Machine
Learning (ML) and Artificial Intelligence (AI): Will AI/ML intelligence surpass humans? Part
Seven
by
Dr. RGS Asthana
Senior Member IEEE
Figure 1: Playing games with Robot [31]
Summary
Will ML/AI surpass human
intelligence in near future is the question, we wish to discuss here. There is a strong possibility that ML/AI will
replace lot of routine jobs. Further ML
and AI may not generate any solid negative outcome like extinction of humanity
as feared by many known people in today’s scenario. Use of AI may not result in
the replacement of humans for some time to come but it will happen sooner than
expected.
There is a strong possibility
that AutoML tech from Google and technologies from Facebook and Microsoft with
or without Human-in-the-loop or HITL may prosper in near future. AutoML
is without HITL tech and is good for reproduction of itself. In fact, AI is redefined with introduction of
AutoML. Further, use of ML/AI in retail is
a positive outcome. The ultimate outcome
is obtained by the fact that who uses this tech and for what purpose.
Graceful degradation of
performance is another jointly developed technology from Microsoft and
CrowdFlower where HITL is used till AI is 100% confident of the outcome. But
ultimately it will be AI without HITL as we can see on date.
The only drawback today is that
ML/AI using ANN based solutions are unable to explain why they have arrived at
a particular decision. Such an
explanation justifying the outcome is something considered necessary for wide
scale adoption of AI in certain industries like Healthcare where ethics must be
followed. Further, this is a subject of research and may take time to resolve.
Keywords
Machine
Learning (ML) Tools, Artificial
Intelligence (AI), Neural
Networks, Internet
of Things (IoT), Deep Mind, IBM’s Watson
Prelude
Computer or any automation is hardware and it becomes general purpose
only by running software as then it shows functionality provided in the
software. Similarly, ML, Deep Learning, big data and AI Technologies
produce only software with special functionality which when run on hardware, makes
it general purpose. Deep learning runs well on parallel processing machine
but today one can put a USD 1000 GPU based card on computer and make it
suitable for deep learning. Deep
learning can solve voice recognition, computer vision, machine translation and
many other problems by tuning ANN [2] provided one has enough training data.
One question is that can hardware change shape as situations require?
The answer is yes but can it reproduce itself and that too in a different shape
as required in a given scenario is a question mark. You can produce hardware in
different shapes and use it in specific scenarios. This aspect, however,
we will not discuss in this paper.
The other question is that can software be general purpose with human in
the loop (HITL) or without HITL? The answer to this question is yes in both
cases. We will discuss this aspect further in this paper. It is assumed
that HITL provides a control to ensure that everything is always in control and
any decision arrived by ML where confidence level is less than 100% is passed
on to human to satisfy many ethical issues too. But do we really need to
worry while developing AI based systems and let research in AI take its
course.
20
years ago… “Chess World
champion - Garry Kasparov - lost to computer referred to as,”IBM Deep Blue”. As per the champion [31], “Machines have a decisive advantage because, you may
call, they have steady hand. Humans are vulnerable because we cannot keep the
similar vigilance required to play with the machines, even if we understand
chess as well as machines do, even if we can survive the brute force of
calculation. At the end of the day, the pressure on the human player facing the
machine is simply unbearable”. Now the expertise of IBM machine can be put in
robot (see figure 1). In simple words,
AI is the capability of a machine to imitate intelligent human behavior. AI is
a broader concept of advanced computer intelligence on par with the smartest
human minds ever. CEO of CrowdFlower [28,
29] – which mainly solves data problems relevant to cancer research to image moderation
to sentiment analysis, combining the best of human and machine intelligence to
collect, clean, and label data, no matter its size or complexity - presented an
interesting definition of AI through the following equation:
AI = TD + ML + HITL, i.e.
Artificial Intelligence = Training Data + ML + Human in the
Loop.
Knowledge is delivered to the machine or AI Automation through
repeated explanations; the ML algorithms are so designed as to give the machine
extraordinary cognitive ability. With these inputs the machine or AI Automation
simply learns, understands and predicts. As AI/ML grows in power, it is likely
to surpass human intelligence sooner than expected.
The rule for maintaining good ratio of training
and test data is amid 65:35 to 80:20.
AI = (Tr+te) D + ML + HITL, where Tr is training and Te is
test data.
Other golden rule is that never mix training and test data. The combination
of Big Data and ML/AI techniques play an interesting and decisive role in
analyzing and handling enormous amount of data (which medical fraternity
possesses) and deriving meaningful and more accurate results.
Google and ML/AI
At Google’s
I/O conference held at Shoreline Amphitheatre, Mountain View, California from May
17 to 19, 2017, the
company put up its plans in the field of AI. Microsoft is the leader of cloud computing, but recent developments have
prompted Google’s CEO to see ML/AI as an economic opportunity and perhaps a
challenge. The announcement of second generation cloud tensor processing units
(TPU) is making cloud computing and open
source ML framework not only faster but also well-organized as compared to
generation one. The use of ML enables rapid development and promises better
Google’s future in the AI sector, putting the company at the frontline of the
ML area [14].
Figure
2: A graph with edges colored to
illustrate path H-A-B closed (green), path or walk with a repeated vertex
B-D-E-F-D-C-B (blue) and a cycle with no repeated edge or vertex H-D-G-H (red):
[16]
Long short-term memory (LSTM)
[15] (see figure 2) is recurrent neural network (RNN) architecture. RNN
is a class of artificial neural network [2] where connections between units
form a directed cycle [16] which is defined by Wikipedia as follows:
“A simple cycle may be defined either as
a closed walk with no repetitions of vertices and edges allowed, other than the
repetition of the starting and ending vertex, or as the set of edges in such a
walk. The two definitions are equivalent in directed graphs, where simple
cycles are also called directed cycles:
the cyclic sequence of vertices and edges in a walk is completely determined by
the set of edges that it uses. In undirected graphs the set of edges of a cycle
can be traversed by a walk in either of two directions, giving two possible
directed cycles for every undirected cycle”.
A LSTM network with enough network
units can compute anything a computer can
do provided it
has the proper weight matrix which, in fact, is its program. Figure 3 shows a typical RNN [17].
Modern AI systems are outpacing the intellectual capabilities of
their creators in a wide variety of fields. From beating grandmaster Go players to outguessing
cardiac surgeons, lip reading to audio transcription,
neural networks and machine learning have already surpassed humans [1-5, 6-10] -- and that list is only going to grow longer
with time.
Researchers from the University of Nottingham in the UK [10]
recently learned a ML algorithm capable of
predicting a patient's tendency for a heart attack or stroke better than both
the American College of Cardiology (ACC) and American Heart Association (AHA)
guidelines.
Figure 3: A typical RNN [17]: where BTSXVPE at the bottom of the figure denotes the input example and CONTEXT UNIT gives the output of the previous moment. So RNNs have two sources of
input, the present and the recent past, which combine to form a response to new
data. The accuracy in RNN is achieved by reducing
the back propagation of error through gradient descent.
Google's
AphaGo AI powered system has defeated the world's best human players [19] in recent years. Facebook’s AI [20] can predict how those blocks in
board games will fall just as well as people can.
Google’s [21] new Smart Reply feature uses ANNs [2] to come up with
appropriate replies to email messages. Google
says 10% of Inbox replies on the mobile app use this feature. A new approach to A.I. referred to an “AutoML” for
“auto-machine learning,” [23] it lets one AI to
become the designer of another, and direct its development without any human
intervention. It sounds like a situation
which can give the right environment for the evolution of the singularity
[24] - as
defined in futurist Ray Kurzweil's book - but it’s
actually Google’s way to place incredible power of ML in the hands of humans.
Definition of
Singularity:
'A future period during which the pace of technological
change will be so rapid, its impact so deep, that human life will be
irreversibly transformed.'
Figure 4: Pepper robots at an insurance company’s
Taipei branch, Taiwan
It's
commonly understood to be a coming convergence between humans or machines. The
Google’s self-driving cars, IBM Watson are a few known examples of an
AI system. Designers say they envisage the ML a range of usages, from working
on assembly lines to care-giving [31]. Intel
has also put their hat in AI market [33] and may compete mainly with IBM Watson
[12]. Taiwan has already
introduced a troop of mini-robots going by the name of "Pepper" [34] into
workplaces (see Figure 4). A few areas where AI/ML is playing an interesting
role are:
·
Natural Language
Processing NLP which helps in designing very
interactive and natural UIs
·
Knowledge
representation to store information
·
Automated reasoning to
use the stored information to answer questions and to draw conclusions.
·
Machine learning to
adapt to new circumstances and to detect and extrapolate patterns.
Google’s AutoML Technology [22]
Strategy
of AutoML – a project from the company’s Google Brain AI
research group [39]
– is to exploit ANNs to design other neural networks without HITL Google
allows programs to edit the code of other programs which is well within the definition of ML.
AutoML is a more full-featured version of what normal “ML”
was always supposed to be.
As per CEO of Google, “We hope AutoML will take ability that a few
Ph.D.s have today and will make it possible in three to five years for hundreds
of thousands of developers to design new neural nets for their particular
needs.”
To solve a problem with ML, a human expert provides
a starting neural network. AutoML, however, tries and tests many neural network
architectures and finds appropriate one to solve the given problem by computing
scores obtained by each network against the goal. The main idea is to find the
best mathematical approach to find a perfect solution for the problem without
the need of any human involvement. The final ANN need not use any of the
algorithms found so far but instead can include individual elements derived
from multiple architectures it has tested so far, thus, AutoML approach is able
to design more efficient neural nets.
Figure 5: Network developed by Google
Engineers (on the left) and the one suggested by AutoML (on the right) [22]
To
classify images in a large database, AutoML advised a better network (see right
picture in figure 5) though it was similar but slightly superior to the one
designed by Google’s engineers as shown in the picture on left side of figure 5).
Barriers to adoption of Machine Learning
ML is being adopted by many key companies of today viz. the
Amazon, Apple, Facebook, Google, Microsoft, Tesla and Uber who have massive
resources and are implementing projects like self-driving cars, AI assistants
and autonomous drones. For ML to become commercially viable within
mainstream businesses, we needed to address two main barriers to adoption.
Lack of High Quality Customized Training Data
ML Models need training and good test data. The results
become better if there is more and more of training and test data. Without training and test data the model
cannot learn. It’s like buying a car and there are no gas stations.
This is supervised learning as human is in loop.
ML
Models Failing Safely
So how can companies cross this deadlock? The solution is
in keeping HITL where the model predicts where it’s confident but human predicts
when it’s unsure. ML without HITL is
only possible when model has 100% confidence lavel. HITL will ensure that
there are no bad outcomes.
Machine Learning without HITL may lead to bad outcomes
Let’s take as an example classification of support tickets and
to begin with it is 100% human. The ML ability to convert TrD into a
predictive model is applied to the new inputs in this case model is further
trained on TeD, i.e., new support tickets with unstructured text that model has
not seen so far. You want the ML model to apply its predictive power to
create new outputs – in this case the “severity level” label. So HITL can
accept or reject the prediction based on the Machine’s own assessment of its
confidence level. For example, if a support ticket has words and phrases
which haven’t been seen in the training data, or seen very infrequently, then
the Machine will objectively assess its own confidence level as being low for
that particular prediction.
Then we introduce ML model and caters for only the high
confident cases which is 10-20% of the volume but humans still handle the
majority because the model not trained fully and is learning. HITL is the
critical third component of commercially viable AI. If the ML model is
not confident in its prediction it can route it to humans to review and answer.
Over time the model is trained on new training and
test data and it slowly becomes ready not only for commercial exploitation but
also precise and also sure, this in turn increases work done by the model and
reduces work done by human. Additionally, the volume of work that can be
handled by this semi-automated process dramatically increases.
The model is
AI = (Tr +
Te) D + ML + HITL (To oversee the performance and outcomes)
Microsoft has launched the joint solution with “CrowdFlower and AI
powered by Microsoft Azure Machine Learning” to leverage the joint strength of both
companies.
Digital Transformation of General Electric (GE) [36]
GE came in existence in 1897 and
established itself as an industrial and consumer products and financial
services firm. In the few past years, GE
has become a “digital industrial” company with a strong focus on the
“Industrial Internet”. In 2016, it has $7 billion in software sales. GE has achieved this feet by exploiting AI and
ML powered by Big Data.
GE’s software Predix [38] processes continuous
data acquired using scalable technologies on the cloud. It delivers actionable
insight into assets and operations, revealing new business opportunities as it
uses Big Data technology to handle data that is being captured by its
industrial device and interprets this data using AI and analytics. It has
helped GE in becoming a digital company in a big way.
Retail business and
AI
The online stores collect [32] almost all useful info about their
customers. The brick-and-mortar store
sales are slowly declining. However, use of cameras and AI tools by big
retailers like Wal-Mart to identify shoplifters are being
used to gather the same kind of data in physical stores also. Retailers are using big data and computer
vision technologies along with AI to help shoppers find what they need.
The
new point-of-sale (POS) system developed by a company called Target will be
able to search real-time inventory across the organization, arrange for
shipping, and take payment from the customer. It will record all credit card
data. The POS is expected to be in all key stores in U.S. by the end of the
year. Brick-and-mortar stores may
also use technology to get useful data about customers from their social
profiles and send customers custom-made offers. At REX, a real estate service platform, the complex problem seems deceptively
simple at first glance.
Realtors
[37] face regularly problem of qualifying leads. As the seriousness, commitment
and timeframe when sellers start thinking about putting their home on the
market to the time they actually do varies substantially and this can lead to
wasted time and lost sales. These are
situations where ML/AI algorithms help by analyzing data and giving guidance on
qualifying leads.
Healthcare [43]
In the year 2015, a
research group at Mount Sinai Hospital in New York applied deep learning to the
hospital’s vast patient records database. This data had hundreds of variables
on. The resulting program was named deep Patient. The 700,000 patient data was used for training
and when tested on new records, it proved incredibly good at predicting diseases.
Without any expert instruction, Deep Patient had discovered patterns hidden in
the hospital data that seemed to indicate when people were on the way to a wide
range of ailments, including cancer of the liver. This model gave good results
but there is no way to know how they work.
It appears that deep patient is able to correctly predict the onset of
psychiatric disorders like schizophrenia but new tool offers no clue as to how
it does this.
An MIT professor at
age of 43 years was diagnosed breast cancer but cutting-edge statistical and ML/AI
methods were not being used to help with oncological research or to guide
patient treatment. After undergoing cancer treatment it is felt that AI has
huge potential to revolutionize medicine, provided one analyzes raw imaging
data, pathology data, etc. and ML/AI system is able to give reason for its
arriving at a particular conclusion.
The most significant application of AI and ML in genetics
understands how DNA impacts life and also what influences life and biology; we
need to first understand the language that is DNA. This is where ML algorithms
play a big role and big players, such as; Google’s Deep Mind [11] and IBM’s
Watson [12] are in the fray. Technologies like big data detect patterns
from enormous amounts of data (e.g. patient records, clinical notes, diagnostic
images, treatment plans).
Deep Genomics [44] is developing the capability to interpret
DNA by creating a system that predicts the molecular effects of genetic
variation. Their database is able to explain how hundreds of millions of
genetic variations can impact a genetic code. This will provide personalized
insights to individuals based on their biological dispositions.
This trend is indicative of a new era of “personalized genetics,” whereby
individuals are able to take full control of their health through access to
unprecedented information about their own bodies. As is the case with any
application of AI/ML and also Big Data, the technology must have access to vast
amounts of data in order to better curate lifestyle changes for individuals.
Most
diseases are far more than just a simple gene mutation. Although the healthcare
system generated enormous amounts of (unstructured) data—which is gradually refining
and improving in quality—but we did not possess the necessary hardware and
software in place to investigate it and produce telling results.
Disease
diagnosis involves a number of factors, like the texture of a patient’s skin to
the amount of sugar level in his blood. So far, we have worked on
symptomatic detection, e.g. if one has a fever and stuffy nose, he is diagnosed
most likely to have the flu. Systems can now predict the molecular effects
of genetic variation [44], opening a new and exciting path to discovery for
disease diagnostics and therapies.
Way forward
Deep learning is evolving as a way to pull
intelligence out of Big Data to an extent well beyond a human brain can do or
even imagine. Google and Facebook personalize search results and Facebook pages
respectively on the fly with algorithmic approach and without HITL. Thus,
Google is able to show different search results to people for the same query. This
may have ethical issues but we are not discussing it in this article.
Google and Facebook have become true AI based
companies and together are introducing ML/AI in almost all their products and
that too without HITL. These companies get
85% share of incremental digital ad revenue and have automated video
personalization process by using ML/AI algorithms [42]. Video will become
crucial to businesses as we move forward and as per Cisco - 82% of all web
traffic will be video by 2020 - so anyone not exploiting video is likely to lose on ad revenue in the long run. Facebook
[41] is also attempting to use AI to help it remove every terrorist message and
propaganda from its platform without HITL.
It is
assumed that reader has seen references [1-5,
7-12]. The Microsoft [25] announced recently a new version of Cognitive Toolkit
(previously known as CNTK) which has lot of new features, including beta
support for Keras [35] - an open
source neural network library written in Python and used
extensively in TensorFlow and Theano - a deep learning library. This update will allow data scientists to port their computer
programs among backends of choice. Developers
can now play with ML based systems as now they have option to choose from wide
variety of tools. With AutoML technology we can revisit and define
AI through the following equation:
AI = (Tr + Te) D + ML, i.e.
Artificial Intelligence = (Training + Test) Data + ML i.e.,
now no human is needed in the loop.
To make sure that ML models fail
safely, a human is kept in Loop and till AI system is 100% confident all
doubtful decisions are vetted by the human.
Human‘s role will be close to zero when AI system becomes 100% confident
even in the Microsoft’s approach which then becomes in a way close to AutoML
model.
No one really
knows [43] how the most advanced deep learning ML/AI algorithms do what they do
and nor the AI algorithms can give any reasoning how and why they arrived at a
particular decision. That could be a problem and also area of further research.
AI is an add on technology so it can only displace human but
can’t replace them [45]. The ML algorithms are so designed as to give the
machine extraordinary cognitive ability including ability to think, read, learn,
remember, reason, and pay attention as human brain does. With these inputs
the machine or AI Automation simply learns, understands and predicts.
Figure 6: shows milestones of AI
replacing jobs from Year 2016 [26]
As per Figure 6, most of
human tasks will be substituted by AI based systems in next 120 years. As per
the figure, teachers may need to be ready for machine-written essays by 2026
and truck drivers may be replaced by 2027. AI is likely to surpass human capabilities
in retail by 2031. AI will be capable of
writing a best-seller by 2049, and perform a surgeon's work by 2053.
The outcomes shown in figure 6 are "possible but as
on date prospects of ML/AI [27] look both
exciting and scary as it is likely to affect our careers. It may have positive and efficient
outcomes as mentioned in use of ML/AI in retail and progress shown by Google in
AutoML which has capacity of reproduction in a sense that too at a reduced cost.
But there is a five percent chance of an "extremely bad" outcome also,
like human extinction as dreaded by many experts [1, 40].
There are already 120 potential game-changing ML applications
in 12 industries — and the evolution is accelerating.
References
[1] Progress and
Perils of Artificial Intelligence (AI)
[2] Invited Chapter 6 - Evolutionary Algorithms and Neural Networks, Pages 111-136, R.G.S. Asthana
in book, Soft Computing and Intelligent Systems (Theory and Applications),
Academic Press Series in Engineering, Edited by:Naresh K. Sinha, Madan M. Gupta
and Lotfi A. Zadeh ISBN: 978-0-12-646490-0
http://www.sciencedirect.com/science/book/9780126464900
[3] Future 2030
by Dr. RGS Asthana, Senior Member IEEE
[4] Machine
Learning (ML) and Artificial Intelligence (AI) – Part 1, by Dr. RGS Asthana, Senior Member IEEE
[5]
Machine Learning (ML) and Artificial Intelligence (AI) – Part Two, by Dr. RGS
Asthana, Senior Member IEEE
[6] Internet of
Things (IoT)
[7] Machine Learning (ML) and Artificial Intelligence (AI): Cognitive Services
and Robotics – Part Three by Dr. RGS Asthana, Senior Member IEEE
[8] Machine
Learning (ML) and Artificial Intelligence (AI): Big Data and 3 D Printing
– Part four by Dr. RGS
Asthana, Senior Member, IEEE.
[9]
Machine Learning (ML) and Artificial Intelligence (AI): Drones and
Self-driving Cars– Part Five by, Dr. RGS Asthana, Senior Member IEEE
[10]
Machine Learning (ML) and Artificial Intelligence (AI): Healthcare– Part Six
by, Dr. RGS Asthana, Senior Member IEEE
http://newblogrgs10.blogspot.com/2017/05/machine-learning-ml-and-artificial_26.html
[11]
Deep mind website
[12]
IBM Watson Website
https://www.ibm.com/watson/
[13] Google's
DeepMind to peek at NHS eye scans for disease analysis
[14]
Why Google is betting big on AI
[15] Long
short-term memory
[16] Cycle
(graph theory)
[17] A Beginner’s Guide to Recurrent Networks and LSTMs
[18] AI is
already beating us at our own game
[19] Google’s AlphaGo AI defeats
the world’s best human Go player
[20] Learning Physical Intuition
of Block Towers by Example
[21] HOW GOOGLE’S AI
AUTO-MAGICALLY ANSWERS YOUR EMAILS
[22] Google reveals Automatic ML: A.I. can create itself
[23] Using Machine Learning to
Explore Neural Network Architecture
[24]
Singularity
[25]
Microsoft gives developers more machine learning ammo
[26] AI
experts predict the future: Truck drivers out of jobs by 2027, surgeons by 2053
[27] The
rise of AI makes Emotional Intelligence more important
[28] Why is
AI still so incredibly powerful?
[29] CrowdFlower
Website
[30] “Garry Kasparov
on AI, Chess, and the Future of Creativity” @Kasparov63 @IBM
[31] Chess-Playing Robot Takes Centre-Stage at Computex 2017
http://www.news18.com/news/tech/chess-playing-robot-takes-centre-stage-at-computex-2017-1416773.html
[32] 3 Ways Retailers Are Using Artificial
Intelligence to Help Save Stores
[33] Intel just painted a target on IBM Watson’s back
[34]
Mini 'Pepper' robots start new jobs in Taiwan
[35] Keras
[36] How
AI and ML are helping drive the GE Digital Transformation
[37]
Machine learning that could make realtors extinct (VB Live)
[38] GE
Imagination at work
[39] Why Google’s CEO Is Excited
About Automating Artificial Intelligence
[40] How long it will take for your
job to be automated?
[41] Facebook’s AI for targeting
terrorists will go beyond Muslim extremists
https://venturebeat.com/2017/06/16/facebooks-ai-for-targeting-terrorists-will-go-beyond-muslim-extremists/
[42] Publishers
fight Facebook and Google with Video Personalization
https://www.forbes.com/sites/dadehayes/2017/04/14/publishers-fight-facebook-and-google-with-video-personalization/#49f03ef72eaf
[43] MIT Review: Intelligent Machines: The Dark
Secret at the Heart of AI
https://www.technologyreview.com/s/604087/the-dark-secret-at-the-heart-of-ai/
[44]
Changing the course of genomic medicine
https://www.deepgenomics.com/
[45] AI can displace human buy can’t replace them
