PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

The economy for artisanal products such as Navajo rugs or Pashmina shawls are often threatened by mass-produced fakes. We propose the use of AI-based authentication as one part of a larger system that would replace extractive economies with generative circulation. In this case study we examine initial experiments towards the development of a cell phone based authentication app for kente cloth in west Africa. We describe the context of weavers and cloth sales; an initial test of a machine learning algorithm for distinguishing between real and fake kente, and an outline of the next stages of development.
Content may be subject to copyright.
Authente-Kente: Enabling Authentication for Artisanal Economies with Deep Learning
Accepted August 19, 2020 to AI & Society
DOI: 10.13140/RG.2.2.27020.95362/2
Corresponding author: Kwame Porter Robinson (, Graduate Student,
University of Michigan, School of Information, 3360 North Quad, 105 S. State St., Ann Arbor, MI
48109-1285. ORCID#0000-0003-2663-571X
Ron Eglash (, Professor, University of Michigan School of Information.
ORCID# 0000-0003-1354-1300
Audrey Bennett, (, Professor, University of Michigan Penny W. Stamps
School of Art and Design. ORCID# 0000-0002-6763-2622
Sansitha Nandakumar (, Graduate Student, University of Michigan School
of Information. ORCID# 0000-0001-8651-1415
Lionel Robert (, Associate Professor, University of Michigan School of
Information. ORCID# 0000-0002-1410-2601
The economy for artisanal products such as Navajo rugs or Pashmina shawls are often
threatened by mass-produced fakes. We propose the use of AI-based authentication as one
part of a larger system that would replace extractive economies with generative circulation. In
this case study we examine initial experiments towards the development of a cell phone based
authentication app for kente cloth in west Africa. We describe the context of weavers and cloth
sales; an initial test of a machine learning algorithm for distinguishing between real and fake
kente, and an outline of the next stages of development.
Keywords: human-machine collaboration; machine learning; artisanal economy; generative
justice; industrial symbiosis; ethnocomputing
This research is supported by National Science Foundation (NSF) grant #1640014 and Mcubed
grant #8330
Conflicts of Interest
The authors have no conflicts of interest to declare. All co-authors have seen and agree with the
contents of the manuscript and there is no financial interest to report.
Availability of data and material
Data used within this work is available at
Code Availability
Code used within this work is available at
Mass production economies have introduced many ills into social life, including high rates of
mental and physical illness from dull repetitive jobs; high rates of environmental degradation;
and a “junk food” approach to both physical and informational over-consumption (Michelsen and
Bildt, 2003; Coccia 2017; Hunt et al 2018). Artisanal fabrication, in contrast, tends to embody
the opposite effect. Artisans often report that they are drawn to their craft because it is an
enjoyable and rewarding form of labor. Traditional fabrication methods often use locally sourced
and sustainable supply chains. And (at least traditionally) artisanal items were purchased in
more thoughtful ways, often establishing a personal relationship between buyer and seller. In
our prior work (Eglash et al 2019), we suggested that AI, robotics and other forms of
automation, if properly designed and implemented, could gradually scale these beneficial
systems towards the development of an artisanal economy. One small step in that direction
might be AI guides that help connect consumers with artisanal producers. In this paper we
explore a prototype, Authente-Kente, to help guide consumers toward selection of authentic
hand-woven kente cloth, and thus diminish income loss due to mass produced fake cloth.
The Problem Context
Traditional artisanal items often compete with mass-produced fakes. M’Closkey (2010)
estimates that out of roughly 2 billion in annual sales of “Native American” goods, about 50% is
not actually of Native origin. Similar problems arise elsewhere: for example, Mehra (2019)
reports that due to competition from mass produced fakes, the number of Pashmina shawl hand
weavers had dropped from over 100,000 in 2007 to about 10,000 in 2017. The artisanal product
of concern in our case study is Kente cloth, a traditional fabric from the West African nation of
Ghana. Kente is of great symbolic importance as a symbol of African heritage, not only in
Ghana but also abroad. African American students often wear a kente stole at graduation. In
January 2018 members of the Congressional Black Caucus wore kente cloth wraps and scarves
to the State of the Union Address to protest the rising racism in populist politics (figure 1).
Figure 1: Members of the Congressional Black Caucus wearing kente on Jan 30, 2018.
As in other cases of artisanal production, authentic hand-made kente cloth is in competition with
lower cost imported fakes. Howard et al (2016) note the steep decline in Ghana’s textile industry
due to fake imports: from about 25,000 employed in 1977 to under 3,000 by 2005. Revenue loss
from fake imports of all products (including textiles) is now estimated to be over $1 billion in
Ghana (Boateng 2019). On the positive side, Ghana’s tourism industry contributed $2.7 billion,
or 6.2% of the national GDP, in 2017 (Oxford Business Group, 2019). If tourism could be tied to
selective purchase of authentic, locally produced textiles, declines for that industry could be
reversed. Thus a cell-phone based AI guide for directing tourists towards purchasing authentic
textiles seemed like a promising prototype for experimentation in this area.
Prior work on automated authentication of artisanal textiles has not focused on the tourist trade.
Rather they are typically made for professional export industries. One exception is the case of
Pashmina shawls: the Kashmir government set up a six million dollar testing facility to ensure
that the fibers are purely composed from hair of the Pashmina goat (Capra hircus), and issued a
microchip-enabled authentication seal (Parvaiz 2017). However these are mainly serving large
high-end tourist shops, where shawls are sold for as much as $1000 each. Even then, decoding
the chip requires an infrared light reader before its code can be checked with an online
database. On the whole, the system is somewhat akin to visiting a diamond merchant and
asking for authentication of the gem’s intrinsic value with respect to provenance and crystalline
On the one hand top-down authentication systems are very helpful in two circumstances. Firstly,
in cases of high intrinsic value. Diamonds have a laser-etched serial number if they have been
certified (Carl 2016). Individual diamonds are worth thousands or millions. Secondly, for a highly
regulated industry. Pharmaceuticals may be low cost as individual packets, but the industrial
wealth and the power of medical institutions are such that they can afford to maintain and
manage a highly regulated system of suppliers and retailers.
Kente cloth weavers, on the other hand are in neither category. Their individual weavings sell
for relatively low cost and as a group they are neither regulated nor wealthy. Thus preventing
someone from appropriating a barcode meant to be used by weavers and applying it to factory
fakes would be very difficult. Indeed one of the best known modern texts on kente cloth is titled
"That Copyright Thing Doesn't Work Here" (Boateng 2011).
Kente cloth, in contrast, is generally created by low-income communities in Ghana and sold at
relatively low prices. The confusion for tourists comes primarily not from duplicate weavings, but
rather kente prints. Mass-produced in factories at very large scales, it replicates the geometric
patterns and colors as an image printed on ordinary cloth. We will refer to these as “fake” kente,
because from a weaver’s perspective it is marketing a phoney version of their craft for lower
prices than they can compete with.
What motivates tourists to take efforts to guarantee the
authenticity of weaving in the case of Pashmina--the perception of a high intrinsic value due to
specific types of fibers and process--is clearly not present in the case of kente prints, at least in
the context of large markets where we have observed this confusion first hand.
On the other hand, tourists in the presence of a weaver clearly value the cloth for its hand-made
character, even when the fake print version is also available at lower cost. Communities
devoted primarily to kente weaving have had some significant success. For example, Edusei
and Amoah (2014) examined the labor demographics for the Kwabre East District of Ghana,
which includes the villages best known for kente weaving. Out of 27,000 households, an
estimated 10,000 jobs were in the textile industry (which includes weavers as well as thread
suppliers, tailors, retail vendors, etc.). While we tend to think of tourists visiting a weaver as
somewhat artificial or contrived, such personal encounters between maker and buyer are
arguably closer to the Indigenous economic tradition than, say, shopping at Ghana’s Accra mall
(which is modeled on American malls, including food courts, security guards and so on). In the
original Indigenous economy, value circulated in the form of collaborative labor groups, rituals of
gift exchange, a shared resource commons, and other elements of what is sometimes referred
to as “relational economies”. Curry and Koczberski (2012), examining how aspects of relational
economies still exist today alongside commodity economies, note:
In these transactions, an intrinsic relation exists between the item of exchange and its
donor, in that the exchange item can be conceptualised as constituting ‘parts of persons’
themselves.... From this perspective, the gift contains the embodiment of the donor and
is never fully alienated from the person as in commodity transactions.
A case can be made that the print is not intended to be mistaken for handwoven, and therefore
legitimately sold. But we have observed tourists failing to grasp the distinction, so the impact on weavers
is the same. There are also distinctions between “authentic prints” made in Ghana, and those produced in
foreign countries and smuggled across the border, but that is beyond the scope of this paper.
This helps us see what is happening when hand-woven kente is selected by tourists in areas
where they see it made, in contrast to the commodity transaction with cheaper kente prints in
places like Accra’s National Center for Culture, which is Ghana’s largest tourist market. Unlike
Pashmina, which can claim value in the material properties (as a commodity), exporting hand-
made kente to sites of mass consumption like a mall breaks the relational link that was (and is)
a crucial source of its value. Thus it makes sense to seek an authentication process that can
operate in more generative modes. To merely distinguish between real and fake might be
appropriate for commodity marketing, but if AI is to play a role in the transition to a generative
economy, it needs to facilitate a richer set of producer/consumer relationships.
If we think about the commodity-based, mass manufacturing system as something that has
broken relationships, then a system for more generative forms of authentication might be
understood metaphorically as a kind of prosthetic for replacing missing parts and restoring
functionality to the social organism. To this end we envision a system in which artisans can use
cell phones (ubiquitous in Ghana) to upload making-in-action videos, conversations with elders,
comments on favorite weaving patterns, symbolic meanings of colors and shapes, or other
media representations that tell their story from the production side. From the buyer side,
automated identification of a textile would enable locating a best guess for its point of origin,
access to the producer’s media (including identification of the pattern’s symbolic meanings), and
perhaps even opportunities for two way communication (“please be a guest speaker in my
class”, “my wife wants to show you the beadwork she sells”, etc.). And there is no reason this
must be restricted to textiles; any purchase could be approached in this way, allowing more
insightful and purposeful buying and selling.
Setting aside for the moment this ambitious vision of what generative authentication might
become, a basic functionality for distinguishing real handwoven versus fake printed patterns
seemed like the most fundamental first step. Below we describe our initial experiments using an
Authente-Kente task specific machine learning pipeline to make this basic distinction.
In selecting the AI method, we began with the observation that the problem of distinguishing
between authentic handwoven and fake (printed) kente patterns is broadly related to anomaly or
fraud detection (Hodge and Austin 2004). There are many possible approaches to automated
anomaly detection (Das 2009); we narrowed our choice by taking into account two constraints.
First, we can expect the proportion of authentic to fake instances will vary greatly by site.
Typically, tourists would encounter far more instances of printed kente in the Accra mall than
would in the village of Asonomaso, within the Kwabre East District, where authentic kente is
ubiquitous. In other cases the ratio may be closer to 1:1. Thus it is important that Authente-
Kente perform well against both authentic and fake instances of kente cloth, independent of
authenticity prevalence. This raises an unusual challenge in that most statistical anomaly
derivations assume a fixed ratio of anomalous/normal cases derived from a static application
domain (Das 2009). Our framework for generative authentication, in contrast, requires that
solutions work across a variety of contexts with varying and unknown ratios of anomalous
cases. Thus we began prioritizing success metrics by assuming equal prevalence, keeping in
mind that re-assessing this assumption might be needed as data comes in from real-world
Second, as a practical matter we had a limited number of authentic kente cloths available as
samples. Extensive field work might slowly accumulate more, but nothing like the thousands of
cases required if we wanted to use whole cloths as the samples. The relatively high number of
sample cases is typical in machine learning, particularly Deep Learning (LeCun, Bengio, and
Hinton 2015), where convolutional neural networks (CNNs) have facilitated empirical success
for complex visual recognition tasks. Some strategies have reduced the training set size: in
Yadav and Jadhav (2019) for example, automated medical image classification for rare disease
diagnosis out-performed trained doctors, using only 5,232 training images. However that is still
magnitudes of order greater than our sample size of a dozen or so cloths. To resolve this issue
we created many local samples of each cloth; we refer to these as “swatches.” This was
facilitated by the visual complexity of kente, which has a wide variety of geometric design
elements, grouped in particular arrangements, and organized in arrangement patterns. We
followed exactly the same process with fake whole cloths, deriving sample swatches from them
in the same experimentally controlled manner. Finally, we note that in a real-world situation,
tourists will likely find it easier to casually take cell phone photos of folded cloth (swatches) than
ask that each cloth be fully unfolded and extended.
Pipeline Overview
As noted above, our initial prototype addresses the problem of guiding a tourist towards
authentic hand-woven kente cloth as a socio-technical decision problem to be supported by an
efficient machine learning pipeline. The problem domain defined by this envisioned future
application, illustrated in Figure 2, shows the user taking a close up photo of a whole kente
cloth--i.e. a swatch--and getting back an authentic/fake classification from the application. In our
laboratory mockup, we begin with a whole cloth, which is broken into swatch images (artificially
creating what real-world users would do) and provide the swatch to the machine learning
pipeline. Training on swatch samples for both authentic and fake cloths then produces a reliable
classification system using a CNN. We experimentally investigated the success of the pipeline
on randomly generated swatch images as described in later sections.
Figure 2: Current Experimental Problem Domain
Whole Cloth Validity
We generated swatches from 16 whole kente cloth samples, given in Figure 3: eight authentic,
and eight fake (i.e. machine printed), and represented in Figure 4: sample grid and tonal color
distribution of fake and authentic swatches. Authentic cloths included 7 museum pieces, and
one collected by one of the authors in the village of Bonwire in 2011. The fake samples were
sourced from a variety of manufacturers. The typical whole cloth sample was imaged at least
2000 x 2000 pixels, with the smallest sample spanning 633 x 633 pixels (museum, authentic)
and the largest sample spanning 2981 x 2981 pixels (in-hand, fake). Environmental lighting
conditions for fake whole cloth photos were kept similar to that of the authentic whole cloths
through omnidirectional lighting without occlusions. All whole cloth samples were square. Table
1 contains the whole cloth specifications.
Figure 3: Eight whole kente cloth samples in each group (fake and authentic)
Environmental Conditions
Sample Size
Dates &
Times of
By Type
1/23/2020 @
~ 5pm
5/1, 5/15 &
6/18 2020 @
~5 pm
Table 1: Sample environmental condition and sizes
* Omni refers to omnidirectional natural lighting without lighting hotspots
Swatch Construct Validity
Figure 4.A and Figure 4.B radially lay out
swatches in a mosaic from lighter to darker tones
without replacement starting from the center. In Figure 4.C the red, green and blue (RGB)
swatch tonal histogram of both authentic and fake swatches share similar shapes although the
fake swatch color is more peaked at the mid-dark tones. Taking a standard model of human
color perception into account, in Figure 4.D, the LAB
histograms also indicate similar shapes.
The LAB differences include a peak in L* (lightness) values for fake swatches as compared to
authentic swatches, wider a* (green/red) distribution in fake swatches and a sharper b*
(blue/yellow) distribution in authentic swatches. In Figure 4.E, as a simplified cylindrical volume,
fake and authentic swatches are virtually the same. Only the center of mass differs and mainly
along the L* component.
In aggregate, these differences are visually identifiable as in Figure 4.A and Figure 4.B as
authentic swatches having less “saturation” relative to the artificial inks of fake swatches and
that fake swatches have darker inks that are somewhat darker than authentic swatches. In
addition, we can see that the range in 4.A is relatively smooth, while that of 4.B changes by
discrete steps. Adherence to traditional patterns (4.B) versus introduction of invented patterns
(4.A) may be causing this effect. Another possible cause is that weavers are purchasing thread
from similar supply chains, and these threads tend to be offered in relatively discrete categories
(the 6 primary and secondary colors). In contrast textile inks are marketed to manufacturers in a
continuous range of colors. We have endeavored to ensure that these color range differences
are not accidentally added by lighting, camera angle or other external factors, as noted in our
Whole Cloth Validity section above.
The dissimilarity of Figure 4.A and Figure 4.B may give the impression that it will be easy to
identify a particular swatch as in either the fake or authentic category. But these are showing us
the aggregate characteristics. Any single swatch, as a user would examine in a market, is a
single realization of color variables drawn from largely overlapping distributions. Viewed
statistically, the overlapping distributions induce similar mean and variance statistics that
prevent LAB components and tone from being discriminative enough to achieve high
classification accuracy. This discriminative difficulty is mirrored by our original problem
Using a mosaic layout algorithm by dvdtho available at
The CIE L*a*b* color space model, abbreviated as LAB, is a non-linear model based on human
color perception and expresses color in terms of L*, lightness, from black (0) to white (100), a* from
green to red and b* from blue to yellow.
motivation that tourists, using color alone, are not reliably able to determine if a cloth is
authentic or fake. This suggests that a more powerful model is needed.
Lighter Tone
Darker Tone
(A) Sample fake kente swatches
Lighter Tone
Darker Tone
(B) Sample authentic kente swatches
(C) Histogram of sample tonal intensity
(D) Histogram of LAB components
(E) Approximate LAB volume extents and center of mass for kente swatch samples4
Figure 4: Sample grid, tonal color, LAB color histograms, sphere of fake and authentic swatches
In addition to the challenge of making the fake/authentic distinction from these sets of swatches,
we can expect that tourists casually taking images will introduce many sources of natural
variation. Data augmentation (Shorten and Khoshgoftaar 2019) provides a controlled method for
post-processing to be consistent with sources of natural variation. We endeavored to replicate
the natural image variation through data augmentation that randomly applies minor image
transformations to each of the 2,000 swatches. These mimicked the real-world variability in the
angle of the photo, the type of phone, and the environmental lighting.
Unless the user’s wrist holds the phone perpendicularly the resulting swatch will be oblique to
the camera lens. Therefore we randomly rotate a swatch along three dimensions (the axes
, ɸ,
). The rotation of the plane formed by axes
and ɸ simulate rotation parallel to the kente cloth
and the rotation along the axis
simulates cell phone camera forward rotation formed by the
natural tit of the wrist. The image sensor in the phone has a unique color profile such that
swatch color can subtly differ across phones. We randomly and slightly perturb the red, green
and blue color channels of the swatch image to induce variations in color. Finally, the user
Photo Reference Credit “About Color Management,” Sony. 2020. Accessed:
location creates the possibility of shadows falling on the swatch. Each swatch image has a
randomly projected shadow on it ranging intensity from 0% (invisible) to 10% in lightness.
Source code for generating this dataset from scratch
is openly available. We account for these
expected sources of natural visual variation through data augmentation. Figure 4 lays out
swatches from lighter to darker. The red, green and blue (RGB) swatch color distribution of both
authentic and fake swatches share similar shapes although the fake swatch color is more
saturated (e.g. vivid). That the tonal color distributions largely overlap another means that color
(including hue , lightness and saturation) alone is not discriminative enough to distinguish
between authentic and fake kente.
Dataset Generation
We took care to create a balanced dataset of authentic and fake kente swatches. From each
whole cloth 125 swatches are generated, per Swatch Construct Validity. Swatches of size 224
by 224 pixels are randomly extracted from whole cloth samples to create a balanced dataset of
1000 authentic and 1000 fake swatches. The data are split into training (750 cases), validation
(750 cases) and evaluation (500 cases) datasets. Experimental Data Evaluation discusses the
partitioning of the dataset in additional detail.
A Brief Background on Convolutional Neural Networks (CNNs)
Progress on ImageNet (Russakovsky et al. 2014), a classic computer vision object recognition
challenge involving thousands of objects in natural settings, had largely stalled in the field of
computer vision in early 2010. However, convolutional neural networks (CNNs), first used by
Alex Krizhevsky et al. (2012), showed the potential to dramatically improve state of the art
performance (in their case on the ImageNet Object recognition task). CNNs are a type of deep
learning modelled after the human visual cortex and achieve their high performance by mapping
pixels to hidden neuron layers. The hidden layers provide a variety of convolutions
(subsampling processes to generate a set of data driven features). Finally, it correlates those
features to class probabilities (e.g. classification decisions) within the final layer. Massive
numbers of samples and stochastic gradient descent are techniques required for efficient
learning of what features correlate to class probabilities in deep learning, including CNNs.
Using an already trained CNN, one can take advantage of the initial layers, which have already
learned to extract visual features such as edges and textures, and limit training for the particular
case at hand to its task-specific needs. One advantage of that is reducing training time; another
is environmental. Typically, hardware acceleration in the form of graphical processing units
(GPUs) are utilized, and GPUs require significant amounts of power while running the
backpropagation neural network training process over many iterations (Strubell, Ganesh, and
Mccallum 2019). We therefore used an already trained CNN, MobileNet v2 (Howard et al.
2017), designed for efficient real time prediction within mobile phones.
Specifically, we carry out an indirect form of transfer learning by using the MobileNet object
classification likelihoods as a feature vector passed to the next stage of Authente-Kente,
principal component analysis. The feature vector is the output layer of the MobileNet CNN and
has 1280 * 7 * 7 = 62,627 feature vector variables that take on particular values (e.g. predicted
object likelihoods) for a given 224 x 224 pixel subsection presented at the input layer. We
discuss how the feature vector is reduced in the following section.
Dimensionality Reduction
Given that we only have 2,000 kente swatches the feature vector size exceeds the number
samples and presents an ill-posed statistical learning problem (e.g. p >> n) that requires a
reduction of the feature vector size. We apply the standard practice of principal component
analysis (PCA), a statistical technique that orthogonally transforms a larger matrix into a smaller
matrix of linearly uncorrelated components called principal components. In our case we have
1,500 samples across the training and validation dataset (see Experimental Data Evaluation).
We opt to reduce the feature vector to a standard
principal components that together
explain 77% of the total variance. The component eigenvalues are all larger than 1 so we retain
them as is standard. The PCA decomposition explains a large percentage of the input feature
vector variance while reducing the input matrix to dimensions that present a well posed
statistical problem.
Logistic Regression
Logistic regression is a common statistical model for binary outcomes and is used to model the
decision of whether a subsection is from fake or authentic kente cloth (e.g. -1 for fake kente, 1
for authentic kente). In our formulation we use the PCA decomposition of the MobileNet feature
vector as independent variables from which to learn a logistic regression model. Regression is
particularly applicable in an applied setting because it is computationally efficient to evaluate.
Regression models are learned by statistically fitting model coefficients through an optimization
process that uses a loss function to numerically express the difference between predicted
outcomes and actual outcomes (e.g. error). Domain specific loss functions often better
generalize fitted model coefficients by controlling for known sources of error in the problem
Authentication error can be characterized by the cost and type of misclassification. A type of
kente may be misclassified as a false positive or a false negative, with related statistical
measures sensitivity and specificity. In a random marketplace the prevalence of fraud is
unknown and so we assume an equal cost of misclassification to avoid bias. To control for equal
misclassification cost and type we use the F1-macro loss function to induce a logistic regression
model that equally minimizes and balances error. Here we assume that authentic and fake
kente are equally likely and their misclassification equally costly.
Experimental Data Evaluation
Experimental evaluation in machine learning situations is complicated by the difficulty of
estimating the standard error of a model. The standard approach is to use cross validation to
estimate appropriate model parameters and performance, using a training dataset for training
the machine learning model and using a separate dataset, the validation dataset, to validate
chosen model parameters. A third, separate dataset of unseen whole cloths, the evaluation
dataset, is used to estimate the out of sample model error (Kohavi 1995). To maximize external
and internal validity we separate the data into training (750 cases), validation (750 cases), and
evaluation (500 cases) datasets. The training and validation swatch datasets are generated
from the same whole cloth images. The evaluation swatch dataset is generated from a different
set of whole cloth images; no swatch in the evaluation dataset has been seen by the proposed
model derived from the training and validation datasets. Source code for generating our dataset
from scratch and running the machine learning experiment
is openly available.
To claim significance we evaluate our pipeline along a gamut of standard binary classification
metrics: precision, recall, macro-F1 and weighted-F1 and accuracy (Tharwat 2018). We use
precision, recall, macro-F1 and weighted-F1 to support our claims of significance and set
purposively low thresholds a priori as given in Table 2 since the problem of visual kente
authentication has not been addressed in the literature. We answer the experimental question of
whether it is possible to significantly and efficiently decide if a swatch belongs to an authentic or
fake kente whole cloth. This is done through a computer science experiment involving training,
validation and evaluation datasets, as detailed in Table 3. Computational efficiency is supported
by using a pretrained MobileNet (Howard et al. 2017) model, a CNN specifically designed for
use on cell phones, and logistic regression as our decision algorithm, which reduces to matrix
multiplication with coefficients. Several efficient algorithms for PCA are commonly available.
A Priori Metric
A Priori Threshold
Per Class Precision
Per Class Recall
Macro F1
Weighted F1
Table 2: Threshold criteria for establishing significance
Unseen Cases
Shared Cases
Experimental Step
Fit PCA on Training
1 Run PCA(375 dim)
PCA(Eval) → EvalPCA
PCA(Val) → ValPCA
PCA(Train) → TrainPCA
2 Transform datasets
CrossValidation(TrainPCA, LR) →
3 Train, cross-validate
Logistic Regression
LRtrain(ValPCA) → PredVal
4 Predict against
validation swatches
Significance(PredVal) > A Priori Thresholds
5 Examine significance
metrics, proceed
CrossValidation(ValPCA+TrainPCA, LR) → LRtrain+val
6 Train, cross-validate
Logistic Regression on
all shared cases
LRtrain+val(EvalPCA) → PredEval
7 Predict against
evaluation swatches
(never before seen
Significance(PredEval) > A Priori Thresholds
8 Examine significance
metrics, report any
results exceeding
Table 3: Initial Authente-Kente Experimental Steps
Following the experimental setup given in Table 3, a logistic regression model was first cross
validated against the training dataset and results were observed to exceed the a priori
threshold. Then those parameters were used to re-train the same logistic regression model on
the combined training and validation dataset. The re-trained model was then evaluated on the
evaluation dataset, as indicated in step 7 of Table 3 and the resulting predictions scored as
indicated in step 8 with all metrics exceeding the a priori threshold. A summary of results is
given in Table 4. The training and validation datasets had 750 cases each and the evaluation
dataset had 500 cases. The number of true cases per class is indicated in the support column.
There are 250 fake and authentic cases within the evaluation dataset.
Evaluation Dataset
F1 Score
F1 Score
Macro Average
Weighted Average
Table 4: Summary of Authente-Kente machine learning pipeline experimental results
* all metrics exceed a priori threshold values
Relative to the a priori metrics given in Table 2 the Authente-Kente machine learning pipeline
exceeds the thresholds by large margins. Noticeably, a large majority (recall of 88%) of
authentic swatch cases are identified and when authentic cases are identified they are
accurately predicted to be authentic (precision of 88%). This balance between recall and
precision is reflected in both macro and weighted averages of 88%. It is important to note that
the precision and recall scores are the same. Per the Logistic Regression section, the loss
function induces a learned model trained to equally minimize across error types. Precision and
recall can be the same when the number of swatches that were actually authentic, false
negatives, are the same as the number of swatches that are actually fake, false positives, as
observed in Table 4. This is because precision and recall only differ in their denominator, using
false positives and false negatives, respectively.
Discussion & Future Work
These results against a large number of unseen kente cloth swatches suggest that our
implementation significantly captures part of the visual distinction between authentic and fake
kente cloth. Although the results are significant, in future work the application needs to be
connected to its larger social-technical context. One important social context is the local
prevalence of fake kente where the application is in use. As an example, imagine 500 whole
kente with a 5% counterfeit rate. In this case there are 25 fake whole cloths. In this low
prevalence example, with a recall of 0.88 Authente-Kente would incorrectly identify 60 authentic
whole kente as false. If the user received the authentication decision “likely fake'' there is only a
29% chance that it was accurate in this example. An application that was connected to local
artisans and regional data could continue to train as a way to improve accuracy, or selectively
account for prevalence to improve overall decision reliability. By design the application makes
no assumption and assumes the prevalence rate is equal. This can be addressed in future work
with the mobile application that contains the machine learning pipeline.
The components of this machine learning pipeline are computationally efficient and we expect a
real world solution to run in real-time. The nearly 1/3 improvement over the a priori threshold
provides an indication of experimental significance to the user and the possibilities for becoming
one element in a larger platform connecting buyers and artisans. Another element for such
platforms might be in STEM education, “whiteboxing” the system to provide an example for how
AI can be of social benefit. Our prior work in Ghana shows statistically significant improvement
in STEM lessons anchored in this kind of merger between computing and culture (Babbitt et al
2015; for access to the culture-based learning software see
A third element might be kente pattern identification. There is a wealth of symbolic meanings,
histories, aphorisms and other details woven into the cloths, but easily forgotten (if they were
communicated at all). Adding an AI application connecting, for example, a video clip of an elder
explaining the meaning of a particular symbol, pattern or style would give such a platform value-
added significance. Further extensions could be obtained by ensuring the sustainability of raw
material sources. We have explored this in our work on Ghanaian adinkra textiles
(, and it could be further enhanced using GIS,
IoT and other systems.
A final element for consideration is enabling the sharing and collection of video recordings of
textiles being created by artisans. The ability to connect to someone who creates the kinds of
kente cloth and patterns just purchased creates a "value added" that benefits both consumer
and producer. In future work, our envisioned system will enable partners in Ghana to directly
collect images of kente cloth and patterns. Generally, AI and associated ICT could be utilized as
a means to create an ecosystem of services that enhance unalienated labor, sustainable
feedstocks and the expressive values of cultures past and future, as we have outlined
elsewhere (Eglash et al. 2019).
Babbitt, W., Lachney, M., Bulley, E., & Eglash, R. (2015). Adinkra mathematics: A study of
ethnocomputing in Ghana. Multidisciplinary Journal of Educational Research, 5(2), 110–135.
Boateng, D. Ghana loses over GHC 5.7 billion to fake and counterfeit goods. Panavest, May 20
2019. Accessed online 2/23/2020 from
Coccia, M. (2017). Global Negative Effects of the Technological Change on Human Health: The
High Incidence of Cancers in the Anthropocene.
Curry, G. N., & Koczberski, G. (2012). Relational economies, social embeddedness and valuing
labour in agrarian change: An example from the developing world. Geographical Research,
50(4), 377-392.
Das, K. (2009). Detecting patterns of anomalies (No. CMU-ML-09-101). CMU Pittsburgh PA
Machine Learning Department.
Edusei, J., & Amoah, P. A. (2014). Cultural Tourism and Employment Creation Nexus: Evidence
from Kente Weaving and Wood Carving Industries in Kwabre East District, Ghana. Journal of
Economics and Sustainable Development, 5(8).
Eglash, R., Robert, L., Bennett, A., Robinson, K. P., Lachney, M., & Babbitt, W. (2019).
Automation for the artisanal economy: enhancing the economic and environmental sustainability
of crafting professions with human–machine collaboration. AI & SOCIETY, 1-15.
Hodge, Victoria J., and Jim Austin. 2004. “A Survey of Outlier Detection Methodologies.”
Artificial Intelligence Review 22 (2): 85126.
Howard, Andrew G., Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang,
Tobias Weyand, Marco Andreetto, and Hartwig Adam. 2017. “MobileNets: Efficient
Convolutional Neural Networks for Mobile Vision Applications.” ArXiv:1704.04861 [Cs],
Howard, E., Aboagye, I., & Quarcoo, J. (2016). Causes and Effects of the Dwindled State of the
Ghana Textile Industry. International Journal of Advanced Scientific Research & Development,
3(4), pp 1-12.
Kohavi, Ron. 1995. “A Study of Cross-Validation and Bootstrap for Accuracy Estimation
and Model Selection,” August, 7.
Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E Hinton. 2012. “ImageNet Classification
with Deep Convolutional Neural Networks.” In Advances in Neural Information
Processing Systems 25, edited by F. Pereira, C. J. C. Burges, L. Bottou, and K. Q.
Weinberger, 1097–1105. Curran Associates, Inc.
LeCun, Bengio, and Hinton. 2015. “Deep Learning.” Nature 521 (May).
Melissa G. Hunt, Rachel Marx, Courtney Lipson, and Jordyn Young (2018). No More FOMO:
Limiting Social Media Decreases Loneliness and Depression. Journal of Social and Clinical
Psychology: Vol. 37, No. 10, pp. 751-768.
M’Closkey, K. (2010). Novica, Navajo Knock-Offs, and the ’Net. In Moberg, M. (Ed), Fair trade
and social justice: Global ethnographies. NYU Press.
Mehra, P. (2019). The rise of the machines: Kashmir's weavers fear for their future. Al Jeerza,
July 19. Accessed 2/22/2020 from
Michelsen, H., & Bildt, C. (2003). Psychosocial conditions on and off the job and
psychological ill health: depressive symptoms, impaired psychological wellbeing, heavy
consumption of alcohol. Occupational and environmental medicine, 60(7), 489-496.
Oxford Business Group (2019). Ghanaian tourism grows with increased visitors and new
funding strategies.
Parvaiz, A. (2017). How a nanotechnology based microchip can help you identify real Pashmina
Strubell, Ganesh, and Mccallum. 2019. “Energy and Policy Considerations for Deep
Learning in NLP.” ACL 2019, June.
Tharwat, Alaa. 2018. “Classification Assessment Methods.” Applied Computing and
Informatics, August.
Yadav, Samir S., and Shivajirao M. Jadhav. 2019. “Deep Convolutional Neural Network
Based Medical Image Classification for Disease Diagnosis.” Journal of Big Data 6 (1):
Special Thanks
The authors would like to thank Connor Esterwood for referring the author to valuable
iconography as well as Chris Kerr for the Noun Project mobile phone icon in Figure 2.
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
Abstract Medical image classification plays an essential role in clinical treatment and teaching tasks. However, the traditional method has reached its ceiling on performance. Moreover, by using them, much time and effort need to be spent on extracting and selecting classification features. The deep neural network is an emerging machine learning method that has proven its potential for different classification tasks. Notably, the convolutional neural network dominates with the best results on varying image classification tasks. However, medical image datasets are hard to collect because it needs a lot of professional expertise to label them. Therefore, this paper researches how to apply the convolutional neural network (CNN) based algorithm on a chest X-ray dataset to classify pneumonia. Three techniques are evaluated through experiments. These are linear support vector machine classifier with local rotation and orientation free features, transfer learning on two convolutional neural network models: Visual Geometry Group i.e., VGG16 and InceptionV3, and a capsule network training from scratch. Data augmentation is a data preprocessing method applied to all three methods. The results of the experiments show that data augmentation generally is an effective way for all three algorithms to improve performance. Also, Transfer learning is a more useful classification method on a small dataset compared to a support vector machine with oriented fast and rotated binary (ORB) robust independent elementary features and capsule network. In transfer learning, retraining specific features on a new target dataset is essential to improve performance. And, the second important factor is a proper network complexity that matches the scale of the dataset.
Full-text available
Ghana's once booming Textile industry has seen a downturn in recent times with multiplicity of adverse effects on the nation at large. Major research findings of the Ghana Textile Industry reveal poor operational and output performance with less than 4,000 workforce, average production output of 46 million yards per annum and foreign exchange and revenue generation of 3.173 million US dollars across the country. Whiles some stakeholders blame the government for lack of goodwill and policy direction for the industry, others argue that the industry operates on obsolete machinery and technology. This research identifies and examines the causes and effects of the dwindled state of the Ghana textile industry and makes projections toward revitalization of the textiles sub-sector. It targeted key players of the textile sub-sector from industry and academia. Purposively, Textiles Ghana Limited (TGL), Akosombo Textiles Limited (ATL), KNUST and Takoradi Polytechnic were selected for the study. However, stratified sampling technique was used in selecting members from the parent population for data collection through interview and questionnaire. The study combines qualitative and quantitative research approaches through Ex Post Facto analysis to arrive at the results. Stiff competition and trade liberalization were found as the major causes of the dwindled state of the industry. The rippling effects includes; high rate of redundancy, upsurge in graduate unemployment, low exports and high imports of textiles, decline in textiles education, etc. Restructuring, recapitalization, retooling and policy direction are critical projections recommended for the industry to build competitive edge against its offshore counterpart to survive.
Full-text available
Introduction: Given the breadth of correlational research linking social media use to worse well-being, we undertook an experimental study to investigate the potential causal role that social media plays in this relationship. Method: After a week of baseline monitoring, 143 undergraduates at the University of Pennsylvania were randomly assigned to either limit Facebook, Instagram and Snapchat use to 10 minutes, per platform, per day, or to use social media as usual for three weeks. Results: The limited use group showed significant reductions in loneliness and depression over three weeks compared to the control group. Both groups showed significant decreases in anxiety and fear of missing out over baseline, suggesting a benefit of increased self-monitoring. Discussion: Our findings strongly suggest that limiting social media use to approximately 30 minutes per day may lead to significant improvement in well-being.
Full-text available
Classification techniques have been applied to many applications in various fields of sciences. There are several ways of evaluating classification algorithms. The analysis of such metrics and its significance must be interpreted correctly for evaluating different learning algorithms. Most of these measures are scalar metrics and some of them are graphical methods. This paper introduces a detailed overview of the classification assessment measures with the aim of providing the basics of these measures and to show how it works to serve as a comprehensive source for researchers who are interested in this field. This overview starts by highlighting the definition of the confusion matrix in binary and multi-class classification problems. Many classification measures are also explained in details, and the influence of balanced and imbalanced data on each metric is presented. An illustrative example is introduced to show (1) how to calculate these measures in binary and multi-class classification problems, and (2) the robustness of some measures against balanced and imbalanced data. Moreover, some graphical measures such as Receiver operating characteristics (ROC), Precision-Recall, and Detection error trade-off (DET) curves are presented with details. Additionally, in a step-by-step approach, different numerical examples are demonstrated to explain the preprocessing steps of plotting ROC, PR, and DET curves.
Full-text available
We present a class of efficient models called MobileNets for mobile and embedded vision applications. MobileNets are based on a streamlined architecture that uses depth-wise separable convolutions to build light weight deep neural networks. We introduce two simple global hyper-parameters that efficiently trade off between latency and accuracy. These hyper-parameters allow the model builder to choose the right sized model for their application based on the constraints of the problem. We present extensive experiments on resource and accuracy tradeoffs and show strong performance compared to other popular models on ImageNet classification. We then demonstrate the effectiveness of MobileNets across a wide range of applications and use cases including object detection, finegrain classification, face attributes and large scale geo-localization.
Full-text available
div> This paper details the development and evaluation of software that allows middle school students to explore the mathematical aspects of Ghanaian Adinkra symbols. We tested the effectiveness of this simulation in a Ghanaian junior high school by conducting a randomized quasi-experiment. We begin this paper by framing culturally responsive math education within the interventionist tradition of ethnomathematics. We draw this tradition together with an empirical exploration of the mathematics embedded in Adinkra symbols. We follow this with a methodological explanation for how we translated the mathematical significance of Adinkra into the design of our software, “Culturally Situated Design Tools.” Finally, we describe the quasi-experimental evaluation of the software using a randomized assignment of students in control and intervention groups in Ghana. We found statistically significant improvement for students using the culture-based software in comparison to similar software with no cultural content. </div
Outlier detection has been used for centuries to detect and, where appropriate, remove anomalous observations from data. Outliers arise due to mechanical faults, changes in system behaviour, fraudulent behaviour, human error, instrument error or simply through natural deviations in populations. Their detection can identify system faults and fraud before they escalate with potentially catastrophic consequences. It can identify errors and remove their contaminating effect on the data set and as such to purify the data for processing. The original outlier detection methods were arbitrary but now, principled and systematic techniques are used, drawn from the full gamut of Computer Science and Statistics. In this paper, we introduce a survey of contemporary techniques for outlier detection. We identify their respective motivations and distinguish their advantages and disadvantages in a comparative review.
"This outstanding collection not only serves as an accessible introduction to Fair Trade but illuminates the gap between the sunny rhetoric and the actual practice. With ethnographic richness and nuance, the authors complicate our understanding of the market as a means of achieving economic and social justice."