World Summit A.I. 2017 Summary
A list of insights I took away from WSAI.
I was pleased to see industry leaders dispelling a lot of hype, though there was
still a lot of hype and a lot of corporate participants trying to promote
themselves relevant in the A.I. space.
Current limitations of A.I:
- Most ‘A.I’ machine learning models start learning from scratch and thus require
large amounts of data and large amounts of computational resources.
- We lack the generalisable building blocks to do machine learning efficiently.
- Even experts machine learning engineers go through trial and error in building
models, we don’t yet have good best practices of exactly what kind of
architecture to use when.
- Models are highly sensitive to the data they were trained on. Commonly, they do
not adapt well to changes in input data.
- We lack the understanding and practical implementation of the sub-components of
learning processes to mimic humans in reasoning and generalising based on
similar amounts of data.
- Bias and stereotypes must be manually removed from the result of unfair training
data.
A.I. Landscape and interesting uses
- It will be a long time before A.I. replaces the jobs requiring creativity and
right brain thinking.
- The best uses seemed to be integrating AI technology with humans in the loop,
not replacing them altogether.
- A lot of research still into both fine tuning existing machine learning
implementations and thinking about new ones.
- While the majority of machine learning and A.I. is used for advertising, there
are people out there trying to improve humanity. From apps as a clinical aid for
depression, apps that offset suggested diagnoses of medical illnesses, automated
project management, etc. Democratising access to these resources for more people
than they could help otherwise.
- As data becomes more valuable the need to have open access and interoperability
while preserving privacy becomes more important. Restricted access to data by
only certain big corps shouldn’t stifle other research and applications.