More Related Content Similar to AI-SDV 2021: Nils Newmann - AI – Who is in control and why is that important? (20) More from Dr. Haxel Consult (20) AI-SDV 2021: Nils Newmann - AI – Who is in control and why is that important?1. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 1
Nils Newman | October 5, 2021
AI
Who is in Control and
Why is that Important
2. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 2
The Magic of AI
• AI is often sold as a magical
tool
• Click a button - - - get an
answer
• Better yet – no clicking
required
• AI just does the work for you…
3. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 3
The Reality
• AI can work like magic
But…
..
• Behind the scenes, a lot
of work goes into
making these systems
work
4. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 4
An Example
• A university is tasked to build an AI system which reads articles and
automatically tags them as AI articles or not.
• The system then further tags the AI articles based on seven sub-
domains in AI.
• The team starts with Google BERT (Bidirectional Encoder
Representations from Transformers).
• Then moves to SciBERT – the BERT trained on scientific publications.
• Next they modify SciBERT by using tens of thousands of manually
human tagged AI records so their AI- SciBERT would “understand”
the categories of AI they were trying to find.
5. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 5
An Example: continued
• Next they build another AI model that
could adapt as terms changed over time
so the system would be robust in the
future.
• Finally, they run the model against millions
of publications from different sources.
• The system produces F- scores above 0.8
for most AI categories with no human
intervention.
• So…
. Magic!
Plus a big budget for manual tagging.
6. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 6
What people ignore about AI
• The training piece of AI, whether it is
unsupervised, semi- supervised, or
supervised, is the bit that most people
tend to gloss over when talking about AI.
• Training is also one of the biggest barriers
to the wider deployment of AI.
• It is one of the big reasons why people are
suspicious about AI. Things do not work
because the training is poor.
• Ignoring the training issue contributes to
the overpromise and underperformance
of AI.
7. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 7
The SME’s
• Areas where specialized
subject matter experts
contribute to knowledge work
have been particularly
problematic for AI.
• The problem is not that AI is
not useful.
• AI can be very useful to SME’s,
but previous generations of AI
have not addressed the SME’s
primary issue…
.
8. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 8
SCALE: Small Data
• SME’s typically operate on Small Data.
• A researcher at a chemical company
might use patent information that has
been categorized using Big Data AI
such as Patent Classifications or
Chemical Tagging.
• But for the impactful work, the SME
usually operates at the Small Data level
– searches of a few thousand records.
• At the Small Data level, the SME has to
be in control of the training process.
9. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 9
The problem
• Big Data algorithms do not work in a small
data world.
• For example, most Deep Learning
algorithms need training sets in the
10,000’s or 100,000’s records.
• These types of approaches simply will
never work for the SME who uses 100’s or
1,000’s of records.
• Big Data algorithms also mean the SME has
little control over the process.
• They just use the output and have little
control over training.
10. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 10
Some Learning Solutions…
• Fortunately, this recent round of AI
development has produced some tools
to allow developers to create effective
AI for the SME working with Small Data.
• Few- Shot Learning
Deployed in VP’s Smart Trainer
• Zero- Shot Learning
Deployed in VP’s Find Similar Records
• Transfer Learning
Portable Knowledge Bases in VP
11. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 11
More Solutions…Composite AI
• In addition to new algorithms that
require only small training sets, the
latest round of AI has produced
improvements in how AI algorithms
work with each other to form
Composite AI.
• In the past, most AI systems relied on
only one approach.
• Now different AI systems can work
together to address different aspects of
a problem.
• For example, VantagePoint uses around
half a dozen different AI approaches.
12. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 12
The Possible Future: Generative AI
• Generative AI is AI that creates something new from
what it has already learned.
• Current models typically require very large training
sets so are currently of limited utility in a Small Data
world.
• But, we have experimented with using Generative AI
to write a summary abstract about a collection of
documents.
• This opens the possibility of using Generative AI to
create something approaching a proxy for
“conceptual understanding”.
• This is much work to do, but we see some
interesting potential for Generative AI in a Small Data
world.
13. Copyright ©1997
- 20 21Search Technology, Inc. TheVantagePoint.com | 13
AI, Small Data, and SME’s
• Many Big Data AI designers are trying to
replace humans.
• The work of SME’s is too important and
too “human” to be simply replaced by AI.
• AI can make the SME’s life easier and
allow them to be more effective in their
work.
• But the design of the AI systems in the
Small Data space has to start with the
human and the recognition of the value of
human expertise.
• Regardless of which AI approaches are
deployed, in the Small Data space, the
Subject Matter Expert has to be in control.