 |
[H]ow
many different automata or moving machines can be made by the industry of man
[...] For we can easily understand a machine's being constituted so that it can
utter words, and even emit some responses to action on it of a corporeal kind,
which brings about a change in its organs; for instance, if touched in a
particular part it may ask what we wish to say to it; if in another part it may
exclaim that it is being hurt, and so on.
I recently saw the film Ex Machina. Leaving aside what I thought about the story, the characters etc. (although there was a lot to like & dislike in both), there were two aspects of the movie that were very interesting (IMHO) & which bear further thought & comment. The first was an updated idea of a Turing Test, & the second was the idea that the infrastructure & function of a general search engine could serve as the basis for a functioning artificial intelligence. An examination of these topics will cover quite a bit of current work on various AI related topics.
First, the Turing Test[3]… In
1950, Alan Turing proposed an “imitation game” for determining a machine’s
(computer’s) ability to exhibit intelligent behavior equivalent to, or at least
indistinguishable from, a human being. There are many versions of this test,
but the essence of it is that natural language is appropriate for this test
& that a human interrogator will not be able to distinguish the
conversation they are having (by text) as generated by a human or a machine.
Turing originally proposed that if 30% of interrogators were not able to
determine which other player was a human & which was a machine after five
minutes of “conversation”, the machine would be said to have passed the test.
There have been many competitions to determine if a machine could pass this
test, & in fact the Loebner Prize competition has been held every year
since 1991. No machine has yet won the full version of this prize.
In 2014, a competition held at the University of Reading
(UK) was said to have been won by the Russian chatbot Evgenyi Goostman. This result was highly controversial (surprise,
surprise) as the chatbot was engineered to represent a 13-year old Ukrainian
boy who was not a native English speaker. Nevertheless, Evgenyi convinced 33%
of the contest’s judges that it was human in a series of five-minute
conversations.
Although also controversial, the vast majority of these
contests were conducted under the model that the interrogator did not
necessarily know that one of the players was a machine. In one of the few
competitions to test this principle, the 2008 Loebner Prize (Reading
University, UK) examined the interrogator’s ability to distinguish between
machine & human even in machine/machine & human/human pairs. No
significant difference was found among interrogators who knew that a machine
was part of the test & those who did not. Perhaps more interesting was that
interrogators used criteria such as spelling errors to distinguish humans
(machines did not make such errors) & speed & length of response to
distinguish machines.
What about the movie you say… Oh yes, the movie. The primary
difference in the movie is that a human is tasked with performing a Turing Test
on a humanoid robot that is clearly not human. The interrogator & robot
meet face-to-face for the test, & even though the interrogator knows the
subject is a robot, he winds up becoming emotionally attached with disastrous
consequences. My point is not the moral & real world consequences of
emotional attachments to robots, but that this type of “Turing Test” will be
upon us sooner rather than later, & that we’ll have to have not only a
technical but also a cultural context for dealing with intelligence exhibited
by non-humans. Oh, & incidentally, not all of these non-humans will be
attractive lifelike robots – most will be entities such as massive networks of
devices, or intelligent programmatic agents… How will we deal with these
entities? Entities that, in their ability to communicate & analyze
information may not be distinguishable from human beings. Perhaps our only
clues will be just that these entities will be much better at communicating
& analyzing information. What kind of sociocultural adaptations will we
have to make in order to function in this “brave new world”? How will people,
humans, work & live alongside these entities. That brings me to search & back to the
movie.
As I have already said, one of the most interesting (to me)
aspects of the movie was the conceit that a massive, general search engine,
that was optimized for certain types of intelligence relevant to natural
language, personalization, deep learning & search, could serve as the basis
for an artificial intelligence that could pass the modified Turing Test
described above. About ten years ago (1/2006), I wrote a paper titled: If Search is the Answer, What’s the Question[4].
In this paper I predicted that in the 2012 timeframe, peoples’ work process
would be primarily knowledge & model based, & that “search” would
provide the overall structure for this work process that would emphasize
information curation & problem solving. I wrote a number of these
predictive papers in that time period & I generally got the direction right
& the timeframe wrong – I always estimated that change would happen faster
than it actually did. That’s the case here too. Now in 2015, we are not yet at
the point where models & problem solving are the primary work context for
knowledge workers, but there are several trends taking us in this direction. In
addition, there is the overhyped, but nevertheless immensely important trend of
ultra-large data set analysis (big data) that is also restructuring the concept
of work & work process. How has search changed since 2006 & where are
we in this evolution? Of course, the question raised by the movie is “does
search provide an adequate basis for general artificial intelligence”? Lot’s of
stuff to address here…
A lot has happened in search since 2006, or has it. This is
true technically, but from an end-user perspective the most visible occurrence
has been the emergence of Google as the world’s preeminent search utility. A
recent report from AYTM Market research found that 74.3% of worldwide consumers
use Google as their primary search engine. SearchStatsBrain[5]
reports that Google performed 2.1 trillion (that’s with a T) searches in 2014 –
almost 6 billion a day! The number of searches not performed on Google, Bing or
Yahoo is trivial. All three engines have deployed almost all new function, at
the end-user level, over a relatively short period of time. New capabilities
such as local search &/or vertical search are available on the “big three”,
but also through specialized apps such as Yelp or De.Li.cious. Perhaps the most
interesting, & relevant for this article, new feature is natural language
search, as offered by Google (Google Voice), Bing (Smart Search, Microsoft) as
well as by such personal assistants as Siri (Apple) & Cortana (Microsoft)
& a growing number of others. Natural language provides the underlying
interaction context for not only a modified Turing Test, but also for our
long-term adoption & adaptation to intelligent systems.
What about the other dimensions I predicted search would
have to proceed in to be able to serve as the primary structuring agent for
people’s work that was based on model interpretation & problem-solving. In
2006 I said:[6]
“During this
transition, search will have to evolve itself from the ubiquitous web &
enterprise engines of today that still mainly operate on key word & page
rank algorithms to much more deeply focused tools. These tools will be able to
refine their operation by using coarse & fine-grained models, not just of
business, but also of more general knowledge categories. They will initially be
able to structure work process because of their interactions with such models
& eventually, in the 2010-2012 timeframe, facilitate work organization
& problem solving as well as location of general or specialized knowledge in
the context of a person’s work (or personal) process. This will require the
integration with search of such areas as classification, ontology-based &
advanced metadata (currently RDF & geospatial but also evolving quickly)
modeling, rule-based reasoning & non-deductive reasoning of various forms –
this is just the beginning, but we are already seeing some of these advances in
products such as Mooter, Clusty or Grokker or in the integration of rule-based
reasoning with business process management &/or ontology-based modeling
(Protégé, Swoop). This work is just at its beginning.”
Most of this has not happened. Search engines, at least the ones
used by the vast majority of people, do not currently use models or rule bases
or any other type of reasoning, to facilitate work organization & problem
solving. Classification is provided by a very few engines that are hardly used
(compared to the Big 3). Mooter & Grokker no longer exist except as Source
Forge downloads last updated in 2007. Clusty is currently called Yippy (www.yippy.com). It appears to do
the same thing that Clusty did in 2007 – provide a sidebar of topics that the
search results can be sorted into. There are certainly specialized geospatial
search engines (GEOSS, GSE etc.), & geospatial search has been integrated
into apps such as Google Maps etc. The larger picture is that search has
progressed a lot, but its primary direction has been the facilitation a
different types of monetization, thus emphasizing the transactional nature of
search engine use.
That’s not to say there hasn’t been some movement in this direction.
Two areas do stand out: semantic search & big data. Semantic search has
been in development for a long time & can be described as follows:
“Semantic
search seeks to improve search accuracy by understanding searcher intent
and the contextual meaning of terms as they appear in the searchable dataspace,
whether on the Web or within a closed system, to generate more relevant
results. Semantic search systems consider various points including context of
search, location, intent, variation of words, synonyms, generalized and
specialized queries, concept matching and natural language queries to provide
relevant search results.”[7],[8]
The big 3 search engines all use some
aspects of semantic search. Apart from attempting to use language related
techniques such as synonym mapping, semantic search engines use additional
techniques including, but not limited to: RDF search & RDF path traversal,
keyword to concept mapping, analysis of graph patterns to identify
relationships, analysis of other (more complex) patterns, use of ontology-based
inference (OWL), use of other nonstandard logics[9].
The following figure is a relationship matrix generated by the
semantic search engine SenseBot for the query “semantic search”. It represents
important references by text size & provides hyperlinks for each topic.
Other semantic search engines
BARBARA STARR BING CONFERENCE CONTEXT DISAMBIGUATION ERIN EVERHART GOOGLE INTENT KEYWORDS KNOWLEDGE GRAPH LISTS MARKETING MARKUP MEANING MICROSOFT NAVIGATION ONTOLOGIES QUERIES QUERY SEARCH ENGINE SEARCH ENGINES SEMANTIC SEARCH SEMANTICS SEO SOCIAL MEDIA SOCIALPRO STANDARDS STRUCTURED DATA TECHNOLOGY UNDERSTANDING
Relationship Matrix (Sensebot) for Semantic Search, accessed 15 July 2015
represent results differently, some as
sidebar topics, some as graphics, but all represent relationships or semantic
groupings in some way. The user can choose what semantic dimension to explore.
In the above example, the user could investigate companies, data topics or even
languages. Search is optimized in the sense that the user gets to emphasize
what aspect of the information they were looking for is explored. Each of the
mainstream search companies have semantic search projects, & each
introduces some aspect of semantic search in updates to their product. I’ll
discuss other AI directions next.
Big data – I already hear you rolling
your eyes… I’ve been doing a project based on “big data” analysis with
healthcare safety-net organizations, & I have a much better appreciation of
the strengths & weaknesses of this approach now that I’ve been using
various aspects of it for the past year & a half. Why, you are saying, are
we talking about it here – this is supposed to be about search. Yes – correct…
the underpinnings of current big data analysis was initially all about search.
Sometime in 2002, Doug Cutting & Mike Cafarella were working on an Apache
search project called Lucerne at the University of Washington. They developed a
web indexer called Nutch that eventually was able to run on up to 4-5 nodes
& was indexing hundreds of millions of web pages, but still was not
operating at “web-scale”, even for 2003-2004 timeframe. Engineers at Google
published several seminal papers around this time[10],[11]
on the Google File System & MapReduce, a programming model &
implementation for processing very large data sets. Cutting & Cafarella
decided to use this set of technologies as the basis for an improved indexer
& rewrote their systems in Java (Google had implemented them in C++).
Cutting then joined Yahoo & over time Hadoop, the system that evolved from
the Nutch project was the basis for all search & transactional interaction
for Yahoo. By 2011 it was running on 42,000 servers with hundreds of petabytes
of storage. Yahoo spun out the distributed file system & MapReduce as open
source projects under Apache, & many other companies, research groups &
universities started developing tools, apps & applications forming the
Hadoop ecosystem. Several companies developing the Hadoop ecosystem were also
spun out, either directly or as engineers left Yahoo including Cloudera &
Hortonworks.
OK – so back to today when most
ultra-large scale projects, whether they are directly search based or analytic,
are layered on some flavor of Hadoop (or some flavor of Hadoop-inspired
software such as IBM Spark). The point, however, is not that Hadoop is the
ultimate answer for search or for analytic processing in general[12]
(it’s not…). It is that we have moved from enterprise distributed environments
that include relational databases to shared-nothing clusters with massively
parallel file & analysis systems. Those systems may be Hadoop based, or
Spark[13]
based or use Dremel[14]
for stream processing or visualization tools for presentation & visual
analysis. We are now in an era of massively parallel storage & analysis
architectures & these architectures enable a type of processing not
previously possible.
ABILITY
ALGORITHMS
APPLICATIONS
ARTIFICIAL INTELLIGENCE
BRAIN
DEEP
LEARNING FACEBOOK
GOOGLE
HINTON
IMAGE RECOGNITION
IMAGES
INDUSTRY
LANGUAGE
LEARNING
MACHINE LEARNING
MACHINE LEARNING CLOSER MACHINES
MICROSOFT
MIT TECHNOLOGY REVIEW
NEURAL NETWORKS
ORIGINAL GOALS
PATTERNS
RECOGNITION
REPRESENTATION
SOUND
SPEECH TECHNIQUES
TECHNOLOGY
There is no agreed upon, single definition of deep learning, but most people working on its development would agree that it is a type of machine learning characterized by:[16] 1) the use of multiple layers of nonlinear processing units (usually a neural net), 2) the supervised (through examples) or unsupervised learning of feature sets (patterns) in each layer with the layers organized in a hierarchy from low to high level features where the output of one layer serves as the input for the next higher layer. OK – that didn’t mean a whole lot… Actually what this technology is about is the recognition of patterns or features as information is fed through a neural net, with each layer of the net developing a more & more detailed description of the feature until an interpretation of the information can be made. A good example would be automated facial recognition. The most abstract layers of the net recognize that the overall pattern is a face, then subsequent layers identify & resolve additional “features” or “patterns” such as the mouth, the nose etc. & eventually the entire face is resolved. At this point there is a detailed digital record of the patterns so that this face could potentially be identified from a repository of faces & additional faces could be resolved, as the abstract patterns that make up a face have been defined.
There are specialized architectures for
deep learning networks & many different mathematical models & associated
algorithms for training & learning. To date, much of the focus in this area
has been on image recognition (classification) & speech processing. An
example from image recognition is the Google Brain project in which a deep
learning network learned to recognize human faces & cats from entirely
unlabeled data[17].
What does this mean for search? Imagine
a search that instead of keying on specific words or terms instead was able to
determine abstract patterns in a request & then respond by specializing
that abstraction in a particular context & return information (in whatever
form) relevant to the overall pattern or context, instead of just what it was
able to match syntactically. Some time ago, I was the lead for a project that
built a system[18]
that used much more naïve pattern recognition to return information that it
determined was “analogous” to a description given by the requester. While this
system was not very powerful compared to today’s learning systems, it did often
surprise the requestor with an analogy that it had identified. The system
provided an explanation of why it had suggested the analogy & often the
requestor would be puzzled at first but then would agree with the system once
it saw the explanation. Wouldn’t it be nice to be surprised like this by a
search engine? Deep learning will provide that capability & much more.
I’ve talked about semantic search, big
data & machine learning with reference to where search is going. It’s
currently mid-2015 – I’m only three years late with respect to the context for
search I predicted back in 2008. We are currently at search on a cusp – a cusp
that could easily push it over to the broad-scale service based on patterns,
models & analogies that assists in structuring our work (inquiry) &
facilitates problem solving in ways we might not have developed or thought of.
The ability of search to use deep learning capabilities means that pattern
recognition of all sorts will lead to models that allow hypothesis testing (as
already done by IBM Watson) & the facilitation of problem solving in
context. This latter will require frameworks (ontologies, rule sets &
trained networks) that will allow problems to be represented & reasoned
about. An additional approach will be the recognition of patterns & models
in ultra-large scale data sets & subsequent data characterization or
reasoning about the empirical data. Search using these mechanisms will be
different than it is today, I’d say it will be better than it is today; better
at finding results that match not only the content of our requests, but also
match & potentially expand their context. At that point, 3-5 years from
now, will it be capable of supporting an independent artificial intelligence.
My very strong feeling is that unless such programs are allowed to evolve on
their own, this will not happen, but then again, I guess we’ll just have to ask
it.
Up next:
·
Intelligent search, big data, deep
learning in healthcare information technology (& healthcare in general)
·
Design as a model for the evolution of
work… What will future knowledge workers be like?
·
& further in the future… a
“meditation” on the evolution of information technology using the “cluster of
terms” model[19]
[1] Image originally of the
Orion Nebula (NGC 1976), (Hubble) Space Telescope Science Institute,
postprocessed by deepdreamr.com (16 July 2015
[2] Rene Descartes. 1637. Discourse on the Method of Rightly
Conducting One's Reason and of Seeking Truth in the Sciences. Leiden,
Netherlands.
[3] Turing, Alan
(October 1950), "Computing Machinery and
Intelligence", Mind
LIX (236): 433–460, doi:10.1093/mind/LIX.236.433,
ISSN 0026-4423, retrieved 2015-07-05
[4] With respect to Danny Bobrow for his 1985
paper If Prolog is the Answer, What’s the Question?
IEEE
Trans. Softw. Eng. 11(11). – perhaps the most insightful paper on the logic of
AI languages
ever
published (with the possible exception of Doug Lenat’s paper on why AM
worked…), presented to MIT ESD Seminar Series, 3/2006.
[5] http://www.statisticbrain.com/Google-searches/
[6] I also started a piece of
work, initially in 1994, looking at design practice as a model for knowledge
work. Much of my thinking on how work would evolve came from this study. c.f.
Knowledge Work as Design: A Description of a Post-Convergence Work Paradigm.
PostTechnical Research Strategist. May 2002.
[7] https://en.wikipedia.org/wiki/Semantic_search
[11] J. Dean & S. Chemawat.
2004. MapReduce: Simplified Processing on Large Clusters. 6th
Symposium on Operating Systems Design & Implementation. 2004. 137-149. San
Francisco, CA.
[13] http://www.computerworld.com/article/2856063/enterprise-software/hadoop-successor-sparks-a-data-analysis-evolution.html
[15] NY Times: 21 May 2015, 12
June 2012; The Economist: 1 February 2014, 13 May 2015
[17] Ng,
Andrew, et al. "Building High-level Features
Using Large Scale Unsupervised Learning”. 29th Internaltion
Conference on Machine Learning. Edinburgh, Scotland. 2012
[18] GNOsys, Digital Equipment
Corporation, also see:
Hartzband, D.J.
1987. The provision of inductive problem solving and (some) analogic learning
in model-based systems. Group for Artificial Intelligence and Learning (GRAIL),
Knowledge Systems Laboratory. Stanford University. Stanford, CA, USA. 6/87.
Hartzband, D.J. 1987. A discussion of inference
and problem solving in the GNOsys knowledge model. Problem Solving Systems
Group. Artificial Intelligence Technology Group. Digital Equipment Corporation.
5/87.
Hartzband, D.J., L.
Holly, and F.J. Maryanski. 1987. The provision of
induction in data-model systems: I. Analogy. International Journal of
Approximate Reasoning (IJAR) 1(1):1-17.
[19] Foster, Hal. 2015. Bad New Days: Art, Criticism, Emergency.
Verso. NYC. 208 pp.
