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

Abstract and Figures

We present the first unsupervised deep learning method for pollen analysis using bright-field microscopy. Using a modest dataset of 650 images of pollen grains collected from honey, we achieve family level identification of pollen. We embed images of pollen grains into a low-dimensional latent space and compare Euclidean and Riemannian metrics on these spaces for clustering. We propose this system for automated analysis of pollen and other microscopic biological structures which have only small or unlabelled datasets available.
Content may be subject to copyright.
arXiv:1908.01866v1 [cs.CV] 5 Aug 2019
Unsupervised Representations of Pollen in Bright-Field Microscopy
Peter He * 1 Gerard Glowacki 2Alexis Gkantiragas * 3
We present the first unsupervised deep learning
method for pollen analysis using bright-field mi-
croscopy. Using a modest dataset of 650 images
of pollen grains collected from honey, we achieve
family level identification of pollen. We embed
images of pollen grains into a low-dimensional
latent space and compare Euclidean and Rie-
mannian metrics on these spaces for clustering.
We propose this system for automated analy-
sis of pollen and other microscopic biological
structures which have only small or unlabelled
datasets available.
1. Introduction
Pollen is formed as the male gametes of all flowering
plants. Palynology, the study of pollen (and spores), pro-
vides critical insights for a number of fields including foren-
sic science, ecology and agriculture (Mildenhall,2006;
Blackmore,2006;Zbrodsk & Vorlov,2014). For example,
pollen grains on the personal affects of an individual can
be used to ascertain whether they were at the scene of a
crime (Mildenhall,2006). However, traditional methods of
pollen analysis such as microscopic analysis of morphol-
ogy are time intensive and require trained specialists. As a
result, pollen analysis remains mostly inaccessible for the
majority of large scale applications. Making the analysis
of pollen fast, scalable and accessible can therefore open
the door to a great many opportunities hitherto unfeasible
in both commercial and academic domains.
While automated methods of pollen analysis have been de-
scribed in the literature as early as 2002 (Ronneberger et al.,
2002) the methods have not been implemented as signif-
icant tools in research or industry. Early attempts were
*Equal contribution 1Department of Computing, Imperial Col-
lege London 2Hampton Court House, London, United Kingdom
3Department of Molecular Biology, University College London.
Correspondence to: Peter He <>,
Alexis Gkantiragas <>. With
thanks to Stewart McGown, University of St Andrews.
Workshop on Computational Biology at the International Confer-
ence on Machine Learning, Long Beach, CA, 2019. Copyright
2019 by the author(s).
largely scanning electron microscopy (Ronneberger et al.,
2002;Treloar et al.,2004) while later research used a
wide variety of microscopy types including dark field
(Lagerstrom et al.,2014;Pedersen et al.,2017), and fluo-
rescence microscopy (Mitsumoto et al.,2009). All of these
techniques require expensive equipment and most require
relatively skilled operators.
The introduction of deep learning techniques for pollen
analysis has shown much promise, however, have been
plagued by a lack of a significant labelled dataset. In a re-
cent publication, a supervised deep learning algorithm was
used to segment and then analyse pollen in honey samples
(He et al.,2018). However, the scope of this paper was lim-
ited by its requirement for the manual labelling of datasets.
Here we provide the first example of the use of an unsu-
pervised deep learning algorithm for the analysis of pollen
using bright-field microscopy.
2. Methods
Unlike previous almost-entirely supervised approaches to
pollen identification (which require large labelled datasets
of pollen to train), we propose a new unsupervised pipeline
(see Figure 1) for generating representations of pollen and
learning groupings of morphologically (and thererby often
taxonomically (Oswald et al.,2011)) similar pollen grains.
Figure 1.
Overview of pollen identification. (A) Image of slide.
(B) Downsampled image. (C) Object detection network. (D)
Bounding boxes. (E) Full resolution pollen grain crops. (F) En-
coder. (G) Decoder. (H) Further dimensionality reduction. (I)
Cluster assignment.
Unsupervised Representations of Pollen in Bright-Field Microscopy
Given an input image of a microscope slide containing
pollen, XR3328×3328×3
+, we first apply downsampling
such that XR416×416×3
+. The downsampled Xis then
passed through a YOLO-based (Redmon et al.,2016) ob-
ject detection network from (He et al.,2018) to obtain a set
of bounding boxes BR4
+×Nfor pollen grains present on
the slide containing indexed tuples (xi, yi, wi, hi)giving
the positions and dimensions of the ith particular bound-
ing box respectively. Each tuple in Bis then scaled up
component-wise by a factor of 8 and recombined with the
original image to obtain centered full-resolution crops ciof
the pollen. Each cipassed through an encoder based on the
VGG16 architecture from (Simonyan & Zisserman,2014)
pre-trained on ImageNet (Deng et al.,2009) with the final
fully-connected layers replaced with a 7×7max-pooling
layer resulting in a latent representation ziZR512 .
We then further reduce the dimensionality of zito df inal
Nand cluster the results using the k-means algorithm. In
our experiments, we compare principal component analysis
to the Isomap algorithm as dimensionality reduction meth-
ods as well as a Euclidean metric on Zto a Riemannian
metric induced by the latent space. The Riemannian met-
ric is approximated using a fixed-point algorithm described
in (Yang et al.,2018) and reflects the curvature that arises
from the non-linearity of the latent space which can be
characterised as a Riemannian manifold (Arvanitidis et al.,
In this way, we are able to generate representations of the
pollen without requiring a large dataset of pollen to learn
from. Furthermore, the clusters discovered can now be
used for a number of new tasks including semi-supervised
learning for pollen classification.
3. Experiments
Samples from a range of honeys (eucalpytus melliodora,
acacia and manuka) were spread onto a microscope slide
and immediately imaged. Images were taken once every
2 seconds on a Solomark compound microscope at 320x
zoom. This generated a dataset of 650 unlabelled bright-
field images of individual pollen.
3.1. Clustering
The system was run with parameters k= 10 and df inal =
3for ease of visualisation. The value of kwas selected
based on the ratio between decreases in cluster variance for
adjacent values of k. The system was in most cases able
to differentiate pollen morphology on (at the minimum) a
family level. For example, one cluster (see A in Figure 2)
shows a strong resemblance to pollen from the Myrtaceae
family (Sniderman et al.,2018).
There are marked differences in the latent spaces gener-
Figure 2.
Randomly sampled members from different clusters of
pollen imagery with PCA and a Euclidean metric on Z.
ated by the two different dimensionality reduction meth-
ods. Qualitatively, clusters are far more clear in the spaces
where Isomap was used (see Figure 4). Nonetheless, the
embeddings in the PCA spaces were sensible (see Figure
3). The clusters were not too different in terms of overall
concepts - the labels in Figure 5are matched to potential
”conceptual counterparts” in Figure 2. There were, how-
ever, qualitatively fewer obviously incorrect cluster assign-
ments from the Isomap-Riemannian space compared to the
other spaces.
We further visualise the curvature of the latent spaces by
sampling geodesics between random points on each man-
ifold. The geodesics observed on the PCA manifolds are
(unsurprisingly) effectively linear while some more inter-
esting curvature can be observed on the Isomap manifold
in Figure 6.
3.2. Comparison with Human-Defined Cluster
A set of 30 bright-field images were collected at 320x zoom
using a combination of an AmScope bifocal microscope
and a Solomark compound microscope. Non-specialist vol-
Unsupervised Representations of Pollen in Bright-Field Microscopy
Figure 3.
Samples along a principal component of Zseemingly
encoding colour, surface texture and roundness.
Figure 4.
Latent space representations of Isomap (left) and PCA
(right). Cluster assignments are illustrated with lines from each
image point in the latent space to its respective centroid. Clusters
are computed using a Euclidean metric on Zin both visualisations.
Point size corresponds to depth.
unteers were instructed to categorise a random selection
of these images into the already generated clusters given
4 randomly chosen images from each cluster (similar to
Figure 2). The images were then reclassified by the sys-
tem (using PCA and an Euclidean metric on Z). Cases of
agreement between the human and system was 63%. Calcu-
lating Cohen’s Kappa coefficient, we find moderate agree-
ment κ= 0.576 (95% CI, 0.378 to 0.775).
4. Discussion
While previous research had used unsupervised methods
to analyse pollen samples in grass (Mander et al.,2013)
we used a sample size which was significantly larger and
did so with less expensive tools, making our method more
accessible. Additionally, we were able to utilize a signifi-
cantly larger dataset and observe an entire pollen grain, as
opposed to simply surface patterns. Moreover, agreement
between humans and our system was relatively high at 63%.
This implies that the clusters and assignments at large are
to some extent human-interperatable. An important future
benchmark would be comparison to human specialists who,
with the help of reference material, are able to visually dif-
ferentiate pollen at a species level with reasonable accuracy
though such resolution is not necessary for many applica-
tions (Mildenhall,2006).
While our system provides a new tool for working with
small or unlabelled datasets, it has significant limitations.
Figure 5.
Randomly sampled members from different clusters of
pollen imagery with Isomap and a Riemannian metric on Z.
Primarily, the small sample size and poor quality of some
images in our dataset limits the classification of pollen to
the family level. A larger dataset and better equipment
may be required to achieve genus and species level clas-
sification. Luckily, this should be relatively straightfor-
ward as no labels are needed. Alternatively, applying the
same system used here on datasets generated using other
microscopy types such as dark field or phase contrast mi-
croscopy could provide powerful new tools for rapid pollen
analysis, though this would potentially not confer the same
benefit with regards to accessibility.
Although the latent spaces observed in the experiments
were relatively sensible, often with the obvious clusters vis-
ible, it is likely that the embeddings are sub-optimal due to
the pre-training of the encoder on ImageNet. This is due to
the fact that biological imagery at the micro-scale is visu-
ally very different to much of the contents of ImageNet. It
would therefore perhaps be beneficial to representation to
use an encoder trained instead on a large dataset of micro-
scope imagery.
Unsupervised Representations of Pollen in Bright-Field Microscopy
Figure 6.
Visualising geodesics between random points on Zwith
Isomap and a Riemannian metric. The red lines are geodesics
while the blue dots are image embeddings from the dataset. They
are clearly non-linear over larger distances.
4.1. Applications
Beyond accelerating existing uses of pollen analysis, two
socially beneficial applications of large-scale automated
pollen analysis include combatting the proliferation of
fraudulent honey and as well as ecological monitoring.
Fraudulent honey is a global issue and has seen honey be-
come the third most faked food in worldwide (Moore et al.,
2012). Honey fraud is most commonly carried out through
mislabelling, dilution with cheaper honeys or sugar syrup
and a host of unethical beekeeping practices such as over-
harvesting. Existing authentication methods for honey have
been inaccurate (such as sugar or chemical testing) or
highly specialist and expensive (such as qPCR, LC-MS or
manual inspection of the pollen in honey (Mcdonald et al.,
2018;Kato et al.,2014;Sniderman et al.,2018)). For this
use case, we describe a system whereby producers of honey
may upload scans of their honey taken through low-cost
bright-field microscopy apparatus to an online database.
With such a database, consumers and retailers down the
supply chain will be able to verify the origin or authentic-
ity of their honey from its pollen profile without special-
ist knowledge. This will help combat the harmful effects
of fake honey on both the ecosystem and local economies
described elsewhere (F. Fairchild et al.,2000;Cairns et al.,
2005). Furthermore, the system will have made up the
cost of a microscope purchase within 3-4 jars of genuine
manuka honey.
In addition, we propose honey as a means for the large-
scale monitoring of flora biodiversity. Previous stud-
ies have used honey as a way of monitoring pollutants
(Smith et al.,2019) and conducting analysis on pollen
would allow for simultaneous generation of data on flora
biodiversity. This would go hand-in-hand with the honey
authentication system described above, using the pollen
profile database as a source of spatio-temporally tagged
samples. Two caveats, however, ought not be overlooked:
firstly, that this application would require a small quantity
of expertly labelled pollen data in order to associate pollen
grains with their respective botanical sources in a semi-
supervised fashion; and secondly, that honey bees have spe-
cific foraging habits which would need to be corrected for.
Nevertheless, pollen obtained from other sources such as
dedicated collection units can be equally analysed by the
system. We present a potential architecture for how both
proposed applications can co-exist in Figure 7.
Figure 7.
A potential architecture for ecosystem monitoring
through large-scale honey authentication infrastructure. The
database can not only be used for honey authentication, but also
for long-term analytics on changing pollen distributions.
Finally, the system could be used for the analysis of any un-
labelled microscopy dataset in the environmental and life
sciences. Soil fungi, for example, are considered to be an
excellent indicator of soil fertility (Kranabetter et al.,2009).
The use of the system for such applications could acceler-
ate the development of prototypes and proofs-of-concept
by removing the need for a large labelled dataset.
5. Conclusions
Despite its limitations, the system we have described forms
the groundwork for a powerful, scalable and accessible tool
for pollen analysis. The acquisition of a larger dataset and
an application-specific encoder are needed to bring the sys-
tem to a level of maturity which, once reached, will open
doors to novel techniques for ecological monitoring, honey
authentication and a variety of other use cases.
6. Acknowledgements
Many thanks to Stewart McGown at the University of St
Andrews for assisting in the design and prototyping of the
large-scale honey authentication system. We also acknowl-
edge Raghvi Arya for her advice regarding the poster pre-
sentation at ICML.
Unsupervised Representations of Pollen in Bright-Field Microscopy
Arvanitidis, G., Hansen, L. K., and Hauberg, S. Latent
space oddity: on the curvature of deep generative models.
In ICLR, 2018.
Blackmore, S. Pollen and spores: Microscopic keys to
understanding the earths biodiversity. Plant Systemat-
ics and Evolution, 263(1-2):312, 2006. doi: 10.1007/
Cairns, C. E., Villanueva-Gutierrez, R., Koptur, S., and
Bray, D. B. Bee populations, forest disturbance, and
africanization in mexico1. Biotropica, 37(4):686692,
2005. doi: 10.1111/j.1744- 7429.2005.00087.x.
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and
Fei-Fei, L. Imagenet: A large-scale hierarchical image
database. 2009 IEEE Conference on Computer Vision
and Pattern Recognition, 2009. doi: 10.1109/cvpr.2009.
F. Fairchild, G., Capps, O., , O., and Nichols, J. Im-
pacts of economic adulteration on the u.s. honey industry.
Western Agricultural Economics Association, 2000 An-
nual Meeting, June 29-July 1, 2000, Vancouver, British
Columbia, 01 2000.
He, P., Gkantiragas, A., and Glowacki, G. Honey au-
thentication with machine learning augmented bright-
field microscopy., Dec 2018. URL
Kato, Y., Fujinaka, R., Ishisaka, A., Nitta, Y., Ki-
tamoto, N., and Takimoto, Y. Plausible au-
thentication of manuka honey and related prod-
ucts by measuring leptosperin with methyl syringate.
Journal of Agricultural and Food Chemistry, 62
(27):6400–6407, 2014. doi: 10.1021/jf501475h.
PMID: 24941263.
Kranabetter, J. M., Friesen, J., Gamiet, S., and Kroeger,
P. Epigeous fruiting bodies of ectomycorrhizal fungi as
indicators of soil fertility and associated nitrogen status
of boreal forests. Mycorrhiza, 19(8):535548, 2009. doi:
10.1007/s00572- 009-0255-0.
Lagerstrom, R., Holt, K., Arzhaeva, Y., Bischof, L.,
Haberle, S., Hopf, F., and Lovell, D. Pollen image clas-
sification using the classifynder system: Algorithm com-
parison and a case study on new zealand honey. Sig-
nal and Image Analysis for Biomedical and Life Sci-
ences Advances in Experimental Medicine and Biology,
pp. 207226, 2014. doi: 10.1007/978-3- 319-10984-8 12.
Mander, L., Li, M., Mio, W., Fowlkes, C. C., and Pun-
yasena, S. W. Classification of grass pollen through
the quantitative analysis of surface ornamentation and
texture. Proceedings of the Royal Society B: Biologi-
cal Sciences, 280(1770):2013190520131905, 2013. doi:
Mcdonald, C. M., Keeling, S. E., Brewer, M. J., and Hath-
away, S. C. Using chemical and dna marker analy-
sis to authenticate a high-value food, manuka honey.
npj Science of Food, 2(1), 2018. doi: 10.1038/
Mildenhall, D. Hypericum pollen determines the presence
of burglars at the scene of a crime: An example of foren-
sic palynology. Forensic Science International, 163(3):
231235, 2006. doi: 10.1016/j.forsciint.2005.11.028.
Mitsumoto, K., Yabusaki, K., and Aoyagi, H. Classification
of pollen species using autofluorescence image analysis.
Journal of Bioscience and Bioengineering, 107(1):9094,
2009. doi: 10.1016/j.jbiosc.2008.10.001.
Moore, J. C., Spink, J., and Lipp, M. Development and
application of a database of food ingredient fraud and
economically motivated adulteration from 1980 to 2010.
Journal of Food Science, 77(4), 2012. doi: 10.1111/j.
Oswald, W. W., Doughty, E. D., Neeman, G., Neeman,
R., and Ellison, A. M. Pollen morphology and its
relationship to taxonomy of the genus sarracenia (sar-
raceniaceae). Rhodora, 113(955):235251, 2011. doi:
10.3119/0035- 4902-113.955.235.
Pedersen, B., Bailey, D. G., Hodgson, R. M., Holt, K., and
Marsland, S. Model and feature selection for the classi-
fication of dark field pollen images using the classifyn-
der system. 2017 International Conference on Image
and Vision Computing New Zealand (IVCNZ), 2017. doi:
Redmon, J., Divvala, S., Girshick, R., and Farhadi, A. You
only look once: Unified, real-time object detection. In
The IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), June 2016.
Ronneberger, O., Schultz, E., and Burkhardt, H. Auto-
mated pollen recognition using 3d volume images from
fluorescence microscopy. Aerobiologia, 18(2):107115,
Jun 2002. doi: 10.1023/a:1020623724584.
Simonyan, K. and Zisserman, A. Very deep convolu-
tional networks for large-scale image recognition. arXiv
1409.1556, 09 2014.
Smith, K. E., Weis, D., Amini, M., Shiel, A. E., Lai, V. W.-
M., and Gordon, K. Honey as a biomonitor for a chang-
ing world. Nature Sustainability, 2(3):223232, 2019. doi:
10.1038/s41893- 019-0243-0.
Unsupervised Representations of Pollen in Bright-Field Microscopy
Sniderman, J. M. K., Matley, K. A., Haberle, S. G., and
Cantrill, D. J. Pollen analysis of australian honey. Plos
One, 13(5), 2018. doi: 10.1371/journal.pone.0197545.
Treloar, W. J., Taylor, G. E., and Flenley, J. R. Towards
automation of palynology 1: analysis of pollen shape
and ornamentation using simple geometric measures, de-
rived from scanning electron microscope images. Jour-
nal of Quaternary Science, 19(8):745754, 2004. doi:
Yang, T., Arvanitidis, G., Fu, D., Li, X., and Hauberg, S.
Geodesic clustering in deep generative models. CoRR,
abs/1809.04747, 2018.
Zbrodsk, B. and Vorlov, L. Adulteration of honey and avail-
able methods for detection a review. Acta Veterinaria
Brno, 83(10), 2014. doi: 10.2754/avb201483s10s85.
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
Urban geochemistry is an emerging field in which key scientific and societal challenges, including rapid urbanization and population growth, compel investigation of readily accessible biomonitors to determine the source, transport and fate of heavy metal pollutants in cities. Lead isotopic analyses of honey have recently proven its efficacy as a biomonitor for Pb source apportionment applications. We collected honey directly from hives in six geographical sectors in Vancouver, British Columbia (Canada) to investigate the presence of potential pollutants from varying zoning districts: urban, industrial, residential and agricultural. Systematic variations in trace element concentrations and Pb isotopic compositions of the honeys reflect proximity to anthropogenic land-use activities such as shipping ports and heavy traffic. Honey sampled from downtown hives, near the Port of Vancouver, shows elevated trace element concentrations compared with suburban and rural honey, and distinctly higher 208Pb/206Pb (that is, less radiogenic) compared with local environmental proxies (for example, oysters, Fraser River sediment and volcanic rocks), indicating possible input from Asian anthropogenic sources. This study presents the first Pb isotope data for North American honey, and supports the combined use of trace elements and Pb isotopic compositions in honey as a geochemical biomonitor.
Full-text available
Ensuring the authenticity of food is a rapidly emerging issue, especially in regard to high-value products that are marketed through increasingly complex global food chains. With the ever-increasing potential for mislabeling, fraud and adulteration, governments are increasingly having to invest in, and assure, the authenticity of foods in international trade. This is particularly the case for manuka honey, an iconic New Zealand food product. We show how the authenticity of a specific type of honey can be determined using a combination of chemicals derived from nectar and DNA derived from pollen. We employ an inter-disciplinary approach to evaluate a selection of authenticity markers, followed by classification modelling to produce criteria that consistently identify manuka honey from New Zealand. The outcome of our work provides robust identification criteria that can be applied in a regulatory setting to authenticate a high-value natural food. Our approach can transfer to other foods where assurance of authenticity must take into account a high level of natural variability. High value speciality food products are threatened by mislabeling, fraud, and adulteration, which calls for practical authentication measures. Research led by the Ministry for Primary Industries in New Zealand was used to develop an analytical strategy that combines four chemicals and a DNA marker to authenticate mānuka honey, a speciality export from New Zealand. They found that a combination of chemicals from the mānuka plants nectar and species-specific DNA from mānuka pollen could be used as authenticity markers, when tested against a variety of plant and honey samples. Based on quantitative measurements of the markers, they used a classification modelling method to establish the identification criteria for both monofloral and multifloral mānuka honey. Combining multiple marker data and statistical classification analysis may also help develop other food authentication measures.
Full-text available
Pollen analysis is widely used to verify the geographic origin of honeys, but has never been employed in Australia. In this study, we analysed the pollen content of 173 unblended honey samples sourced from most of the commercial honey producing regions in southern Australia. Southern Australian vegetation is dominated by Eucalyptus (Myrtaceae) forests and, as expected, most Australian honeys are palynologically dominated by Eucalyptus, while other important components include Myrtaceae taxa such as Corymbia/Angophora and the tribe Leptospermeae; plus Brassicaceae, Echium, Macadamia, and Acacia. An important feature of the honeys is the number of Myrtaceae pollen morphotypes per sample, which is generally high (mean = 4.6) compared to honeys produced outside of Australia, including Eucalyptus honeys produced in the Mediterranean region, and honeys produced in South America, which has its own rich indigenous Myrtaceae flora. In the latter regions, the number of Myrtaceae morphotypes is apparently generally ≤2. A high number of Myrtaceae morphotypes may be a feasible criterion for authenticating the origin of Australian honeys, since most Australian honey is produced by honey bees mainly working indigenous floral resources. Myrtaceae morphotype diversity is a convenient melissopalynological measure that could be applied even where detailed knowledge of the pollen morphology of the many component genera and species is absent. Palynological criteria developed in Europe for authenticating Eucalyptus honeys should not be relied upon for Australian honeys, since those criteria are not based on samples of Australian honey.
Full-text available
Taxonomic identification of pollen and spores uses inherently qualitative descriptions of morphology. Consequently, identifications are restricted to categories that can be reliably classified by multiple analysts, resulting in the coarse taxonomic resolution of the pollen and spore record. Grass pollen represents an archetypal example; it is not routinely identified below family level. To address this issue, we developed quantitative morphometric methods to characterize surface ornamentation and classify grass pollen grains. This produces a means of quantifying morphological features that are traditionally described qualitatively. We used scanning electron microscopy to image 240 specimens of pollen from 12 species within the grass family (Poaceae). We classified these species by developing algorithmic features that quantify the size and density of sculptural elements on the pollen surface, and measure the complexity of the ornamentation they form. These features yielded a classification accuracy of 77.5%. In comparison, a texture descriptor based on modelling the statistical distribution of brightness values in image patches yielded a classification accuracy of 85.8%, and seven human subjects achieved accuracies between 68.33 and 81.67%. The algorithmic features we developed directly relate to biologically meaningful features of grass pollen morphology, and could facilitate direct interpretation of unsupervised classification results from fossil material.
Deep generative models provide a systematic way to learn nonlinear data distributions, through a set of latent variables and a nonlinear "generator" function that maps latent points into the input space. The nonlinearity of the generator imply that the latent space gives a distorted view of the input space. Under mild conditions, we show that this distortion can be characterized by a stochastic Riemannian metric, and demonstrate that distances and interpolants are significantly improved under this metric. This in turn improves probability distributions, sampling algorithms and clustering in the latent space. Our geometric analysis further reveals that current generators provide poor variance estimates and we propose a new generator architecture with vastly improved variance estimates. Results are demonstrated on convolutional and fully connected variational autoencoders, but the formalism easily generalize to other deep generative models.
Technical Report
In this work we investigate the effect of the convolutional network depth on its accuracy in the large-scale image recognition setting. Our main contribution is a thorough evaluation of networks of increasing depth, which shows that a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16-19 weight layers. These findings were the basis of our ImageNet Challenge 2014 submission, where our team secured the first and the second places in the localisation and classification tracks respectively.
Manuka honey, obtained from Leptospermum scoparium or L. polygalifolium flowers in New Zealand and a part of Australia, has strong antibacterial properties. In this study, plausible authentication of the manuka honey was inspected by measuring leptosperin, methyl syringate 4-O-β-D-gentiobiose, along with methyl syringate. Despite a gradual decrease in methyl syringate content over 30 days at 50°C, even at moderate 37°C, leptosperin remained stable. A considerable correlation between nonperoxide antibacterial activity and leptosperin content was observed in 20 certified manuka honey samples. Leptosperin and methyl syringate in manuka honey and related products were analyzed using HPLC connected with mass spectrometry. One noncertified brand displayed significant variations in the leptosperin and methyl syringate contents between two samples obtained from different regions. Therefore, certification is clearly required to protect consumers from disguised and/or low-quality honey. Because leptosperin is stable during storage and specific to manuka honey, its measurement may be applicable for manuka honey authentication.