related, emerging movement in support of model transparency, explainable machine learning, and the right to explanation. We'll then wrap up the overall three-course specialization with a summary of the ethical issues, the technical pitfalls, and your options for continuing your learning and career path in machine learning.