Data Science is the MVP for AI Products
Update: Originally published on Medium. I've been receiving a number of questions about Step 1 from corporate innovation teams which struggle to find business cases for AI (this is a common problem for a good reason). The **key** to finding these opportunities is getting comfortable with supervised learning. Just reach out to me if you're still confused and we can jump on a quick call.
Google and Baidu are investing millions in AI to solve fundamental computing problems like speech recognition. They’re in an arms race to build AI products to revolutionize their businesses.
Your company probably doesn’t have Google’s resources. You can use Lean Startup techniques to incrementally invest in AI products and mitigate your risk.
Here is a 4-step process for using Lean Startup techniques to systematically generate value in AI projects.
Step 1 — Hypothesize the business value from Outputs
Since all AI business applications use supervised learning you can hypothesize potential Outputs.
The goal is to answer, Do we have a business case for AI?
This is the lowest-risk (and most important) first step. Just stop if the answer is “no”.
For example, CapitalOne can hypothesize fraud alerts based on past customer purchasing activity.
Inputs are the questionable purchase, past behavior, and other signals. Outputs are the alerts.
Step 2 — Run data science experiments
AI products live or die based on data.
Is the Input data correlated with our desired Outputs?
Do we have enough Input data?
Are the results good enough to generate business value?
Data scientists can use clustering and regression analysis to answer the question, Does our data support our business hypothesis?
Machine learning won’t work if Input data doesn’t predict Outputs.
Sometimes data science alone is an AI MVP. You can deploy the results without machine learning.
Step 3 — Go live with traditional machine learning algorithms
Here’s a good heuristic:
Start with the simplest algorithm which generates results.
Traditional (e.g. not deep learning) machine learning techniques like linear regression models require less data and are easier to implement.
Furthermore, engineers can troubleshoot problems by creating visual graphs of the results.
Traditional machine learning techniques reveal whether AI is feasible. Can we dynamically build models which generate Outputs from Inputs?
In many applications like Stripe’s fraud system traditional machine learning techniques work just fine.
Step 4 — Migrate to deep learning at the right time
Deep learning models use very large neural networks on high-performance computers.
Deep learning is generating so much interest because results keep improving as more data is added.
Unfortunately this performance comes at a high cost of complexity, data, and processing power.
Start by adding more data and testing the performance of neural networks to answer the question, Does adding more Input data improve our Outputs?
AI is still in its infancy and we’re all collectively learning with each new application. Lean Startup techniques can mitigate your adoption risk and deliver early wins to justify continued investment.