MTC2018 - Learning to think like a Machine: How Machine Learning changes the way that we solve problems

Speaker: Jeff Hara

There are many obvious use-cases of Machine Learning like text and image analysis, but how about subtle AI solutions to everyday issues? At Mercari US, the ML team wants to change the way we solve problems, using a "bottom-up" data-driven approach. I would like to showcase our project that uses simple machine learning algorithms in a creative way and share a new paradigm for engineering solutions.

Transcript

1. Learning to think like a Machine: How Machine Learning changes the way that we solve problems Jeff Hara Software Engineer (Machine Learning)
2. Software Engineer Machine Learning Jeff Hara
3. Learning to Think Like a Machine How machine learning changes the way that we solve problems
4. You might be wondering... … What does a machine learning engineer do? … And what is he doing here?
5. Think like a machine?
6. “Neural networks are not just another classifier, they represent the beginning of a fundamental shift in how we write software. They are software 2.0” Andrej Karpathy, Director of AI at Tesla
7. Software engineering is top-down problem-solving def bake_a_cake(eggs, flour, sugar, ...): # beat the eggs eggs = beat(eggs) # make the batter batter = mix(eggs, flour, sugar, ...) # put in oven cake = bake(batter) return cake def beat(eggs): ... return eggs def mix(...): ... return batter def bake(batter): ... return cake Advantages: - closed-form solution - consistent output - human-readable Disadvantages: - not robust - closed-form solution may not exist
8. input ・X 1 ・X 2 ・X 3 ・X 4 machine learning model output ・Y 1 ・Y 2 ・Y 3 ・Y 4 database Machine learning engineering is bottom-up problem-solving
9. What data do we have? What story is being told? What problem can we solve?
10. Item & status (sold, trading, cancelled) Can we predict future status? 1. Sellability 2. Cancellability
11. Search queries Can we measure query similarity? 1. Smart search 2. Normalization
12. user embeddings item embeddings text similarity sellability cancellability user similarity search relevance item categorization item suggestion query normalization brand normalization catalog taxonomy friend suggestion fraud detection 2018 2019 subcommunities Machine learning building block road-map
13. 1. Find stories in the data 2. Solve new problems 3. Make a smarter, data-driven product The future of machine learning at Mercari