How the Obscure Features Hypothesis Leads to Innovation Assistant Software
 Tony McCaffrey Lee Spector1
 Cognitive Psychology Cognitive Science
 University of Massachusetts Amherst Hampshire College
 Amherst, MA 01002 USA Amherst, MA 01002 USA
 amccaffr@psych.umass.edu lspector@hampshire.edu
Abstract
A new cognitive theory of innovation, the Obscure
Features Hypothesis (OFH), states that almost all innovative
solutions result from two steps: (1) noticing a
rarely noticed or never-before noticed (i.e., obscure)
feature of the problem’s elements, and (2) building a
solution based on that obscure feature (McCaffrey
2011). Structural properties of the human semantic
network make it possible to locate useful obscure features
with a high probability. Innovation Assistant (IA)
software interactively guides human users to the most
likely obscure features for solving the problem at hand.
 Introduction
The OFH articulates a core principle for solving the problems
requiring innovation used in psychology experiments
(i.e., insight problems) as well as real world problems in
engineering and design. As an example of an insight problem,
consider the Two Rings Problem in which you have to
fasten together two weighty steel rings using only a long
candle, match, and a two-inch cube of steel. Melted wax is
not strong enough to bond the rings together so the solution
relies on noticing the obscure feature that a wick is a string
that can be extricated from the candle by scraping away the
wax on the cube of steel before tying the rings together.
Two lines of evidence suggest that all insight problems
used in the psychology literature follow the OFH. First, the
key features for solving known insight problems are not
listed among that object’s common associates in the Association
Norms Database (Nelson et al. 1998) and are thus
obscure (e.g., string is not listed as a close associate of
either candle or wick). Second, Chrysikou (2006) had subjects
list features from the key objects used in insight problems
and found that the key feature from these objects was
listed by only 9% of the subjects and are thus obscure.
The OFH opens up a research program to improve innovation
based on two questions. What inhibits humans from
noticing the obscure? What techniques can help overcome
these sources of inhibition? In general, the normal processing
of our perceptual, motor, and semantic systems leads
us to notice typical features. Our everyday experience entrains
us towards the typical, which is efficient for everyday
life but an enemy of innovation. Moving humans from
the typical to the obscure seems to be a promising way to
improve innovation. Immediately, however, a major challenge
presents itself. The number of features of an object
(e.g., a candle) is potentially unlimited (e.g., my aunt collects
candles, candles are generally smaller than a toaster)
and perhaps infinite (e.g., a candle generally weighs less
than 100 pounds, weighs less than 101 pounds, etc.: Murphy
and Medin, 1985). What techniques can lead humans
to obscure features that have a high probability of leading
to innovation? Specifically, can a computer program steer
human users through the potentially infinite search space to
the most promising areas of the human semantic network?
McCaffrey (2011) analyzed the semantic networks of all
insight problems used in the psychology literature plus a
collection of engineering design problems and discovered
several important structural properties relevant for innovation.
In diagramming a semantic network graph for an object,
a box is drawn around the object name and all its
commonly noticed features. Analysis shows that the key
obscure features for solving the examined problems are
just outside the box. The Just Outside the Box Hypothesis
(JOTB) posits that the key feature for most innovative solutions
is one or two steps outside the box of commonly
noticed features. Searching through a semantic network, a
computer could help locate those concepts that are at the
proper distance from the original concept according to the
JOTB. This technique helps narrow the search a bit, but
another principle related to the direction of the search from
the original concept can narrow the search even more.
Although there are potentially an infinite number of features
for an object, there is a finite taxonomy that characterizes
the most prevalent types of features and relations
for an object that are known to be useful for innovation.
McCaffrey (2011) has developed such a taxonomy called
the Feature Type Taxonomy (FTT) and uses it to characterize
the most promising direction to search in from the initial
concept. The most promising direction will vary depending
on the kind of problem being addressed. For example,
an examination of all known insight problems involving
concrete objects reveals that the key obscure feature
for 68% of them resides in noticing the material composition
of the parts (e.g., examining the material make-up
of a wick and wax for the Two Rings Problem) (McCaffrey
2011). So, given a concrete object insight problem, looking
Proceedings Proceedings of of the the Second Second International International Conference Conference on on Computational Computational Creativity Creativity 120
at the material category of the parts is a good idea. Other
feature/relation types are useful for other innovative tasks.
For example, a candle is often present during romantic
events such as dinner dates. A comedian might take this
information and create a funny situation by placing a candle
in an unromantic event such as divorce court proceedings.
In other words, the kind of innovative problem stipulates
which feature types have the highest probability of
being useful (e.g., Material or Events). For every type of
problem articulated, we can create a probability distribution
that measures the likely usefulness of any given feature
type.
Together, the JOTB and the probability distributions
formed from the FTT help specify the best distance and
direction to search in from the initial concept for a given
problem. The JOTB operationalizes the idea of innovation
standing between order and randomness (Watts 1999). A
concept inside the box is too strongly connected to the initial
concept (i.e., too much order) while a concept many
links away is too weakly connected (i.e., too much randomness).
A concept just outside the box is on the boundary
between order and randomness, which seems to be
ideal for innovation.
Innovation Assistant Software
We created a software prototype, called an Innovation Assistant
(IA), which guides human users to the most promising
obscure features based on problem type. Our overall
goal is to create a human-machine interaction that is more
innovative than a human working in isolation. Whereas
traditional AI programs attempt to solve problems completely
on their own, the IA guides humans to notice the
obscure features that we tend to overlook. The human user
must then construct the solution based on the key obscure
feature.
Currently, the prototype is instantiated with the probability
distribution of the FTT for concrete object insight problems.
Further, presently it only executes a technique to
help users notice the material make-up of an object’s parts;
but other techniques for noticing the obscure members of
other feature/relation types are ready for implementation.
Evidence: Lab and Real-World Problems
We tested the IA prototype on the eight insight problems
used on human subjects in the experiments of McCaffrey
(2011). Because the material composition of the parts is
central to solving these problems, the IA asks the user to
enter this information (e.g., “Can a wick be decomposed
further into parts?” and “What material is a wick made
of?”). While the IA successfully leads users to the key obscure
features for these problems, on three of the problems
the IA exceeded its goal and actually located the key feature
and how it could be used to solve the problem. For
example, for the Two Rings Problem, after the user enters
the goal “fasten rings,” enters the candle’s parts as “wick”
and “wax,” and answers the IA’s question about what material
a wick is made of, the IA is able to generate how the
string can solve the problem. The IA suggests: “a candle’s
wick is made of string, which might be able to tie the
rings.” Basically, the synonyms of the goal verb “fasten”
are intersected with the verbs related to the potential actions
of the known objects and parts. In this case, the verb
“tie” is a synonym of the goal verb “fasten” so the IA can
make the connection and solve the problem. In sum, the IA
is designed to help users notice useful obscure features that
are often overlooked. However, the IA can periodically go
above this standard and actually solve the problem—if the
user has not already solved it.
 Most real-world problems require noticing features other
than the material make-up of an object’s parts, so we will
present how several of the other techniques have helped
with real-world problems.
First, a company challenged us to create an idea to detect
roadside bombs. Most proposed solutions focus on
either detecting the bomb itself through its metal or disrupting
the electronics of the bomb. After shifting focus
from the bomb’s physical make-up to the events that the
bomb is involved in (as suggested by the probability distribution
for real-world engineering problems), a technique
guided the user to articulate a hierarchy of these events and
their sub-events (e.g., building the bomb, transporting the
bomb, burying the bomb, and detonating the bomb). This
hierarchy led to an idea for a mechanical method that could
reliably detect the displacement of the dirt above a buried
bomb rather than trying to detect the bomb itself. No further
information will be disclosed at this time due to this
problem being an ongoing open military problem. The
technique performed its job by shifting attention to an obscure
but promising feature type.
Second, another company gave us the unsolved problem
of adhering a coating to Teflon. After switching focus to
the goal, a technique helped users to reveal the common
assumptions of the verb adhere: including (1) direct contact
between (2) two surfaces using a (3) chemical process.
Questioning these assumptions led to the novel and workable
idea of sticking something “through” Teflon to a
magnetic surface beneath the Teflon (i.e., a sandwich of
three surfaces in which the coating indirectly sticks to the
Teflon due to its attraction to the magnetic surface). Of
course, the coating would need to possess the proper makeup
in order to induce sticking. Again, the technique shifted
our focus to an obscure way to adhere things together.
Third, at the beginning of the BP oil spill in late April
2010, we worked with the goal “contain oil” and a technique
helped suggest a pipe wide enough to enclose the
spill area. After making this “pipe” out of plastic, several
engineers verified the workability of this plastic “curtain”
that reached from the ocean floor to its surface and would
guide the spilled oil to the top in a confined space where it
could be dealt with. We submitted our idea to the BP disaster
website as did other engineers.
Fourth, two candle companies verified the novelty of our
candle designs that were created using a technique which
helped us notice that candles are built out of one piece of
wax and are motionless when they burn. We created novel
Proceedings Proceedings of of the the Second Second International International Conference Conference on on Computational Computational Creativity Creativity 121
candle designs consisting of multiple pieces of wax as well
as designs in which the entire candle moves based on its
own dynamics. These designs are under consideration by
two candle companies so will not be disclosed.
Finally, the U. S. Post Office solicited the public for
money-saving ideas. Working with the goal “deliver mail,”
a technique helped us to consider places to “pick up mail.”
The option of drive-through windows led to the idea of car
commuters signing up to pick up their mail at the drivethrough
window where they get their daily coffee. Train
commuters could pick up their mail at the subway stations
they frequent every weekday. We added more logistics
(e.g., banks of post office boxes at key pick-up locations),
but estimates indicate that this method could eliminate up
to one-third of daily home deliveries without delaying the
reception of this mail.
The Next Phases
Given the promise of the IA prototype on insight problems
and other techniques on real-world problems, we are moving
forward in two directions. First, after implementing our
other techniques, we will conduct controlled experiments
with engineering and design students to further test the
IA’s effectiveness for improving innovation. Students will
be given a design task. Students in a control group will
complete the design problem on their own. Students in an
experimental group will use the IA to help complete the
design problem. Judges will rate the creativity and workability
of the designs. We hypothesize that the IA users
will produce designs deemed by the judges to be more
creative than the designs produced by the control group.
Second, we will record all the runs of the interactions
between human users and the IA program in order to try to
detect patterns in the use of the program that led to successful
innovation, as well as interactional patterns common
to unsuccessful use. To all these run traces, we will
apply a genetic programming (GP) system (Koza, 1994).
Note that GP systems have been shown to be effective on a
variety of difficult problems, including pattern detection
(Koza, 1994). Specifically, when all the parameters to explore
are known, GP systems are competitive with humans
on design problems thought to require innovation (Spector,
2008). A GP system is therefore a good method for evolving
and detecting patterns in the traces of the IA runs.
Further, the IA runs involve accessing the human semantic
network as embodied by WordNet (Miller 1995). A
GP system will be used to detect patterns in the semantic
network searches that led to an innovative (i.e. novel and
useful) connection. Using this method, we may discover
other structural properties of the semantic network that are
important for innovation. Each structural property we uncover
will permit us to craft a new heuristic to exploit the
semantic network for the benefit of innovation.
All our current research focuses on the first step of the
OFH (i.e., helping humans notice obscure features) and
ignores the second step (i.e., building a solution based on
an obscure feature). Because solutions can become intricate
causal networks, human cognitive limitations may
inhibit us from piecing together the solution once we have
noticed all the relevant features. Future research will address
this possible limitation to human innovation.
Conclusion
Initial evidence from laboratory insight problems and a
collection of real-world design problems points to the potential
for the IA approach to improve human innovation.
Testing of the IA in controlled experiments will lead to
scientific evidence measuring the effectiveness of the IA as
well as the truth of the OFH, its underlying principle. Discovery
of new semantic network properties relevant to innovation
will lead to new heuristics that improve innovation.
In sum, the IA approach is a promising one for creating
software that can improve human innovation based on
the psychologically plausible theory called the OFH.
<references_biblio/>
References
Chrysikou, E. (2006). When shoes become hammers:
Goal-derived categorization training enhances problemsolving
performance. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 32(4), 935-942.

Koza, J.R. (1994). Genetic Programming II: Automatic
Discovery of Reusable Programs, MIT Press.

McCaffrey, A. J. (2011). The obscure features hypothesis
for innovation: One key to improving human innovation.
Unpublished doctoral dissertation, University of Massachusetts
Amherst.

Miller, G. A. (1995). WordNet: A lexical database for English.
Communications of the ACM, 38(11), 39-41.

Murphy, G. L., & Medin, D. L. (1985). The role of theories
in conceptual coherence. Psychological Review, 92, 289-
316.

Nelson, D. L., McEvoy, C. L., & Schreiber, T. A. (1998).
The University of South Florida word association, rhyme,
and word fragment norms.

Spector, L. (2008). Introduction to the special issue on
Genetic Programming for human-competitive designs. In
AI-EDAM: Artificial Intelligence for Engineering Design,
Analysis and Manufacturing, 22(3), 183-184.
Watts, D. J. (1999). Small worlds: The dynamics of networks
between order and randomness, Princeton University
Press.

1 This material is based upon work supported by the NSF under
Grant No. 1017817. Any opinions, findings, and conclusions or
recommendations expressed in this publication are those of the
authors and do not necessarily reflect the views of the NSF.
Proceedings Proceedings of of the the Second Second International International Conference Conference on on Computational Computational Creativity Creativity 122