Overview Presentation - Visual AI
Overview Presentation - Visual AI
categorical
+
numeric Raw text
from doctors
notes
date/time
Images
+ from patient
MRI
?
Needs a specialized Data Scientist.. to train a neural network.. that generates a black-box model.
Visual AI that’s Easy to Use, Easy to Explain and Easy to Get Value
DataRobot Image
Embeddings show
images that are similar
to each other
rly e
l
■ Do you have potential use case?
■ Visual AI currently Early Release
feature (GA mid 2020)
■ Wanna play w/ Visual AI?
○
■ Contact your DataRobot team for
more details:
○ Customer Support (Managed AI Caloud)
○ CFDS Team (On-Prem/VPC) Easy to Use, Easy to Explain and Easy to Get Value
○ Field Engineering (On-Prem/VPC)
See the Big Picture
Capabilities: Prebuilt models to detect common objects (e.g. Capabilities: Many solutions today use open-source platforms
human faces), geo locations, celebrities, adult-content, etc with a “do-it-yourself” approach to building custom models
Of Note: Cannot usually customize. Code only. Of Note: Hard to implement. Requires deep learning expertise.
Capabilities: General data science platforms that allow expert Capabilities: Drag and drop UI. Build custom models using
users to build custom image models from scratch automation and deep learning for non-experts.
Of Note: Users still need high level of expertise to use deep Of Note: Only DR Visual AI has true business analyst ease of use, is
learning with images. Need specialty hardware (GPUs) included with AML vs a separate product, does not require GPUs,
and allows all data types to be used together (tabular, text, images,
geo*, audio*). It also runs everywhere (on-prem + PaaS + SaaS)
Confidential. ©2020 DataRobot, Inc. – All rights reserved
Visual AI. Sample Use Cases (from Private Beta)
Risk Assessment for Home Insurance Process Automation for eCommerce Retail
Business Problem: More effectively price policies at subscription. Business Problem: Increase capacity to label product images for
website (i.e. colour, gender, category). This is currently performed
manually.
Business Goal: Increase total insurance premiums.
Business Goal: Reduce manual labelling costs and increase labeling
Visual AI Solution: Improve accuracy of the risk model by adding capacity via automated data entry.
images of properties to the mix of existing data.
Visual AI Solution: Automate labelling as much as possible using a
Implementation Detail: 1 years worth of images. Several images per classifier trained against a set of already labelled images.
row. Pictures of the roof and other exterior views added to “standard” Implementation Detail: Dozens of new products a day, ~4,250 images
model variables (num, cat, text). for training, 3 targets: Colour and Gender x Category (Jeans, Shirt,
T-shirt)
Site Selection for Gas Stations Predictive Maintenance for Oil & Gas
Business Problem: Sales reps visit garages across the US with the goal Business Problem: Oil pump malfunction occur frequently, experts are
of turning them to product sellers. They cannot visit them all so they use rare and preventive maintenance is expensive. Experts don’t trust and
an internal app to identify which ones to visit. can’t validate data-based predictions.
Business Goal: Reduce maintenance costs and downtime. Increase
Business Goal: Generate additional sales revenue by improving confidence by leveraging ML-based decisions in addition to humans.
prospect qualification process and site selection process.
Visual AI Solution: Analyze sensor data and classify defects to avoid
sending an experts on-site. “Build a model that does what our experts do
Visual AI Solution: Improve accuracy of the conversion propensity
- look at charts”.
model by adding images of the garages to the mix of existing data.
Implementation Detail: Multiclass classification (8 classes + 1
Implementation Detail: 17,000 prospects with numeric, categorical, text “Unknown”). 300 charts labelled already. Data augmented dataset
data. Pictures of the garages were downloaded from Google images. available, with 10,000 to 20,000 images. 100,000 to 1M non labelled
images overall. Images are available at various ranges, from 8x8 to
256x256 pixels.
Confidential. ©2020 DataRobot, Inc. – All rights reserved
More Potential Use Cases