Madhura Dudhgaonkar is the senior director of Machine Learning at Workday Inc. She believes that it’s possible to deploy machine learning within your enterprise, but it takes a few steps to get exactly right. She loves to get into unknowns and things we haven’t tried yet, but let’s look at a few ways her team has deployed these solutions.
[Related Article: Watch: Challenges and Opportunities in Applying Machine Learning]
Workday: A Single Cloud System
Workday allows you to leverage API. The power of one helps your silos connect. It can be so challenging to do simple things like finding a budget for next year. It puts the power of different systems right at your fingertips to bring your department together. They had to convince the market to adopt the trend of cloud architecture, but now, they’re a leader in this space.
The Challenge
They needed to build an ML product that could scan receipts and auto-populate fields in a computer program. They only had six months, starting from scratch and working across multiple disciplines. Anyone who’s belonged to an enterprise knows this type of truncated timeline is difficult.
They accepted the challenge and followed the following steps:
- Define the win: for them, it was an ML product that scanned receipts with 80% accuracy and delivered in private and public cloud by a specific timeframe.
- Build transparency: customers are instinctively distrustful of the cloud, so you must establish that trust when you’re building (i.e., they didn’t give out data to third-party developers.)
- Utilize transfer learning: They built synthetic data sets to mimic the noise of receipts. However. It wasn’t accurate enough.
- Run a receipt contest: The prize was an iPhone. People sent in old receipts, so they had an authentic data set. They ended up with around 30,000 receipts, plenty for transfer learning.
- Choose the framework: They had criteria for their own use cases, and she encourages all organizations to decide their criteria.
- They used a deep model network by building a bounding box detection system. They moved to text recognition of what was inside the bounding box. With both, they used a deep neural network. The final stage was mapping. They had rule based options and deep learning ensembles with multiple layers. The map of values received allows them to automate the text fields.
Keep Humans in the Loop
The success is attributed to a core group of humans that provide the data cleansing, transformation, and labeling to allow them to improve their model accuracy. Keeping humans in the loop transforms models from total black box to somewhat transparent. If you’re building products with a lot of cross-functionality, you’ll need a bit of human intervention to keep that validation.
Workday uses a core human team that can access the type of sensitive data they receive from their customer. The UI is developed in house, and the team is in charge of things like proof of concept or validation of results.
The architecture itself moves from UI and non-ML services to Machine Learning as a service. The platform engineering elements are common across all ML services. Those services help build a platform that can deploy at scale.
Scaling from 0 to 1
If you’re going to do your own enterprise solution, you need a framework. Here’s what she suggests:
S – Select one win: Be precise like they were above with the 80% accuracy benchmark.
T – Start with the team: Pick the right leader and then move to other people on the team. Go for an entrepreneur or people who have achieved 0 to 1 already. It doesn’t have to be ML.
A – Articulate and Align: Tell everyone in the company your win and game-plan and make sure your stakeholders are on board.
R – Rally and Support: Keep your team on board. Help with support. Protect your team and make sure they aren’t distracted by other issues on the business side.
T – Take Shortcuts: Operate as if you’re a bootstrapping startup that only has six months. Make your team realize that every day is vital.
[Related Article: Automating Data Wrangling – The Next Machine Learning Frontier]
From 0 to 10: Moving Forward
Once you get from 0 to 1, you have a better position to go from zero to ten. You have some credibility now, and you have some resources. Here’s how you can move further to your scale.
G – Gather more capital: Your resources are a vital part of continuing your work. You can get fill out your team and get more significant projects.
E – Establish repeatable processes and platform: There will be churn, so you must operationalize your process to combat loss of knowledge.
T – Transfer learnings to scale to ten: Choose things that can transfer from project to project.
