Scaled Machine Learning: Reflections from the Front Lines

This weekend, Matt Mollison and I were lucky to join the Scaled Machine Learning Conference at Stanford. This post describes some of the AI advances that have excited the best and brightest researchers in the field — and our thoughts on what these advances mean for talent analytics.

Andrej Karpathy

Who he is: the Head of AI at Tesla, or as he describes it, “the largest operator of robots in the world.”

What he’s excited about: new ways to make AI useful for imbalanced class problems” — situations, where an outcome is rare, but important to identify correctly. For example, when it comes to autonomous vehicles, Tesla’s AI knows how to drive well on a straight highway in good weather. But it will struggle with unusual situations like a blue traffic light, a fractional speed limit, or the route sign below. This is because there are relatively few real-world examples, or “training data”, that the AI can use to learn these situations. But, Karpathy argues, “It is precisely the rare examples that are so critical to get right” for humans to trust autonomous vehicles.

A rare but challenging situation for a self-driving car — or a human driver! Photo Credit: JP Ronan

Why it matters for talent analytics: overcoming challenges predicting rare employee events.

The imbalanced class problem is prevalent in talent analytics. For example, how do we identify the “needle in the haystack” that we want to hire when sifting through thousands of job applications? How can we identify the rare, but explosive, combination of factors that can lead to misconduct? Many of the same methods that Tesla is using to make its algorithms attentive to rare — but important — situations can be applied to talent analytics.

Jeff Dean

Who he is: Senior Fellow at Google Brain. (If you haven’t heard of him, the “Jeff Dean Facts” post on Quora is a must-read!)

What he’s excited about: the rise of sequential machine learning. Early machine learning focused on “snapshot” data. For example, given an image, can AI classify its primary subject? Newer machine learning models can work with longitudinal data. One example is machine translation. Instead of just looking at each word, Google now rivals best-in-class human translation by using algorithms that ingest and understand lengthier sequences of text, like sentences and paragraphs. Similarly, in medicine, sequential models are able to ingest and understand a patient’s medical history, instead of just their latest symptoms, to assist with diagnosis.

Comparison of Google’s new sequential machine translation AI (green) vs. older phrase-based model (blue) vs. expert human translation (orange). Photo Credit: Google

Why it matters for talent analytics: better predictions of employee outcomes.

As an industry, HR is in the early stages of predicting employee outcomes. When we try to understand the future, we often use data from today only as the input. For example, when predicting attrition risk for employees, we might look at the employee’s current salary, position, and engagement. With a sequential AI, we can look at that employee’s full experience at the company — how have salary, position, engagement, and many other factors evolved from the employees first day to the present? Given how inherently complex human beings are, and how many of our decisions are driven by experiences we had long before the decision-point, understanding an employee’s history is critical to understanding their future.

Ilya Sutskever

Who he is: Cofounder of OpenAI, a non-profit discovering and enacting the path to safe artificial general intelligence.

What he’s excited about: “self play”, or systems that teach themselves by generating their own data. The most well-known example of such a “bootstrapped” AI is Google DeepMind’s AlphaGo Zero. This system taught itself to play the board game Go, which is more complex than chess, by playing against a version of itself. Programmed only with the rules of the game — but provided no other data — AlphaGo Zero played 4.9 million games against itself over 3 days, developed never-before-seen strategies, and then beat the the best Go player in the world 100–0.

Photo Credit: KDnuggets

Why it matters for talent analytics: workplace simulations for HR decision-making.

Simulations can be especially powerful for HR, because data is especially precious and expensive in human capital. Data on the impact of changes in talent management can come at real cost to employees’ well-being and the bottom line. The ability of AI to generate realistic human capital outcome data without impacting real people is a potential game-changer. Take compensation as an example: a raise is a simple to enact, but the outcomes are complex and hard to predict (Will employees stay longer? Be more engaged? Perform better?) New AI systems can open opportunities to test tricky hypotheses like these “in silico” before making decisions that impact thousands of people — and making it more likely that the real impact is positive.

If you’re interested in learning more about AI and machine learning for talent analytics, please drop us a line at!