← writing

Neural Networks vs Humans

Neural Networks vs Humans

“All progress depends on the unreasonable man.” — George Bernard Shaw

In 2018, I convinced a Quantium board member to give me 4 NVIDIA V100s and a year to prove neural networks could beat manual feature engineering. I was a big data engineering graduate, six months into my first job. I had no idea this was supposed to be hard.

Peak 24-Year-Old Energy

Fresh out of mechatronics, I’d built a neural network to classify recyclables and thought I was hot shit. Neural networks need big data, so I joined Australia’s biggest data consultancy.

When my manager suggested I do a “Topic of Interest” presentation, I had one thought: I’ll show these analytics guys why they’re using primitive models.

GLM Factor

Quantium ran an internal Kaggle competition called GLM Factor. Seventy analysts and data scientists. Seniors, leads, staff. Spent weeks handcrafting features. That was the entire field.

I entered with categorical embeddings and an MLP classifier. No domain expertise. No handcrafted features. Just: throw a neural network at it.

Top 5 in the company. As a big data engineering grad against the analytics team that had been optimising features for years.

That result became a one-page pitch:

“If a grad with no domain knowledge can compete against 70 analysts using embeddings, why doesn’t Quantium invest in deep learning?”

One page. Straight to the CTO. Who turned out to be a board member. He signed off on the research mandate.

What They Actually Gave Me

The Impossible Assignments

Three research areas, handed down from above:

1. Automated Feature Engineering — prove neural networks beat 12 expert analysts at feature extraction. Actually doable.

2. Neural Network Interpretability — make learned features human-readable. This is mechanistic interpretability. We still haven’t solved it in 2025.

3. Spend Prediction with 1D CNNs — sliding window forecasting across industry categories. Actually worked.

Learning Everything On The Fly

The beautiful absurdity: I was the company’s “deep learning expert” while Googling “CUDA Docker setup” and “multi-GPU Keras tutorial” at midnight.

Spark, Docker with CUDA, multi-GPU training, temporal cross-validation, production deployment. All figured out in real-time, while being the authority on the subject.

What I Built

DeepVec — 5M+ customer embeddings, 127K brand embeddings. Learned transaction representations instead of crafting features. The embeddings clustered by industry without being told to (confirmed via t-SNE). The model was figuring out that Woolworths and Coles are the same kind of thing. Nobody told it that.

1D CNN Forecasting — 12-month sliding windows predicting 276 industry spend categories simultaneously. Outperformed traditional feature engineering on temporal forecasting.

Production — 3GB models deployed to ANZ banking applications for loan default prediction, first home buyer identification, retention modelling.

The Outcome

Neural networks definitively beat manual feature engineering. Embeddings captured latent structure that handcrafted features missed. The research succeeded.

I then joined the analytics team I’d spent a year trying to compete with.

Looking back: the insight was correct, the execution was chaotic, the outcome was inevitable. I wasn’t prescient. I was just too naive to know it was supposed to be hard. That naivety is load-bearing. A more experienced researcher would have scoped it down.

For one year, a 24-year-old grad had 4 V100s, billions of bank transactions, and the audacity to think he could beat 70 analysts at feature engineering.

Epilogue: Quantium’s CEO said recently: “Quantium is now an AI company.”

I’d like to think I planted that seed in 2018.