SATURN 2019

Machine learning (ML) is everywhere, and there is a lot of material that discusses many of the difficulties associated with creating an ML solution from a data science point of view. This work covers aspects related to obtaining data, selecting a particular algorithm, and training and testing the algorithms. There is, however, less information related to architecting the software system where this algorithm will be running once it is in production. In this talk, we want to address this topic, both from a theoretical and practical point of view. Our goal is to help software architects that need to design systems that support ML by identifying common architecture design considerations for various phases of data processing, training, and model serving.

We will cover the following topics:

1. Where does software architecture fit in systems that support ML?

2. A framework for gathering primary architectural decisions of systems that support ML. These decisions include aspects such as
* type of training of the model and model location
* time of training: offline vs. online
* time of prediction: batch vs. on demand
* location of prediction: cloud vs. device
* technological choices
* other considerations

3. Several case studies of systems developed at SoftServe that support ML using the previously discussed framework.

4. Synthesis: What needs to be considered when architecting ML systems?
* Lessons learned
* Design process considerations

See the slides.

Watch the video.

Size of the AI/ML global opportunity
a. It is believed that artificial intelligence (AI) and machine learning (ML) will contribute $13 billion to the global economy by 2030.
b. The following six companies have collectedly invested over $7.7 billion in AI/ML research and development: Google ($3.9 billion), Amazon ($871 million), Apple ($786 million), Intel ($776 million), Microsoft ($690 million), and Uber ($680 million).

Deficit of trained data scientists’ vs. open opportunities
a. According to the August 2018 LinkedIn Workforce Report, there is a national shortage of 151,717 people with data science skills.
b. IBM predicts that by 2020, the number of jobs for all U.S. data professionals will increase by 364,000 openings to 2,720,000.

Development of tools to make it easier for non-data scientists to do AI/ML
a. Six of the most common ML use cases are as follows: Customer Lifetime Value Modeling, Churn Modeling, Dynamic Pricing, Customer Segmentation, Image Classification, and Recommendation Engines.
b. Talk about ML APIs, Keras, and AutoML as enablers of addressing the “turnkey ML” use cases

Introduction to Use Case: Autonomous Inspections of Aircraft
a. Market opportunity for infrastructure inspection is over $45.2 Billion, with predictive maintenance specifically growing 400% by 2022.
b. Similar solutions have seen the following results: 10x reduction in time spent in inspections, 20% to 30% reduction in maintenance costs, and 15% to 20% reduction in unplanned downtime.

Breakout of the Google tools used in the demo and what each one does
a. Pub/Sub
b. Cloud Storage
c. Machine Learning APIs/AutoML
d. App Engine
e. Google Street View API
f. Google Earth API

See the slides.

Watch the video.