Machine Learning Startup
Download as PPTX, PDF•0 likes•79 views
The document discusses various machine learning platforms and tools, highlighting their usability, scalability, and integration challenges within cloud environments. It outlines the machine learning lifecycle, from data wrangling and exploration to model deployment and performance monitoring, emphasizing the need for end-to-end solutions that minimize infrastructure management for users. Additionally, it touches on market dynamics, potential acquisition strategies for startups, and future directions for machine learning in enterprise applications.
Machine Learning Startup
- 1. Machine Learning Startup Ben Lackey
- 2. Cloud • Amazon Web Services (AWS) • Machine Learning • Seems usable but difficult. The most mature of any cloud vendor. • Microsoft Azure • Cortana Analytics • Seems unusable • Incredibly complex • Revolution • Surprised more isn’t going on with this • Google Cloud Platform (GCP) • Cloud Machine Learning – “Today needs a data scientist. Next needs you.” • Based on Tensor Flow • Private Beta • Image Recognition • Extremely usable
- 3. Open Source • scipy – Probably the most usable thing out there. Actually quite nice. Doesn’t scale. • R – Arcane but can be figured out. Doesn’t scale. • Spark MLlib – Not convinced anyone outside of AMP Lab can use it properly. Scales quite nicely if you don’t write bad Scala. Everyone writes bad Scala though. “It’s like Java without semicolons.” • Tensor Flow – Incomplete and only usable by Google people today. Can be used by a lay person to multiply large matrices.
- 4. Enterprise • Dato (a pleasure to use) • H2O.ai (linear regressions) • Context Relevant (featurization) • Arimo • Alpine Data • Skytree • Ayasdi • Palantir (consulting) • Scaled Inference • Data Robot • Predixion • Wise.io • Serial Metrics • BigML • Kensho • Sumo Logic • Dataiku • Brighterion
- 5. Desired Machine Learning Lifecycle (repeat) 1. Batch data wrangling 2. Data exploration • Tell me interesting things about the data. I’m not sure what I’m looking for, but I think there’s some value here. • Ideally Tableau on steroids • In real life sometimes just gplot 3. Characterize a supervised learning problem (biggest gap I’ve seen). 4. Build a model • Feature generation and selection • Training and validation 5. Deploy a model • Real time data wrangling and featurization reusing transformations from batch • Real time scoring • Proactively detect model degradation
- 6. The same old use cases… • Predictive Maintenance • GE Predix is going to wipe the floor with you • Incremental improvements on linear regressions SAS users wrote 30 years ago • Fraud Detection • Up Sell / Cross Sell • Loss Prediction / Actuarial • High Frequency Trading • How about just starting a hedge fund? It’ll be easier than selling software.
- 7. Where is the value? • Used to be in the hardware • Still lingering in the software • Next the data • Later the features • Next the model (or features of features)
- 8. Why does ML matter? • I want to predict the future. • So I can delight my customers by anticipating their desires • Brooks Brothers service at Walmart prices • Lower my operating costs through automation • And perhaps outmaneuver competitors • I can use my data to do it, now all I need is one of these machine learning things.
- 9. Where’s my machine learning thing? • Pulls in the data • Ask it to predict a variable and it figures out how auto magically. • Maybe even ask it to say what is predictable from the data (mutual information) and it says so automatically. • Scales automatically on the cloud so you can throw money at it to make it go faster. • Easily embedded in the enterprise (.jar, REST) • Explicable - you can pull up a web browser, understand what is happening and explain it to your compliance people
- 10. Problems in ML continued… • Training a model isn’t hard. It is computationally intensive. • Featurization is hard and computationally intensive. This is a large part of why neural nets are back (again)
- 11. Goal • Building a standalone business to IPO is unlikely. • Really want to get bought by a big three cloud provider and integrated into their ML portfolio. • Customers want an end to end solution. • Customers don’t want to think about infrastructure. ML people aren’t good at IT.
- 12. Ideas • Automatically characterize ML problems and then plug into existing systems • Machine learning replaces large chunks of my ERP. • Or my core banking system • Or billing system • All very hard. Require lots of domain expertise. • Consultancy – maybe cloud specific? • Generate features? Little software package to do that. Aim to build and get acquired? • Hosted ML • Easiest to build on Google. • Easiest to get bought by Microsoft. • Method might not matter. Use GBT for all classification problems.
- 13. Pricing Model • Docker image – Free, single node • Embeds training engine • Admin webpage • REST scoring endpoint • Cloud • Training – Multitenant, charge for • Storage • Compute • Scoring • REST endpoint, charge per call • Export a jar with a time bomb, charge per export (probably not the best enforce compliance) • Data Clearinghouse • Rebate on storage if data set is made public