FEATURE
“We know that the previously supervised learning approach
has been partially replaced by an entirely unsupervised
learning-to-rank approach”
among developers because it’s relatively
easy to learn. It is currently ranked
fifth on the TIOBE Index, the software
development hit parade for language
popularity and adoption. As with all
successfull v1.0 software projects,
v2.0 delivered dramatic improvements.
Dean, along with a team of scientists,
worked on a refactor of DistBelief,
which became TensorFlow. Despite
Python’s relative slowness, TensorFlow
remains unaffected by this performance
limitation. Its lower-level architecture
is based on the statically
typed C programming language.
Furthermore, this low latency,
C-based interface merely bridges the
gap between the popular, high-level,
almost human readable, Python language
and the real workhorses: CPUs, GPUs
and, more recently, TPUs. This is
where the computationally expensive
operations take place.
Until GPU (graphical processing
unit), TensorFlow was built to use a
combination of hardware depending
on the required model. Specialising in
doing matrix mathematics (the iterative
and highly paralellisable computation
that performs the computationally
expensive matrix mathematics required
to train these deep neural networks),
Google has since developed an ASIC
(application specific semiconductor) against
a number of deep network operations.
Because Google open sourced the deep
learning research framework, TensorFlow
is now more than three times as popular as
its next most popular competitor.
Additionally, there are two trending
projects (at the time of writing) in
GitHub’s most popular Python open
source projects. The geometric
advancement of artificial intelligence
isn’t limited to computation and
storage costs or available labelled data.
As we can see, it may also be observed
as a result of human collaboration.
1) TensorFlow models, which include
some of the most sophisticated publicly
available pre-train models.
32
iGB Affi liate Issue 65 OCT/NOV 2017
2) Keras, which is deep learning
for Python, a layer on top of
Tensorflow that introduces
improvements in usability,
modularity and extensibility
and offers Python compatibility.
Just to keep the current public domain
capability in perspective: Inception-v4, for
object recognition, has better-than-human
capabilities given several computer vision
tasks, including the ability to accurately
identify dog breeds.
To compete with AI we require AI
The previous section illustrates the
rapid evolution of superhuman artificial
intelligence. Importantly, it is impossible
to observe a deep neural network from
the outside looking in. We know from the
Google media team, when it announced
RankBrain, the current ranking algorithm,
that the previously supervised learning
approach has been partially replaced by
an entirely unsupervised learning-to-rank
approach that is widely known to be
serving the search results.
As early as 2008, Google’s Peter
Norvig denied that its search engine relies
exclusively on machine-learned ranking.
Cuil’s CEO, Tom Costello, suggests that
Figure 1: The timeline of median estimates (with 50% intervals) for AI achieving
human performance. Each timeline represents the probability range from
25% to 75%, and circles denote 50%. Each milestone is for AI to achieve or
surpass human expert/professional performance (Source: “When Will AI
Exceed Human Performance? Evidence from AI Experts by Grace et al”,
arXiv:1705.08807v2, published 30 May, 2017)