Fantastic Problems and Where to Find Them: Daryl Weir
4 likes1,281 views
The document presents a talk by Daryl Weir, a senior data scientist, on the practical applications and limitations of machine learning (ML) in solving specific problems. It emphasizes that ML excels at answering narrow questions, particularly when rules are complex but examples are abundant, with examples from companies like Airbnb and Google showcasing successful implementations. The document also outlines the challenges of determining suitable problems for ML, distinguishing between supervised and unsupervised learning, and providing resources for those interested in starting with ML.
Fantastic Problems and Where to Find Them: Daryl Weir
- 1. Fantastic Problems Daryl Weir_ Senior Data Scientist_ And Where to Find Them
- 2. Founded in 2000 400+ employees from 22 countries 8th year in a row profitable growth YOY growth 30% Helsinki Tampere Stockholm Berlin Munich London We help our customers succeed in digital business_
- 3. Daryl Weir_ Senior Data Scientist PhD Computer Science University of Glasgow Research Intern Microsoft Finland Postdoc Researcher Aalto University Data Scientist Futurice daryl.weir@futurice.com @darylweir 2010-2014 2014 2014-2016 2016-Now
- 4. Machine Learning Hype_ AI, big data, machine learning… These have been buzz words for years now This talk tries to answer two key questions: - What can machine learning actually do? - Should I use machine learning to solve my problem?
- 5. AI/Machine Learning is not Skynet (yet…)
- 6. So what is it?_ Machine learning is really good at answering narrow questions It does this by learning from examples The most common method is supervised learning, where problems look like: Given some input A, predict an output B By choosing the right A and B, you can do some amazing stuff A—>B
- 7. For example_ Baby’s first self driving car Narrow question: given the road looks like this, how should I turn the steering wheel to not crash? Input: camera image of the road Output: steering wheel angle (CMU did this in 1995 - the car drove 3000 miles across the US)
- 8. But what kind of problem should we look for?_
- 9. “If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI” - Andrew Ng
- 10. A (somewhat trite) counterexample_ 18 27 131 94 322 1061 Are these numbers odd or even? You can (hopefully) judge this in under a second But many popular machine learning algorithms are terrible at learning this (from the numbers alone) Not all problems that can be tackled with ML should be
- 11. You might have a machine learning problem if: 1 - writing down a set of rules is hard BUT 2 - gathering examples is easy My rule of thumb_
- 12. Is there a cat in this image?_ Writing rules for the presence or absence of a cat is really, really hard BUT Finding cat pictures is really, really easy
- 13. Narrowing the focus_ What kind of things can we do? Predict the future Uncover hidden structureRecommend content
- 15. Scenario_ We have lots of historical examples of system state, each one labelled by its outcome Given a new example of the state, we try to predict the outcome Such problems fall into two main types: regression and classification Regression problems are those where the outcome is a number, for example: — how many potatoes will my supermarket sell next week? — what’s the optimal price to sell this stock? — what speed should I drive given the weather & traffic? Classification problems are those where the outcome is one of a fixed set, for example: — whose face is in this image? — is this credit card transaction fraudulent? — do these scan results indicate cancer?
- 16. Case: AirBnB_ AirBnB is a great example of a company oriented around data They blog regularly about data science and open source some of their tools One problem they’ve talked about: given the user’s search term, should we show a given listing in the results? Basic distance based search for San Francisco shows many rentals in Oakland Users were likely to find these results irrelevant Bayesian model estimates the probability of booking each listing given a specific search This sped up the booking process and significantly increased conversion rate
- 17. Case: Railways_ Two buzz words for the price of one: machine learning + IoT! Modern trains and rails are highly instrumented with sensors (motion, sound, pressure, etc.) Using data from these, we can classify whether or not a part (e.g. a wheel or door) will fail within a given time Lets the operator move from scheduled to on-demand maintenance/replacement VR Group estimates this cuts amount of wheel maintenance by around a third Bonus inception: sometimes the sensor themselves start giving incorrect readings Another classifier can detect these faults before they suggest phantom breakdowns
- 19. Scenario_ Many businesses depend on content of some sort, e.g.: — articles and videos — products in a store — ads It is often valuable to personalise the content shown to your users There are two main methods to predict content users will like Collaborative filters suggest things other users have liked The preferences of others are weighted by how similar they are to you — “Customers also bought…” — Spotify Discover Weekly Content based recommenders suggest, unsurprisingly, based on the actual content — these articles have similar topics to your previous reads — more videos from channels you watched But it’s not always easy/ possible to define similarity between pieces of content
- 20. Case: Netflix_ Netflix is one of the most well known users of recommender systems Like AirBnB, they blog regularly and have some open source tools available Good recommendations increase user satisfaction, encourage continued use, and drive users to the ‘long tail’ Famously offered a $1M prize to anyone who could improve their recommender by 10% Winning solution was never used! Recently changed from star based ratings to a thumbs up/ down system Better UX, avoids differences in users’ rating systems
- 22. Scenario_ What if we don’t have labels for our historical states? Another major ML application area is automatically detecting recurring patterns in data This is called unsupervised learning - it’s about learning what’s interesting about some given data This lets us ask different sorts of questions, like: — how do players of my game group together based on play behaviour? — is this spending behaviour unusual given someone’s past purchases? — what are the common topics in this collection of documents?
- 23. Case: Google News_ Google has been very vocal about its transformation into “a machine learning company” As of 2016, ~10% of their engineers have ML experience and internal training is growing They have stated a goal of systematically applying machine learning in all their business areas The rankings of different topics and outlets can be tailored by the user explicitly The service can also learn user preferences from usage data Google News automatically groups versions of the same story from different outlets Stories are also grouped to form news categories
- 24. Trying it out_
- 25. Starting out with ML_ Only for masochistsDesigned with data analysis in mind Amazing library and community support Not general purpose Excellent tooling - try the Anaconda distribution Jupyter notebooks allow mixed code and Markdown Can be slow
- 26. The real third option_ Fast, powerful language with excellent support for really big datasets (via Spark, Hadoop) Highly parallel Not very beginner friendly Designed with data analysis in mind Amazing library and community support Not general purpose Excellent tooling - try the Anaconda distribution Jupyter notebooks allow mixed code and Markdown Can be slow
- 27. Resources_ “I want tutorials and interesting blog posts pitched at beginners (and I don’t mind Python)” github.com/hangtwenty/dive-into-machine-learning “Python sucks, show me resources for machine learning in <my language of choice>” github.com/josephmisiti/awesome-machine-learning “Coding sucks, I just want to read a book” ‘The Master Algorithm’ by Pedro Domingos
- 28. If you remember 3 things_ 1) Machine learning answers narrow questions 2) Look for complex rules with plentiful examples 3) Don’t use Matlab Thanks for listening! daryl.weir@futurice.com @darylweir