365: Deep Learning Models For Recruitment

Episode Summary

Executive Summary: John Crohn, Chief Data Scientist at Untapped, discusses his path from neuroscience and quantitative trading into data science, and how he applies machine learning to HR candidate matching. He explains the trade-offs between transformer-based NLP accuracy and production latency, emphasizes explainability and bias checks, and shares his approach to teaching deep learning through visuals, books, and free educational content.

Main Topics: Career path into data science: Crohn describes moving from neuroscience research at Oxford to quantitative trading, then discovering data science through a ZocDoc job posting that matched his background. Untapped’s HR machine learning products: He explains how Untapped helps recruiters surface better candidates from millions of applications using deep learning and ranking models rather than keyword search. NLP, transformers, BERT, and GPT-2: The discussion covers transformer architectures, GPT-2 demos, and why Crohn believes these models are powerful but often too computationally expensive for production use. Accuracy vs. compute in production systems: Crohn contrasts academic model optimization with real-world constraints, arguing that sub-second user experience often matters more than marginal accuracy gains. Explainability and bias mitigation: He stresses the need for fairness testing and describes Untapped’s approach to ensuring equal scores across matched male/female candidate test cases. Teaching deep learning and machine learning foundations: Crohn shares why he creates educational content, how his illustrated book uses visuals and analogies, and why he is building a broader machine learning foundations curriculum. Productivity and focused work: He recommends Cal Newport’s Deep Work and says he batches social media into weekly calendar blocks to preserve focus and output.

Key Arguments: Data science is a natural bridge between research and industry because it can automate pattern discovery in very large datasets. In HR, keyword search is too brittle; deep-learning-based semantic matching can identify far more qualified candidates. Production ML should optimize for user experience and latency, not just benchmark accuracy. Transformers like BERT and GPT-2 are impressive, but their compute cost often outweighs practical benefits in real systems. Explainability is important, but if a model is hard to explain, it becomes even more important to rigorously test for bias. A narrow, well-defined AI problem is safer and more effective than broad claims about general intelligence. Teaching is a way to deepen understanding; explaining concepts forces clarity and exposes weak assumptions. Focused, uninterrupted work is a major driver of professional growth and creative output.

Data Points: PhD location: Oxford - Crohn earned his PhD in neuroscience at Oxford, focusing on machine learning for brain imaging and genomic data. Data growth rate: Doubles every 18 months - He cites the rapid growth of global data as a reason automation and machine learning are increasingly important. Candidate matching lift: 21 times more - Untapped reported that its algorithm found 21x more of the best applicants than a client’s keyword search approach. Application volume: Millions per year - Some Untapped clients receive millions of job applications annually across thousands of roles. Dataset size: Hundreds of millions of decisions - Untapped built its ranking models from a very large historical dataset of invite-to-interview decisions. Gender bias test: Same score to the decimal point - Crohn says Untapped tested male and female candidates equally suited to a role and the model produced identical scores. Conference timing: End of June / start of July - The virtual Data Science Go conference mentioned during the interview was planned for late June or early July. Book translations: 6 languages - Deep Learning Illustrated was described as translated into six languages. Deep work recommendation: 2 hours a day - Crohn references Cal Newport’s suggestion that even two hours of deep work daily can produce substantial output.

Pivotal Quotes: "I think that the future is going to be really interesting. In the coming decades, data is going to facilitate enormous changes in society." — John Crohn: On the long-term societal impact of data science and the expanding role of data scientists. "The purpose is almost always to build the most accurate algorithm. And so you come up with these complex ensembles... But in practice... it’s often actually much more important to be efficient." — John Crohn: Explaining why production systems require a different optimization target than competitions or academic benchmarks. "We need to be absolutely sure that it is unbiased." — John Crohn: Describing Untapped’s approach to model validation and fairness testing in HR applications.

Implications: The episode highlights that practical AI success depends on narrow problem framing, fairness checks, and production constraints—not just model novelty. For listeners, the takeaway is to build efficient, testable, and ethically sound systems while investing in deep, focused learning.