1 of 14

What Gena thinks

Steve does.

The AI Mindset��🧠

What you need to know before building an AI or ML Product

AI for Her

2 of 14

What is the AI Mindset?�

🤔

3 of 14

Traditional Software Product

AI or ML Product

v s

4 of 14

Machine Learning

Artificial Intelligence

v s

ML is a subset of AI. We will be using them interchangeably

Computers performing tasks that, until recently, required human intelligence.

Computers learning without being explicitly programmed

E.g. Recommendation engines in apps such as Netflix, YouTube, Spotify

E.g. Visual recognition, natural language understanding

5 of 14

5 Questions you need to ask yourself

6 of 14

Analyze problem

Is AI/ML necessary?

Should we use AI or ML?

Alternate options

Traditional programming, analytics

“No brainers” for AI

Unstructured data, Varied data inputs

We often see tech experts pushing for an AI engine or an ML algorithm to solve a problem. AI and ML are tools that require considerable time invested in identifying the problem and seeking a solution.

It can be easy to jump straight into the most complicated AI implementation, but you should take a step back. Look at the problem you want to solve, break it apart, and then deliberate if AI is best suited for the job. You may find that traditional programming could solve your problem.

When is AI better than traditional programming? AI tends to out-shine traditional programming in two main areas. One, when you are dealing with unstructured data (i.e. not in a database). And two, when you are dealing with fluctuating data inputs.

“Unstructured data” is data that isn’t organized in a database like photos, images, video, and sound. AI allows you to tap into this data by ‘structuring’ it into a usable format for a machine to understand.

Additionally, AI can help deal with fluctuating data points better than traditional programming. Traditional programming techniques used to organize unstructured data require developing a long list of rules that need manual hand-tuning. AI techniques are better suited here to simplify code and increase performance.

On the flip side, sometimes programming techniques can be more efficient than AI. For example, if you want to find the word “dog” in text, you could use a keyword search for that. But if you want to determine sentiment from text, then you are pushing into AI territory. Your data scientists will understand that nuance. As the product manager, you should strive for the simpler method first and be open to discussing trade-offs with your data science team.

However, it is not a binary AI v. traditional programming. Often-times, many applications use AI and programming techniques together. Once a machine understands unstructured data, traditional programming can then take over.

Let’s think about an example of a self-driving car. When AI models can take in unstructured data and identify a potential obstruction, you can program the system to warn the human driver. AI and traditional programming often work together to actualize an end to end scenario for the user. In fact, you can think about defining an AI solution as AI = ML + traditional programming.

Traditional programming and AI together can also help manage the cost of errors. For instance, if you are building a loan rating AI model in finance, you have to account for the legal risk of calling a “poor” loan “good.” You can program a rule that alerts a human to do further expert review on loans that had out of usual range inputs.

7 of 14

Do we have the right data?

80% of the project time

Is spent collecting, cleaning, preparing the data (“data janitorial” work)

1. What data do I need?

Kaggle, UCI ML Repository, Google Dataset search, Data Asset Exchange

2. Any alternate sources?

Any missing data that could bias your decisions? Complete your dataset

3. What’s more important?

Key attributes to be selected for feature extraction. Select a subset of the data for training and testing

(Your hackathon project is only as powerful as the data you bring.)

prepare your data for your machine learning model: good data is: correct, current, consistent and consolidated.

Find your niche: Niche is the intersection of Customer, Business and Data

Be careful to set expectations with the business and give your data scientists enough time to do this work!..... Do not underestimate the time it may take to secure and clean the data.

“Quality” means getting training data as close to what the AI model will see in production.

You can bootstrap this process by building interfaces to gather data or extracting data from sources like logs or search queries.

Deep learning algorithms need more data.

3. Business consequences: Building trust – will this feature help you gain or lose the trust of your users? To help figure this out, Ria suggests using FATE principles:

Fairness

Reliability & Safety

Privacy & Security

Inclusiveness

Transparency

Accountability

4. Data Understanding: Preventing bias

Comprehensive test cases (represent the real world)

Data stratification

Diverse workforce (avoid tech bro AI)

Unconscious bias – review model outputs for correlations to race and gender

During the training process, if you’re doing supervised learning, you’ll have to label the data. SMEs can help label the data, especially if the labeling requires domain expertise. You may consider outsourcing the labeling effort that doesn’t need SME input to save costs. For example, labeling the ‘noun’ in a sentence doesn’t need special knowledge. There are vendors out there that can do that for you like Amazon Turks or Figure 8. But, if you want to label “risk” in a legal document, you will need a SME like a lawyer to label that. After training the AI model is complete, SMEs can then help confirm the output of the AI model. For “unsupervised” learning, you won’t need to label the data since the AI model will find the trends itself. But, SMEs are still important because they can sanity check the output of the AI model.

*The word “label” simplifies the language here. Depending on the algorithm you use, your data scientist may use different terminology. But it’s the same concept — you need to “label” or “identify” the entities you want to extract from your data set.

8 of 14

Do we have the right team?

Diversity

Age, gender, race, income levels, etc.

Multi disciplinary

Machine Learning Engineers, Software Engineers, Domain experts/SMEs, �end-users, buyers

AI Skills and Awareness

Prepare all stakeholders involved

9 of 14

Are we building the right model?

Accuracy isn’t enough

Balance with model performance

Define metrics

Align with business KPIs or user problem

What is good enough?

Don’t aim for 100% accuracy

Is the model good enough?

Align with business or user problem

Your priorities will drive algorithm choice, UX design and AI model design.

Accuracy is key but not enough

If you are building an AI Chatbot, transcribing “A” instead of “The” may not impact that translation! Accuracy will also never be 100%.

In fact, if you do get an accuracy metric of 100%, that’s a red flag. This means that your AI model is only working on your data set and will stop working when it sees new data.

Thus, accuracy is just one data point to think about the AI model. If the bot is wrong 5% of the time but people are completing their calls and getting what they need, then does it matter? How much cost are you willing to take to improve your model? If 6 months of work makes the AI model 1% better — is it worth it? The accuracy of most AI models will also eventually plateau. Increasing accuracy X% at this stage can cost as much as the entire project to date.

10 of 14

How will we test and deploy the model?

Continuous learning

Build a feedback loop in the product

For end users and SMEs

Set Expectations

Maintenance takes time and effort

Any key changes?

For e.g. data changes -> change the model

Model Ops team

Continuously monitors and improves the model

11 of 14

Should we use AI or ML?
Do we have the right data?
Do we have the right team?
Are we building the right model?
How will we test and deploy the model?

12 of 14

I want to build an AI Product

😎

13 of 14

13

You can build AI-powered apps without any R/ Python skills

2. DIY Models

Automated ML (AutoML tools) such as Watson ML

1. Pre-trained models�Visual, Speech <-> Text, Language, Search, Decisions�

IBM’s Full Stack AI Platform�Call to Code: Data sets and challenges for key problems such as climate change, water crisis, etc. �(Check sample projects, code patterns and tutorials)

14 of 14

AI For Her

AI for Her Newsletter

Sign Up at:

https://www.aiforher.com

https://www.instagram.com/aiforhe r /

https://www.facebook.com/aiforher /

https://twitter.com/aiforher

LinkedIn Group

is on a mission to improve the gender diversity gap in the AI/ML Industry.

Join our community of current and aspiring AI professionals