Machine learning is now everywhere, with no shortage of open-source technologies and new and updated tools released every week – from streaming databases to machine learning libraries and pre-trained models.
It’s easier than ever to train a model, but piecing together the components to deploy, optimise and scale machine learning in production is time consuming and expensive. Too many companies are reinventing the wheel by building and maintaining their own infrastructure.
Seldon helps data scientists and developers focus on solving problems by providing a fully integrated open-source machine learning platform and infrastructure. Kubernetes has transformed the way in which Seldon deploys to any cloud platform and on bare metal servers.
At ODSC London, Alex will give a technical overview of Seldon, share its transition to a fully dockerized set of containers running inside Kubernetes, and demonstrate how to deploy a TensorFlow model on Seldon’s machine learning infrastructure.
6. CONFIDENTIAL
Problems we are solving
Seldon provides machine learning infrastructure to enable
data scientists and developers to create smarter applications faster
Shortage of data scientists
Loss of control
Industry standard components.
Faster time to value.
Full control (data and model).
Deploy anywhere.
Inefficiency
Quick start, low configuration
complexity, low capital overheads.
8. CONFIDENTIAL
@seldon_io#ODSC
Demo: creating a digit classifier
1. Create a machine learning cluster
2. Build the neural network model
3. Train the model
4. Create a microservice for the model
5. Deploy the microservice
6. Make predictions
11. CONFIDENTIAL
@seldon_io#ODSC
Full stack recommendation engine and machine learning infrastructure
Quick and easy to integrate with your service:
generic REST API and drop-in JavaScript API.
customisable to work with your meta data.
microservices API for custom algorithms.
Performance optimisation to maximise KPIs.
Seldon Technology
13. CONFIDENTIAL
Rise of the containers
Containers enables efficient use of resources with encapsulated dependencies.
Kubernetes enables orchestration - easy deployment and maintenance of containers.
Platform-agnostic - works on any cloud platform or your own servers.
15. CONFIDENTIAL
@seldon_io#ODSC
Demo: creating a digit classifier
1. Create a machine learning cluster ✔
2. Build the neural network model
3. Train the model
4. Create a microservice for the model
5. Deploy the model
6. Make predictions
16. CONFIDENTIAL
@seldon_io#ODSC
Machine Learning
Stream events in real-time (i.e. metadata associated with transactions)
Create supervised learning pipelines:
Classification - yes/no (binary) or categorize (multi-class)
Regression - predicting a continuous value
Microservices API
TensorFlow
Vowpal Wabbit
XGBoost
Your algorithm!
17. CONFIDENTIAL
@seldon_io#ODSC
Recommendation Algorithms
Algorithms are combined and optimised to maximise your KPIs
1. User Clusters - improve relevance in high churn services.
2. Tag Affinity - focused tag-based associations.
3. Latent Factor Models - best for lower churn service.
4. Item Activity Correlation - built for static slowly changing historical items.
5. Topic Models - built for sites needing long tail recommendation.
6. Association Rules - basket analysis to suggest the next best action.
7. Content Similarity - for services with rich metadata and high sparsity across items.
18. CONFIDENTIAL
@seldon_io#ODSC
Advanced Optimization
Cascade/combine multiple algorithms to cover different users and use cases
control relevance, popularity, diversity
control interactiveness of recommendations
Combine algorithm results - e.g. weighted scores, rank combine.
Run A/B and Multivariate tests with no redeploy
Select algorithm strategies via API tags
to handle user cohorts: mobile users, desktop, tablet
to provide multiple content recommendations per page: site-wide, intersection
Change all configuration in real time with no redeployment.
19. CONFIDENTIAL
@seldon_io#ODSC
Selecting the best model
● Evaluation of multiple strategies
in parallel using multi-armed
bandit.
● Adaptive as context changes - i.e.
time of day, special event.
● The latest winning test strategy
(1...N) is promoted to best.
21. CONFIDENTIAL
What’s next?
Machine Learning Infrastructure as a Service
● More control. Easier
maintenance. Better
performance. Scalable.
● Algorithm and model selection.
Hyperparameter optimization.
● Built in partnership with the
world’s leading cloud providers.