AI Challenges and Learnings from Experience- C2 Montreal

This is a AI Forum at C2 Montreal. This is sponsored by Element AI.

Claude Gay, IBM General Manager talks about AI Challenges.

Iterate be agile. You fail quickly and you learn fast with AI
STart now be very focused
Work with outside parties to do it
Real use cases to benefit from
Sometimes the data is not clean. This is the most common reason for failure is the quality of the data.

Here are key lessons from IBM AI engagements

1. Giving the use case for a mining drilling case. Went from 4-5 days to test a geology case but went to 4-5 hours. 8 month project.

2. AI is not here to replace us. It makes us better.
Let us solve problems we could not do before.
Medical example. Parkinson research.
Took 28 million medical records. 3800 potential drugs. Narrow down to 16 potential treatments using AI.

3. Beware of data ownership in the world of AI

The Facebook example of Cambridge Analytics illustrates the case of being careful about data ownership.

About 20% are very confident in AI benefits but only 20% are confident in their company to develop it.

Stephan Piron, Deep Learni.ng startup in Toronto, CEO

First company a big bank
Build a fraud detection algorithm
A module built by bank statistician,
DeepLearni.ng had a competition to see which algorithm would work.
The Deep learni.ng company won this confrontational setup but this caused problems by winning by an embarrassing margin. It is better to partner

Not just solving AI problems but solving business problems. Businesses have nuances that the experts and users know.

Collaboration is key.

Eric Nguyen
Raymong Chabot Grant Thornton
Senior Manager

Lessons learned applying machine learning

Need to have uncertainty and complexity for AI to bring value
Established and repetitive process to able to rely on the data
Well defined and measurable outcomes to measure the success

Context – Annual concert event, which clients should we target to max ticket sales
Outcome – Due to venue and promoter constraints, the seat preferences were not met and ended up choosing a rule based approach which performed best

The more constraints the less complex it became

Context – Financial market – how does price movement and liquidity effect stock selection
Outcome- data-related events and announcements were not fully captured. (not well-defined process)

Context – Food logistics – food for families in need, optimize route
Outcome – the actual KPI to be optimized was not determined (route time, fuel?) Wasted time.

AI Panel 0 Getting Executive Buyin for your AI project

Sylvain Carle (Real Partner) Moderator

Carolina Bessega – Stradigi AI. Chief Science Officer
Andy Mauro – Automata.ai co-founder and CEO
Foteini Agrafioti Borealis AI. RBC Chief Science Officer
Shelby Austin – Deloitte Canada National ServiceLine Leader

Recent wins

Shelby Austin – helping retailers pick product lines. Helping banks and financial stop fraud. Win is helping businesses achieve goals

Foteini Agrafioti Borealis AI
Building Apollo to monitor events in realtime. Estimating that the events will evolve and impact client companies. Using news and social media. Monitored healthcare topics. Able to predict Ebola breakout in Sierra Leone 48 hours before it was announced.

Cannot get senior exec to use twitter but you can train an AI

Andy Mauro
Business breakthroughs for partners
Work a lot of the beauty industry. L’oreal
Use conversational business diagnostics
Apply conversational AI to create -define new products at higher speed

Learning
Be very clear on your problem is key for AI.
Need more research and exploration on your problem.
Always a research component is a project.

You have to rely on data
Need to get data as clean as possible.

AI and Data Science are different.
There are quick win data science projects.
AI is very powerful. Need to look at anonomyzing data. Other ethical considerations.

Not just the tech stack.

Think about your AI journey.

Tech is second to the human challenges.
More human deployment.

What is the business value? Get a win be specific
Build or buy?
Should you partner? Who to partner with?

Exec does not want a science project.
Not commercialized then it may not have been built or others do not want it

If it has been built, then buy
If it is key to your business then look at building
Then you have to link it all together
Fit the solution into your business
Many companies have custom problems

Partner may be the worst solution
Buy solutions
Buy Companies
Buy a team

Put some thought into data collection up front
Google Duplex was irresponsible to unleash on the world (Andy Mauro)
Andy Mauro was more contrarian.

Do not use AI hammer if you do not need it.
Start on small battles and build on it.
Make sure you can get it solved in a clear amount of time X months and definitely less than 3 years.

RBC AI group functions as a startup with a single client. Build intellectual property for RBC.
Create new business lines and find ways to monetize

12-24 months
Deloitte looking to amass larger team
Look for more unsupervised looking solutions
Look for more validated systems that can be used in regulated environments