Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Revista Brasileira de Computação Aplicada, November, 2023
DOI: 10.5335/rbca.v15i3.15028
Vol. 15, No3, pp. 80–95
Homepage: seer.upf.br/index.php/rbca/index
ORIGINAL PAPER
SnakeFace: a transfer learning based app for snake classification
Jorge Guerra Pires ,1and Luiz Henrique Dias Braga ,2
1Founder at IdeaCoding Lab / JovemPesquisador.com, Brazil, 2Laboratório de Zoologia dos Vertebrados, Departamento
de Biodiversidade, Evolução e Meio Ambiente/ICEB, Universidade Federal de Ouro Preto, Campus Morro do Cruzeiro, s/n,
Bauxita, Ouro Preto 35400-000, MG, Brazil
*jorgeguerrapires@yahoo.com.br; luizbraga.bio@gmail.com
Received: 2023-07-06. Revised: 2023-10-30. Accepted: 2023-11-16.
Abstract
Introduction: deep learning emerged in 2012 as one of the most important machine learning technologies, reducing
image identification error from 25% to 5%. This article has two goals: 1) to demonstrate to the general public the ease of
building state-of-the-art machine learning models without coding expertise; 2) to present a basic model adaptable to
any biological image identification, such as species identification. Method: We present three test-of-conceptmodels that
showcase distinct perspectives of the app. The models aim at separating images into classes such as genus, species, and
subspecies, and the input images can be easily adapted for different cases. We have applied deep learning and transfer
learning using Teachable Machine. Results: Our basic models demonstrate high accuracy in identifying different species
based on images, highlighting the potential for this method to be applied in biology. Discussions: the presented models
showcase the ease of using machine learning nowadays for image identification. Furthermore, the adaptability of
this method to various species and genuses emphasizes its importance in the biological fields, as root for inspiring
collaborations with computer science. On our case, future collaborations could lead to increasingly accurate and efficient
models in this arena using well-curated datasets.
Keywords: Biology; bioinformatics; deep learning; Snakes; tensorflow; JavaScript; transfer learning
Resumo
Introdução: O aprendizado profundo surgiu em 2012 como uma das tecnologias mais importantes de aprendizado de
máquina, reduzindo o erro de identificação de imagens de 25% para 5%. Este artigo tem dois objetivos: 1) demonstrar
ao público em geral a facilidade de construir modelos de aprendizado de máquina de última geração sem expertise
em programação; 2) apresentar um modelo básico adaptável a qualquer identificação de imagens biológicas, como
identificação de espécies. Método: Apresentamos três modelos conceituais de teste que mostram perspectivas distintas
do aplicativo. Os modelos visam separar imagens em classes como gênero, espécie e subespécie, e as imagens de entrada
podem ser facilmente adaptadas para diferentes casos. Aplicamos aprendizado profundo e transferência de aprendizado
usando o TeachableMachine. Resultados: Nossos modelos básicos demonstram alta precisão na identificação de diferentes
espécies com base em imagens, destacando o potencial deste método para ser aplicado em biologia. Discussões: os
modelos apresentados mostram a facilidade de usar aprendizado de máquina atualmente para identificação de imagens.
Além disso, a adaptabilidade desse método a várias espécies e gêneros enfatiza sua importância nas áreas biológicas,
como base para colaborações inspiradoras com ciência da computação. No nosso caso, colaborações futuras poderiam
levar a modelos cada vez mais precisos e eficientes nessa área usando conjuntos de dados bem selecionados.
Palavras-Chave: Aprendizado profundo; bioinformática; biologia; JavaScript; Serpentes; TensorFlow; transferência de
aprendizado.
Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95 81
1 Introduction
Around 2012, deep learning came to the surface, and
became one of the most important machine learning
technologies. The image identification error went from
about 25% to just about 5%. Those values can be higher, or
at least close to human error rates on image classification.
This revolution also went to other areas. Google Translator,
chatGPT, and other tools, they all use deep learning.
China announced heavy investments in deep learning
researches, after they were beaten on their own game Go
using deep learning (FRONTLINE,2020).
Different from human intelligence, machine learning
is stored and easily transferable. When Einstein died, so
died his genius. When we train a model, you can either
give access to the model (e.g. chatGPT) to the public, called
pretrained models, or actually borrow part of the model,
called transfer learning. These approaches make machine
learning easily accessible to anyone, from anywhere. One
of the selling points of chatGPT is making large language
models accessible, models that not even organizations
could afford on their own. Said the CEO of openAI (Altman,
2023): the cost of intelligence will drop drasticallyin the
upcoming years.
Nowadays, we have a common consensus that AI should
be accessible to anyone; not just experts, not just big
techs. Generally, they are released as APIs (Application
Programming Interface), or public libraries. In a close
past, those models were just for experts, and were very
hard to work with, very focused on all aspects, including
programming languages. Good models were generally
locked up on highly-cost proprietary software. With APIs,
applications can talk to each and exchange information
and abilities, or even, different AIs, and share easilywhat
they have learnt, what they do best. chatGPT has been
constantly integrated with other tools using APIs, even,
other Artificial Intelligence (AIs) models. It makes those
models exponentially more powerful, and more accessible.
1.1 Our goal
This article aims at two goals, atits core: 1) showing to the
general public howstraightforward it is nowadays to build
machine learning models with state of the art techniques,
no code required, no expertise required; 2) present a basic
model that can be adapted to any similar case in biology
(i.e., species identification based on images, as long as
the identification can be done from images, the model
should work). Also, we want to draw attention to possible
collaborations.
We also propose a big model, which we hope will be
able to support people on identifying snakes. Our mindset
is similar to chatGPT developers: launch it as soon as
possible, and get feedback, even if negative. Our hope
is that we could grow together with the users, instead of
following the traditional path of just launching when the
app is perfect. We have launched a simple demo, a test
of concept, and share our findings on this article. As we
are going to show, the app misclassifies when the snakes
are similar, and we have a couple of thoughts on how
to improve the predictions. To be fair, where the model
misclassifies, non-experts also may misclassify (e.g., false
coral snakes vs. true ones).
1.2 Basic concepts
Our main bottleneck is a well-curated dataset of snakes
with species and subspecies, we are using Google
Image which can introduce errors of classification from
erroneous annotations. Even though the machine learning
training process is set generally to ignore noise on dataset
(i.e., mistagged snakes), we would like to avoid those
noises when possible. On machine learning community,
we call this process of associating classes to images,
annotation, the process of annotating images for teaching
an algorithm to do it on its own.
Annotation can be a very big problem since it may take
time, and time from experts. That is where the quotation
from the CEO from openAI Sam Altman comes handy,
once those models are trained, they can perform the work
of experts, at zero or low costs; chatGPT costs one cent
of dollar per 1.000 words, no human can beat that. Our
dataset was partially curated by a biologist, but enlarged
with images from Google Images. Since we are using
transfer learning, about 30 images for each class was
enough; image models are trained generally on millions
of images for being good enough.
A nice move we did was using transfer learning: it
makes the annotation less heavy since we need a small set
of samples for each class. A machine learning mayrequire
millions of images for actually achieving interesting
results. MobileNetV1, as an example, was trained on more
than 1 million images.
Transfer learning was an ingenious way found by
computer scientists to solve two problems: i) making it
less heavy to train those models; ii) making it available
already trained models, but, with learning abilities.
On this approach, we have a base model, and a trainable
model. The base model is trained heavily, and made
available online for users, which use this base model as
feature model (see Section 2.7). A trainable model will use
those features for training a local model. Our feature model
has about 1.200 features, which is sampled from images.
Models made available on TensorFlow Hub can have more
than 1600 features, which is a measure on how many
details they can learn from images. TensorFlow Hub is
a collection of pretrained models made available for users,
no charge, feature models are one type of models made
available for transfer learning on this Google platform.
Their goal is making transfer learning accessible, and
easy to use, especially on the browser (i.e., JavaScript).
TensorFlow.js is also part of this pursuit of making
machine learning widely available, it was created by a
group at Google, the same creators of TensorFlow in
python (they are similar, and exchangeable).
As one example on how this tool can be powerful,
we have Pires (2018) reported an example where
researchers applied a general-trained based model for
teeth segmentation, making it possible to achieve
excellent results from a small number ofimages. We are
using the same approach, except they had to build the
model from scratch, requiring a couple of PhD students
and masters in computer science as so they could make
the transfer learning. On its basics, transfer is not
82 Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95
straightforward to implement.
We are going to use TensorFlow.js (Rivera,2020;
Shanqing Cai et al.,2020;Laborde,2021). TensorFlow.js
was developed by Google, based on TensorFlow in
python, also by Google. Their goal was making machine
learning more accessible, also for web developers.
Moreover, taking advantage from the most used
programming environment: the browser; and the most
used programming language, JavaScript (Stack Overflow,
2023). The nice point about JavaScript is that most
modern libraries and frameworks for web applications
are somehow rooted in JavaScript, making it easy any
extension. Our app was developed in Angular, also by
Google, coded in Typescript from Microsoft, a typed
version of JavaScript. Angular can be useful for scientific
computation, as it was pointed out here (Pires et al.,2021).
1.2.1
On the importance of having an app for identifying
snakes even by non-experts
In 2022, 27515 accidents with snakes were notified
in Brazil, with death rate of about 0,40% (Sistema de
Informação de Agravos e de Notificações – SINAN, 2023).
Some numbers given by the organization: about 69% are
attributed to Bothrops genus (simulations showed a good
identification of this genus by our model since they are
very easy to define on images); 1,19% to Micrurus (our
model is not very good at separating from the fake ones;
the goal should be avoiding misclassifying the true one
as false coral snake); 16,6% to another snake genus and
13,4% non-identified snakes.
According to the healthcare agency from Paraná
(Brazil) (Secretaria de Saúde,2022), in Brazil are notified
annually about 20.0000 accidents with snakes, with
death rate of about 0,43%. Some numbers given by the
organization: 1) about 75% are attributed to bothrops
genus (simulations showed a good identification of this
genus by our model since they are very easy to define on
images); 2) Micrurus (0,5% our model is not very good at
separating from the fake ones; the goal should be avoiding
misclassifying the true one as false coral snake). Moreover,
16,6% to another snake genus and 13,4% non-identified
snakes.
In this fact, services to help on identification of snakes
is important and could minimize risks of accidents or
misidentification during the treatment, since some non-
venomous snakes could be morphologicallysimilar with
venomous snakes.
The snake genus Micrurus, commonly known as
coral snake, is a well-known snake by the Brazilian
population for its color pattern: mostly of the species
in the genus presents red, black, and white (or yellow)
rings (Roze,1996). This type of color pattern is
not by chance: its biological evolution for advertise
their dangerous called as “warning coloration” (i.e.,
aposematism, Poulton (1890)). In contrast, certain non-
harmful species (mimics) imitates the visual, acoustic,
and chemical characteristics of toxic species (models).
This phenomenon is called Batesian mimicry (Ruxton
et al. (2004); Hossie and Sherratt (2013); Davis Rabosky
et al. (2016)). Within this context, species belonging
to the Micrurus genus are often associated with being
mimic models for non-venomous or "semi-venomous"
snakes, commonly known as false coral snakes (Wickler,
1968;Greene and McDiarmid,1981;Campbell and Lamar,
1989;Savage and Slowinski,1992). Some of these species
mimic various Micrurus species, showcasing different
color patterns in according to their distribution and of
the mimic model (e.g., Atracus sp., Almeida et al. (2022)).
1.2.2 A brief look at the literature
Even though we were the first to explore transfer learning
for snake classification using JavaScript, the application
of machine learning to snake classification, that we
explore herein, has already been explored also by other
researchers. A quick search on the internet, focusing on
technical works, we can find the following researches.
This topic has the potential to support on avoiding
accidents with snakes, and even protecting the non-
poisonous snakes.
Rajabizadeh and Rezghi (2021) did a work very similar
to ours, and compared several algorithms; they have
used as based model MobileNetV2. They also compared
different methods: one of them is using Principal
Component Analysis (PCA) for supporting the learning
process. The feature model is a kind of PCA, when one
reduces an image to a number of features, it is a sort of
PCA. Their best results was in the same direction as ours:
transfer learning based.
Kalinathan et al. (2021) was able to classify 772 classes
of snake species using the architecture ResNeXt50-V2,
which is an impressive result. In our case, we do not know
our architecture since this is not available on Teachable
Machine, but we know it has about 1200 features by
analyzing their pretrained model locally. Laborde (2021)
suggests that the final model is "a transfer learning
version of MobileNet, PoseNet, or some other practical
model that fits your needs."
Ahmed et al. (2023) tested several base models on snake
classification, and MobieNetwas one ofthe top algorithms,
the one we are using as base model for transfer learning.
Yang and Sinnott (2021) created a model for 11 snake
breeds from Australia, and also made it available as iOS
application; for our case, we are betting on a Reactive Web
Design (RWD), no need to keep two or more codes, the app
supposes to chance accordingly. Their lower number on
classes seems to be due to updated in machine learning,
much has changed in four years since the publication.
Durso et al. (2021) created a model for 45 species of
snakes, they tried several models.
Santos et al. (2021) did something very similar to us.
Their app Snake Classifier aims to help identifyvenomous
and non-venomous snakes, using techniques ofcomputer
vision and machine learning. Their app uses a model
of convolutional neural network trained with a dataset
of 1,200 images of 12 species of Brazilian snakes. They
have also used transfer learning. Different from us, they
have used python, which makes it harder to create web
application, our main selling point. They have used
MobileNetV2, which is essentially whatwe have used.
1.2.3 Apps for biology
With recent technological advancements, several
mechanisms have been employed to link technology
with biodiversity. Many applications and programs have
Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95 83
utilized connections with local communities to bridge
gaps in scientific knowledge related to biology, such as:
recognition of sounds emitted by animals (e. g. FrogID,
BirdNET, Merlin Bird ID; Rowley et al. (2023), Kahl et al.
(2021)), birds identification (e. g. Merlin Bird ID), plant
identification (e. g. PlantNet; Garcin et al. (2021)). In
addition, some resources are produced directly for the
scientific community, although there are instructions
accessible to non-specialists (e. g. online multi-access
interactive identification key; Pezzuti et al. (2021)).
Rowley et al. (2023) built a system for frog
identification, which includes noise they make to
support the identification. Merlin is an app for identifying
birds, similar to ours. We have gathered more information
on Merlin here since it is an interesting app. Basically,
they are using the same approaches as we are: deep
learning + transfer learning. INaturalist has on their
website the possibility to try to identify the species on the
photo: for Brazilian snakes, it does not seem to do a good
job.
1.3 Article content
On this article, we do a test of concept, and our results
are promising for a big model, a model that could from
images, identify a snake. chatGPT has been incorporated
into tools (Brockman,2023) as so it can be used as an
optimized way for teaching and learning, and we envision
a possible integration. Given an image identified, we
could use chatGPT as an addition aiding tool for teaching,
gathering information, and making sure our model got it
right.
We are going to consider three models. Since we are
presenting a test of concept, those three models aim at
showing different perspectives from the app. We believe
that this is generic enough to be used in anysimilar case:
one wants to separate images into classes (e.g., genus,
species and subspecies). One just needs to change the
input images and associate the respective class with each
annotated image.
The first model classifies just false coral snakes, and it
was the best results, achieving almost 100% of accuracy,
with zero misclassification; and it is easy to converge, it
does not require many attempts to get a good classification.
However, when trying to classify with true coral snakes,
the model starts to have problems (model 2) with very
similar snakes, but still having accuracy higher than 50%
(a measure of randomness). For the last model, model
3, it was added different species to see how the model
handles diversity, for a possible big model (see Section 3);
the true coral snakes were removed for avoiding the issue
we already saw on model 2.
Finally, we present on the final section of the discussion
section how we see a possible big model based on those
findings. We are strongly convinced that a model for
snake classification is not just possible, but also useful
for the general public, and maybe for experts depending
on the final accuracy. Apart from possible setbacks we
may face on the way and we are optimistic we can solve
them gradually, we strongly believe that building this app
is imperative, and could contribute to the interception
between biology and computer science (bioinformatics).
2 Methods
On this section, we present details about the training,
and methodologies we have used. We have also used the
opportunity to explain some basic concepts when needed.
2.1 Dataset
"Garbage in, Garbage out"
We have used images from Google Images: given the
species/subspecies, we looked for images on Google Image.
We had a couple of well-curated images for starting the
model, but we have gradually added those extra images.
Neural models, when well-trained, tend to ignore noise.
Thus, even if some images may be mistagged on Google
Images, it seems to be ignored since the training was
a success. We hope to close a possible collaboration for
retraining the model on a better dataset, curated by a
biologist. However, what we got so far is promising.
We should also bring to attention that we already have
a pretrained model for INaturalist available at TensorFlow
Hub, one ofthe biggest datasets on species on the internet.
This means that someone else has already trained a large
model focused on biology, made available as feature model.
It can be accessed here for transfer learning.
2.2 Training the model
We started with a small set of images, well-curated by
a student of biology. Then, we considered the graphs,
especially the matrix of accuracy per class, and added
new images from Google Images when needed, and
retrained the model. Even though we have changed the
defaults parameters from Teachable Machine, the default
parameters were enough. We have used extensively a
feature in which they allowyou to disable a class, which
was useful to figure out which class was making the
learning process unstable, and which classes were well-
classified. In most of the cases, 0.001 as learning rate and
50 epochs were enough.
2.3
Neural networks, Deep learning and transfer
learning
Neural networks are a subset of machine learning
techniques (Fig. 1), focused on numeric algorithms,
different from alternatives from artificial intelligence,
those algorithms are not focused on reasoning; even
though non-experts using those models may think they
reason. Moreover, even though they are inspired by the
brain, they do not replicate the dynamics of neurons with
fidelity. Their only goal is learning from samples, and it
does not matter how.
We are going to use the supervised variety: one presents
a set of inputs, and expected outputs, and the algorithm
should learn, without human interference, except at
annotating the dataset.
On this type of algorithms, they should learn on their
own, we do not interfere with the learning process.
Deep learning is a neural network with too many hidden
layers; that is what ’deep’ stands for, in opposition to
84 Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95
Figure 1: Neural networks are ruler finders. Source: Pires
(2022a)
shallow, which now is used to designate classical neural
models. They are designed to execute certain tasks, now
called narrow neural networks, in opposition to artificial
general intelligence.
Those models revolutionized the machine learning
community, and are everywhere: from Google translator,
chatGPT, DALL-E, to image classification. The drawback
is that they are heavy to train, and require too many
images (millions). The solution for making this technique
widely accessible was transfer learning. In simple terms,
you share your learning, and people can adapt to their
small cases.
2.4 Training and validation
It is a common practice in the machine learning
community to validate the models splitting the dataset
into learning and validation: the ratio will appear on the
figure for accuracy per class (e.g., Fig. 10). In all the
scenarios we tested, the training curve converged.
When we say about ’instability’ of the model, we are
mentioning the validation curve. The validation curve is
used to avoid memorization, overfitting. A model that
memorizes is useless since it will have to predict outside
their training dataset. Since our images are mainly from
Google Image, when testing the model, we tried to make
sure the testing was done with images not used neither
on the training dataset nor on the validation dataset; if
you input an image, say by accident, already present on
the training/testing dataset, the model will give almost
100% of certainty on the prediction. The testing is just a
confidence building strategy, it does not change the model.
A couple of samples is here, for the general model.
Bear in mind that we are paying for a server on Heroku,
those links may change in time. The reader is invited to
get in touch by e-mail for the latest links and resources.
2.5 Teachable Machine
Teachable Machine is a Google AI Experiment, as they like
to call it. It is built on top of TensorFlow.js. What you do
with this platform can be built on TensorFlow.js, except
that with this platform, no code needed. It does all the
training ritual, from uploading image, preprocessing to
showing training details. We have explored extensively
this platform herein. In fact, we have taken advantage
of that and we have actually done tests directly with
TensorFlow.js in order to compare the results: the
results of Teachable machine is competitive with plain
TensorFlow.js.
One possible usage of TensorFlow.js directly is actually
testing on different feature models, with different feature
numbers as output. You may see this feature number as a
granulation from the model. The more features you have
on the base model, the more details it is likely to find;
see that we cannot control which features the model will
extract, dictate where to look on the image. In essence,
the feature model reduces the image into a smaller set of
details, the features; see that other authors, as brought
to attention on the introduction, actually used PCA, but
the result was still better with feature models. Keep in
mind that except for some final touch, the base model is
not trained on the small dataset, it comes already trained.
For using Teachable Machine, one just needs to upload
the images, and ask to train; it takes seconds, therefore,
you can experiment with different parameters, or even
retrain the model if it does not converge. The model
is using local search, thus, it may get trapped on local
minima. There are not too much to setup. We have kept
the default parameters.
The one change we made sometimes was he training
rate. The thumb rule is: small training rate will make
the training slow, careful; whereas big values will make
it faster. Think link this: when you want the model to
be careful, try out small learning rates; when you believe
the model can learn fast, just increase it. In most of the
cases, the default parameters were the best. You may
need to experiment with different training rates for either
replicating our results, or adapting to your case.
2.6 TensorFlow.js
TensorFlow.js is a JavaScript-based library for deep
learning, based on the classical TensorFlow, written in
Python; you can also do simple learning machine, some
simple mathematical operations with tensors and so on.
There are several reasons for using TensorFlow.js
instead of Python, an imperative reason is using justone
language, from the app to machine learning. Another
reason is that you do the calculations on the browser, no
need to have high-performance servers. If your app starts
to gain users, the calculation cost will not grow since each
user is responsible for their calculation load. This last
point is especially interesting for startups, since you can
scale up without also increasing the cost. One reason for
medical applications: your data never leaves the browser,
it is ideal for sensitive data.
A nice point is that they claim it is possible to
transform models in both directions: TensorFlow.js <
– > TensorFlow (Rivera,2020;Shanqing Cai et al.,2020;
Laborde,2021). It is possible to transform models in both
directions. Even the manual transformation is possible
since their notations are similar.
TensorFlow.js provides several ways to be used:
pretrained models, CDN calls, NPM, and even
downloading the models. Each way has their own
advantages and disadvantages, which we shall not cover
here.
Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95 85
2.7 Feature model
Feature model (Fig. 2) is a deep learning trained model,
but just the hidden layers: it is unable to make sense of
an image since its classification layer was removed. It is
made available as so others can use this knowledge to train
small models, like ours: we just place a classification layer,
and train the model on our datasets (i.e., snakes).
Figure 2: Scheme on how a model based on transfer
learning works.
The biggest motivation for making those models
available is that not everyone has access to high-
performace computers. For our case, using machine
learning in the browser, we have a limitation on how
much we can take from a person’s computer. See that
TensorFlow.js also runs in the backend, using Node.js,
which requires GPUs; in fact, since JavaScript is also
present at desktop, you can also use on desktop apps.
Another motivation is that having millions of images,
even say 100 per class may be unreal. Transfer learning
was created to tackle those limitations on actually using
machine learning on images.
2.8 External resources
We cannot make sure those resource will be available when
you read this article, but here goes the current public
resource for learning from this article. You may get in
touch by e-mail for latest resources and links. See that we
have a course teaching the basics of the paper.
Find here on GitHub how to use an already-trained
model exported from Teachable Machine using Angular
(TypeScript).
As an extra feature, teachable machine hosts your
model, and make it available as a link, that can be called
externally. It is ideal since we separate the model training
and updates from the app; we have provided our links
on the upcoming lines. Once you update the training,
the change will automatically be pushed to the model
being used on the link. If you prefer pure JavaScript, here
goes a gist on GitHub. If you need more options, which
includes Android, you should access the export option of
each model, after training.
Below the links for each model reported here. See
that we had to retrain the models, thus, you mayobtain
different results. For the rawmodels, please, get in touch
as so we can find a way to exchange the files.
i.
Fake coral snakes (model 1):
https://teachablemac
hine.withgoogle.com/models/W9_qIu14Y/
ii.
Fake vs. true coral snakes model (model 2):
https://
teachablemachine.withgoogle.com/models/9vw2M7LJw/
iii.
Model with variety (model 3):
https://teachablem
achine.withgoogle.com/models/Sc8mKQsS0/
Tip. if you click on the link, the model will open on
the browser and you can input images. The models of the
paper will not be updated, as so you can test the results
herein. The same links work if you are programmer, for
importing the model inside your code.
3 Results and Discussion
On this section, we shall focus on the results achieved by
running our model with a set of images. We have selected
different snake species. This selection was done without
any predefined preference, biased by the studies of the
authors. We have a preference for Brazilian snakes, the
field of studies of LHDB.
We cannot see how this selection would affect the
results, thus, it does not affect reproducibility (as long
as you use the same images, it likely to be reproducible).
We have used curated images alongside Google Image
samples. By ’same images’, we mean for the same classes:
machine learning is not tied up to specific inputs, when
properly developed.
3.1 model 1: just fake coral snakes
We have selected "randomly" fours snakes classified as
fake corals; numbers range from 50 to 80 different false
coral snakes existent worldwide. This variation has to
do with the observation that those classifications are not
scientific, any snake that looks like a coral and it is not can
be classified as false. Our goal is showing a scenario where
two species or more are stronglyalike, and how the model
would behavior.
Namely: 1) Apostolepis assimilis (8 images); 2)
Erythrolamprus aesculapii (18 images); 3) Oxyrhopus
rhombifer (24 images); 4) Lampropeltis triangulum
triangulum (LTT, 22 images).
We have used chatGPT to get a list as complete as
possible from which we have selected the species used,
and selected the correspondent images from Google Image;
the list is here. We have taken the care to avoid species
with too many variations between subspecies: that is why
we have Lampropeltis triangulum triangulum (subspecies)
instead of just Lampropeltis triangulum (species).
Initial simulations with a possible big model showed
that variations between subspecies can make the model
unstable, and misclassify, generally amongst similar
86 Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95
species; also, the validation curve does not converge, and
can even "blow up" when we place subspecies in the
same class for the model to learn. Those classifications
are constantly reviewed by the biology community, and
subspecies may become species; we are using what is
officially accepted at the time of writing this article.
In some applications, people are using chatGPT to
automatically call apps from text(i.e., it understands what
the user wants and given a set of apps, it calls those
apps to give a well-based response), the results have
been promising (Wolfram,2023;Brockman,2023). We
have tested this in other scenarios, and the results are
promising. As one example, chatGPT was able to suggest a
snake from a short-text based description, even though it
may get it wrong sometimes, and you may need to cross-
references with other sources, it is a good starting point.
INaturalist is also able to identify from images, but,
for Brazilian snakes, it seems to be bad at; sometimes it
gets it right, but giving several suggestions which can be
confusing.
Even the MobileNet (generallyused as base model for
transfer learning) can make an identification, but without
details regarding species (it tags it as king snake, see
Fig. 3); king snakes and corals are two different types
of snakes with distinct characteristics; king snake is
considered a false coral snake. Keep in mind that when
the base model gets it right, it is transferred to the final
model (i.e., the fact that MobileNet can spot a king snake
does not make our model meaningless, our goal is making
those models more accurate on their prediction, especially
for Brazilian species).
Figure 3: King Snake, a false coral
Another possibility is Google image, however, adding
to possible mistagging from users or even from Google
when making "its trick", it tends to associate image with
already-existent image; in fact, we had difficulties to find
species from Google image, using the image as input. The
user will most likely not take pictures from the snakes
close to their images: chances are high the image will not
exist, and chances are it will not necessarily be close to
any existent one on Google Image. As one example how
Google Image can get it totally wrong. We have entered
apapa lesma (Dipsas indica): it gave several suggestions,
some totally wrong, but did not got it right.
Our goal with transfer learning those big models is
actually making them more specific. Since we are using
Teachable Machine, we do not know the base model, it
is not available on their site. One alternative, which we
do not explore here, is actually making transfer learning
using TensorFlow.js manually, see Laborde (2021), chapter
11, for a guiding on how it is possible to use your
own base model of preference. Even though harder
than teachable machine, it may require programming
abilities in JavaScript, it is still much easier than classical
approaches using python.
Still on the base model, we have actually done a test
with INaturalist: there is a pretrained model available as
feature model that was trained on INaturalist. So far, the
only advantage we saw is that the final model converges
faster with INaturalist. Surprisingly, using INaturalist
as base model did not improve the predictions. We have
actually created the model outside Teachable Machine for
comparison using JavaScript. More investigations will be
done in the future for confirming this finding.
Generally, we have started to train with a small number
of images, and increased when needed. We have used
both the confusion matrix and the accuracy per class
(Fig. 10) to decide which class to increase the image
samples. What is interesting: in some of the simulations,
the model converged even for 3 images. This is possible
due to transfer learning, without transfer learning, a
model would never converge with just say 30 images,
unthinkable for 3 images. It also happens because they
are very distinct snakes: the closer two species, the more
images you may need.
We are going to present some results. The goal is
showing a potential of a platform, a concept validation.
We want to show that it is possible to build a model for
snake classification with current free technologies, and
possibly support on learning and even helping people to
avoid accident with snakes.
Innovating with biomathematics is all about making
our models easily available (Pires,2022b). Usually, the
scientific publications are the final goal on a research, but
we believe in a culture where applications/innovations,
should be the final goal. This is the startup culture.
Artificial intelligence, as APIs and libraries, is making
this a reality, with models each time more powerful, and
easier to use in generic applications.
Below the confusion matrix for model 1 (Fig. 4). This
matrix shows that all the species were gotten right: no
misclassification was done; when a misclassifcation is
done, you should see values out of the main diagonal.
Keep in mind it does not guarantee the model will not
make mistakes. If the dataset is biased, say to a subspecies
claiming to be a species, the model will behaviour like it
learnt, but it will fail when presented with subspecies not
represented on the training process.
Biased dataset is a big issue on machine learning: those
models cannot learn what have not been shown. If they
have never seen a species, it will try to guess amongstwhat
it has.
As an example (Fig. 5), the model 2 is misclassifying,
which has two true coral snakes, with 60% the micrurus
lemniscatus which does not exist on our dataset at
the moment as micrurus frontalis (true coral) and
36% as Erythrolamprus aesculapii (the winner on
misclassifications on the current "big model" since it is
quite similar to the true corals). See that we had to retrain
Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95 87
Figure 4: confusion matrix for model 1
the model to run this test. Always keep in mind that you
may obtain results slightly different from ours since those
models are initiated with random weights.
Figure 5: Micrurus lemniscatus, true coral not present at
the current training. Source: Wikimedia Commons
To make sure we are not memorizing the dataset, you
can check out Fig. 6, which shows that both learning and
validation curves converged to lower values; and best of
all, the validation curve is slightly pointing downward,
which means we could improve the model by increasing
the learning time. See that teachable machine allows to
stop the learning process, at any time: you may add an
big learning time, and stop when you see it is becoming
unstable, or converged.
Another useful graph is Fig. 7, which presents how
much the model got it right: training arrived to 100%,
and validation to 99%; it is called accuracy, you divide
the correct classifications by the total attempts. All those
curves help to avoid common traps when teaching those
models (e.g., overfitting).
“With four parameters I can fit an elephant, and with five I can make
him wiggle his trunk” Johnny von Neumann
As a last way to make sure the model is working, we can
Figure 6: learning curves for model 1. Legend: the blue
curve is training, whereas the orange one is
testing/validation.
Figure 7: Accuracy for model 1.
actually test it (Fig. 8); we have actually applied on a more
complex image, which also worked, a person holding it on
his/her hand, but due to possible copyright issues, we have
decided to show this simple case as illustration. The model
gave 80% of chance of being an oxyrhopus rhombifer, with
lower probabilities for the remaining classes; high values
for remaining classes even when the model gets it right
show possible confusion, chances of misclassifcation. It is
a good result for classification. see also Fig. 9. Remember
that the model will give probabilities for each class present
at the training, and we choose the highest one. We are
playing with clustering the probabilities by genus; and also
to use extra information such as geographical.
Still on Fig. 8, different from Merlin, which is focused
on birds, it is not necessary to zoom in: it is done
automatically by the algorithm. We have though about
it, however, if we decide to use zoom in, we are planning
to use machine learning to automatically zoom in. Unless
we have no choice, it is interesting to avoid human‘s
88 Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95
Figure 8: Oxyrhopus Rhombifer. Source: Wikimedia
Commons
Figure 9: Realistic photo for erythrolamprus aesculapii
(71% given by the model, and additionallywith low
probability to other similar classes. See that this image
was chosen for being hard to classify due too much
background and obstructions, being very realistic).
interference.
Last but not least, Fig. 10 shows the classification per
class. Since the confusion matrix was diagonal, we already
expected it.
3.2 model 2: false coral snakes vs. true ones
We have more than 50 true coral snakes, and even more
when it comes to the false ones; keep in mind that it is not
a scientific classification. The numbers are not a common
consensus; for the false ones, the subspecies may increase
Figure 10: acurray matrix per class for model 1
.
even more the final count.
On this model, we are using the following true coral
snakes: 1) Micrurus corallinus (24 images); 2) Micrurus
frontalis (25 images); 3) Micrurus alleni (23 images) ;
4) Micrurus albicinctus (14 images). For the false ones:
1) Apostolepis assimilis (8 images); 2) Erythrolamprus
aesculapii (18 images); 3) Oxyrhopus rhombifer (24 images).
This is a tough classification. Humans generally
cannot make the difference, unless, they are experts,
with specialized training. This paper was born because
JGP could spot a false coral, LHDB can due his technical
training in biology.
This model can be useful for helping people to separate
the fake coral snakes from the true ones. The fake ones
are generally not dangerous (i.e., venomous snakes). The
true coral snakes are well-known for its potent venom, it
can be fatal to humans.
Error type I and error type II are two different types
of errors in statistics: they are called false positive and
false negative. We are calling on this model error type I
as classifying a non-dangerous snake as dangerous; and,
similarly, error type II as classifying a dangerous as non-
dangerous. See that even though one can use our models
to separate venomous snakes from non-venomous ones,
different from Santos et al. (2021), this is not our primary
goal.
We assume we all agree that error type II should be
avoided, if we assume one key application is actually
helping people avoiding accident with those snakes. On
the current version of the algorithm we cannot make sure
we avoid one type of error, but we can see the final training,
and retrain until we get an expected result. For this case, it
may take several attempts, which takes seconds to train.
A possible enhancement of the model is considering
automatization of this best model finding. We see
promising results actually trying an evolutionary
computing. We are able to find even already available
repositories using TensorFlow.js with genetic algorithms,
something we need to study and run simulations.
Different from model 1, this model did not converge
easily. it took several attempts to get the results showed
here. The model has problems to classify similar snakes.
One example was Micrurus corallinus vs. Erythrolamprus
aesculapii (Fig. 11); erythrolamprus aesculapii is the
champion of fooling our model, and humans too (Fig. 11).
We saw similar problems with tests with a model not
Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95 89
Figure 11: Micrurus corallinus vs. Erythrolamprus aesculapii
reported here for just Bothrops genus: similar snakes tend
to induce misclassification between themselves (Fig. 14).
Figure 12: Confusion matrix for model 2
On Fig. 11, just take a look at the Erythrolamprus
aesculapii on horizontal line, and look upward (basically,
all the misclassifications come from Erythrolamprus
aesculapii being misplaced as true corals). Similarly, do
the same on the vertical, and also the misclassifications
come from Erythrolamprus aesculapii being misplaced as
true corals. The errors type II, to be avoided, are on the
lower part of the diagonal. For reading this matrix, on y-
axis (vertical), we have the true class; whereas, on x-axis
(horizontal), we have the model predictions (Fig. 12).
Figure 13: Accuracy for model 2
See that if you try to replicate those results (Fig. 13),
you may find something different, likely similar. Neural
models are initiated randomly, thus, when you start
training, chances are you will start on a different point
on the neural network space that we started. Just try a
couple of times, and the model should converge as we
report (Fig. 15). Fig. 16 shows accuracy per class. See that
most of the misclassifications come from erythrolamprus
aesculapii being misplaced as either micrurus frontalis or
micrurus micrurus corallinus (both dangerous). We are not
sure adding more image will help; even though we may
try in the future.
On the machine learning community, we knowthose
models are generally bad at what we humans do easily;
and tend to be better at what we do poorly. It seems
biologists make the difference on those cases the model
90 Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95
Figure 14: Bothrops alternatus misclassified as Bothrops neuwiedi. Source: courtesy of Luiz Braga
Figure 15: Training curves for model 2
is struggling with by taking a look at the tail, head and
belly. chatGPT is suffering from "common sense" (easy
answers for humans, and it makes a mess when trying
to answer)(Choi,2023). Thus, what we human created
"a common sense", the model cannot see. The model
used as base, for transfer learning, is basically frozen
on our small dataset. Therefore, we cannot be sure it
learnt to actually "look at tails" or head. How those
models learn is something we cannot explain, chatGPT has
been target of several nice discussions (Wolfram,2023).
Those discussions are not just helping us to popularize
AI, but also to question basic assumptions we have been
doing. chatGPT opened the gate for talking about artificial
Figure 16: Accuracy for model 2
general intelligence, smarter models.
3.3 model 3: experimenting with diversity
In the world, 4,038 different species of snakes are listed
Uetz et al. (2022). 435 of these species are registered for
Brazil, being the 3rd country in the ranking of the highest
number of snake species. For building such a model,
it would take time, patience, and lot of collaborations.
Our simulations shows that when the snakes are similar,
not necessarily from the same genus, they tend to be
misclassified amongst themselves.
Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95 91
We are not even sure it is possible to build such model,
at least, not with 100% of accuracy: mistakes are about to
be made on such a generic model. The best model we know
is able to identify something superior to 1.000 different
images. MobileNet, widely used in transfer learning, can
identify 1.000 different objects; even though, they can
make mistakes, or can "get in doubt" ("confused"). The
Inception model also can identify this number of objects
Laborde (2021). They differ on precision and speed of
response. See that those numbers may change since those
models are constantly enhanced. A solution would be
building small models, with species separated by rules
(e.g., fake coral vs. true ones). The selection of the model
to be used can be done by an algorithm, by the user, or
even by chatGPT; something we mustexperiment on.
A good result we can see on the literature, which we are
not sure it is easily replicable is Kalinathan et al. (2021):
they created a model with more than 700 different snakes.
Thus, it may be possible to build such a model, theyused
"old" approaches, we have better methods now.
For the fake ones: 1) Apostolepis assimilis (8 images);
2) Erythrolamprus aesculapii (18 images); 3) Oxyrhopus
rhombifer (24 images). And for the diversity, we have
used: 1) Crotalus durissus (12 images); 2) Bothrops
alternatus (24 images); 3) Bothrops neuwiedi (22 images);
4) Bothrops jararaca (23 images). Since we already know
the difficulties between false and true coral snakes, we left
the true coral snakes out.
Figure 17: Accuracy per class for model 3. Legend: 1 -
100%, no mistakes; 0 - 0%, got it all wrong.
As we can see from Fig. 17, the accuracy per class ranged
from 50% to 100%. Another interesting fact: a small
model tends to be good also on the big model (see that
no fake coral snake was misclassified), which something
we need to investigate. It would be good news if that
observation holds for all the classes.
The mistakes accounts for misclassifications on the
Bothrops genus (Fig. 19). We have a guess that part of this
misclassification may come from the fact that they have
subspecies, which we have verified, with wide variations
Figure 18: Accuracy for model 3
in patterns: we may need to train for subspecies.
Figure 19: Confusion matrix for model 3. Legend: the
numbers mean how may classifications are right, when
on the main diagonal, or misclassification, when outside
the main diagonal. The perfect matrix is diagonal.
The good news is: the model misclassify amongst the
same genus (Fig. 21). The solution we found for now: the
model tend to give high probabilities to similar snakes. We
just sum up the probabilities for genus, and add this as the
predicted genus. We call this probability clustering.
It may help for a potential second level where we
give the first prediction for a more refined model,
something we want to experiment on for improving the
predictions. This strategy could also increase the number
of species/subspecies we could spot.
As you could see, the models can misclassify (and a
potential big model, Fig. 22), which generally happens
when the snakes are similar. Erythrolamprus aesculapii
92 Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95
Figure 20: Training curves for model 3
was the winner, able to fool our model, passing as true
coral snakes (Fig. 12).
As we highlighted before, the dangerous is when the
venomous snake is classified as false coral snake, what we
have called error type II. The images show that they are
very similar in patterns (Fig. 11). A solution that came up
during our discussions is using geographic information.
The collaborator on this research is actually working as
part of his research on a probability distribution of snakes
based on geographic information. However, we are opened
for possible collaborations. We have done some tests
with INaturalist API, that provides this information once
we provide a location. It could be used to support on
predicting the snake: in addition to image identification,
also considering whether the snake was ever seen on the
user’s location.
As alternative, we may consider public APIs that provide
information and more. INaturalist has an API, which
provides information based on API calls. Providing extra
information, we could actually give our model a better
chance to get it right. Another alternative is actually
asking the user extra information, either before or after
the prediction; we are avoiding this type of strategy as
so the system will be as autonomous as possible. We
should mention that chatGPT (Fig. 22) showed promising
behaviors during this app development as an assistant,
which could be used also to enhance the predictions (e.g.,
chatGPT is able to guess, sometimes getting close, the
snake based on a short textual description).
3.4 Conclusion
We have presented a state-of-the-art machine learning
system for classifying snakes, best ofall, no coding during
the model development. It was possible thanks to a
public platform by Google called Teachable Machine. This
approach can be adapted to any classification problem,
as long as the difference between the species is on the
images. Our hope is that we do not just call attention to this
option for non-experts in machine learning, but also call
for collaborators, especially, with well-curated datasets
Figure 21: Misclassification of the species, but classified
properly the genus. This case was taken from our actually
deployed app
on snakes species.
It is hard to make a fair comparison with the other
authors cited herein. The reason is that those models
evolve very fast, therefore, a better result here may be
due to a better routine for image identification. One
strong point of our model is that it was implemented
basically codeless, using an online tool. It makes it easier
to replicate the results by non-experts in programming
(e.g., biologists), or even, for passing the work forward.
Also, the training is decoupled from the app, therefore, it
is possible work on them separately. It will certainly make
it easier to share the work load on possible collaborations.
See that we have an online course teaching the basics of
this paper for the general public. The best result we found
on the literature was able to classify +700 different species,
at least theoretically, we can arrive to 1.000, and with some
tricks, even more. One challenge is that a species may have
several subspecies, on some cases, those subspecies are
not necessarily alike in the image. Some species may have
too much variations, which make it harder to use species
as an way to enhance classification (i.e. clustering the
probabilities).
One insight we have mentioned on the paper is using
geographic information to support on the classification.
Thus, if a snake is classified as probable, we can see if
the snake was ever seen on the geographic location it was
found by the user. This information is part of the work
of LHDB. For creating an initial prototype, we can use
the iNaturalist API. It allows a call with location, and it
gives back several pieces of information, amongst this
Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95 93
Figure 22: Vision of a possible execution plan for SnakeFace based on the presented results
information, we have the frequency the species was seen
on a location. This information can be crossed with the
model prediction to improve the prediction. Another
application of INaturalist, a huge dataset on species, is
using their feature model, already available on TensorFlow
Hub. We did some initial testings and comparisons: we
could not find so far advantages compared to what we did
so far. We hope to make more testing.
In one of the papers mentioned, they have used PCA and
compared with an approach similar to ours, using transfer
learning: transfer learning won. It means that most likely
we are on the right track.
One possible drawback is that we are using JavaScript,
since python is the dominant language in machine
learning. It is not necessarily a problem since we can
convert the models between the languages, thus , if
you train in python, you can just convert the model to
JavaScript. Therefore, if a new model is released, and it is
in python, we can still make the testing. It is possible
that new models will be better at predicting, and will
have more objects (so far, the best is about 1.000 different
objects). Each species/subspecies is an object. See that we
have released our datasets on Kaggle, a python machine
learning community, so far, no one has solved the problem
as we did using python.
Our biggest challenge now is actually adding new
species, and keep training. It may take time, and at the
moment we do not have funding. As we train, we need to
study the curves, and make sure the model is optimized
for a certain goal. We mentioned about error I and error II:
we should avoid misclassifying poisonous snakes.
We never know for sure how our findings will be
applied. One futuristic application of our findings is when
integrated with drones (i.e., their webcam). As far as we
know, drones cannot move around areas full of obstacles
(e.g., wide forests), something that may change fast; even
the fact that we are making this proposal herein could
create the motivation for such researches, that could focus
on that. We are aware of researches that try to use drones to
spot targets on physical areas. Once a drone is able to move
in a wild forest, one could sample all the snake species in an
area, even look for a specific species. No need of human’s
interference, avoiding possible accidents, or even human’s
interfering with the wide life. Teachable Machine allows
to choose between photos and webcam, also, it is already
a standard practice between TensorFlow.js practitioners,
the library behind Teachable Machine, to use webcam for
identifying objects. One just need to replace the computer
webcam with the drone’s one.
4 Author contributions
JGP created the model, ran the simulations, designed
the codes, and wrote the manuscript. LHDB helped
with the snake classifications, initial images, writing the
manuscript, and supported on designing the software to
best fit biologists’ requirements for being friendly.
94 Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95
References
Ahmed, K., Gad, M. A. and Aboutabl, A. E. (2023). Snake
species classification using deep learning techniques,
Multimedia Tools and Applications .
https://doi.org/10
.1007/s11042-023-16773-0.
Almeida, P. C. R., Feitosa, D. T., Terribile, L. C., Freitas-
Oliveira, R. and Prudente, A. L. C. (2022). Co-occurrence
patterns between false coral snake atractus latifrons
(günther, 1868) (serpentes: Dipsadidae) and venomous
coral snakes from the amazon, Anais da Academia
Brasileira de Ciências 94(Suppl. 3).
https://doi.org/
10.1590/0001-3765202220210991.
Altman, S. (2023). OpenAI CEO on GPT-4, ChatGPT, and the
Future of AI. Available at
https://www.youtube.com/watc
h?v=L_Guz73e6fw&t=7260s.
Brockman, G. (2023). The Inside Story of ChatGPT’s
Astonishing Potential, TED. Available at
https://www.
youtube.com/watch?v=C_78DM8fG6E&t=626s.
Campbell, J. A. and Lamar, W. W. (1989). The Venomous
Reptiles of Latin America, Comstock Pub. Associates.
Choi, Y. (2023). Why AI Is Incredibly Smart and Shockingly
Stupid, TED. Available at
https://www.youtube.com/wa
tch?v=SvBR0OGT5VI.
Davis Rabosky, A. R., Cox, C. L., Rabosky, D. L., Title, P. O.,
Holmes, I. A., Feldman, A. and McGuire, J. A. (2016).
Coral snakes predict the evolution of mimicry across
new world snakes, Nature Communications .
https://do
i.org/10.1038/ncomms11484.
Durso, A. M., Moorthy, G. K., Mohanty, S. P., Bolon, I.,
Salathé, M. and Ruiz de Castañeda, R. (2021). Supervised
learning computer vision benchmark for snake species
identification from photographs: Implications for
herpetology and global health, Frontiers in Artificial
Intelligence 4.
https://doi.org/10.3389/frai.2021.
582110.
FRONTLINE (2020). In the AgeofAI, FRONTLINE. Available
at https://www.youtube.com/watch?v=5dZ_lvDgevk.
Garcin, C., Joly, A., Bonnet, P., Affouard, A., Lombardo,
J., Chouet, M., Servajean, M., Lorieul, T. and Salmon, J.
(2021). Pl@ntnet-300k: a plant image dataset with high
label ambiguity and a long-tailed distribution, NeurIPS
Datasets and Benchmarks 2021. Available at
https://da
tasets-benchmarks-proceedings.neurips.cc/paper_fi
les/paper/2021/file/7e7757b1e12abcb736ab9a754ffb
617a-Paper-round2.pdf.
Greene, H. and McDiarmid, R. (1981). Coral snake mimicry:
does it occur?, Science 213: 1207–1212.
Hossie, T. J. and Sherratt, T. N. (2013). Defensive posture
and eyespots deter avian predators from attacking
caterpillar models, Animal Behaviour .
https://doi.
org/10.1016/j.anbehav.2013.05.029.
Kahl, S., Wood, C. M., Eibl, M. and Klinck, H. (2021).
Birdnet: A deep learning solution for avian diversity
monitoring, Ecological Informatics 61: 101236.
https:
//doi.org/10.1016/j.ecoinf.2021.101236.
Kalinathan, L., Balasundaram1, P., Ganesh, P., Bathala,
S. S. and Mukesh, R. K. (2021). Automatic snake
classification using deep learning algorithm, Conference
and Labs of the Evaluation Forum, . Available at
https:
//ceur-ws.org/Vol-2936/paper-135.pdf.
Laborde, G. (2021). Learning Tensorflow.Js: Powerful
Machine Learning in JavaScript, O’Reilly Media. Available
at
https://www.amazon.com.br/Learning-Tensorflow-J
s-Powerful-Machine-JavaScript/dp/1492090794.
Pezzuti, T. L., Leite, F. S. F., Rossa-Feres, D. d. C. and Garcia,
P. C. A. (2021). The tadpoles of the iron quadrangle,
southeastern brazil: A baseline for larval knowledge
and anuran conservation in a diverse and threatened
region, South American Journal of Herpetology 22(sp1): 1–
107. https://doi.org/10.2994/SAJH-D-20-00042.1.
Pires, J. G. (2018). Relatório Final de pós-doutorado
Programa Nacional de Pós-doutorado PNPD/CAPES.
Available at
https://www.researchgate.net/publicati
on/329815289_Relatorio_Final_de_pos-doutorado_Pro
grama_Nacional_de_Pos-doutorado_PNPDCAPES.
Pires, J. G. (2022a). ComputationalThinking: How computers
think, decide and learn, when human limits start and
computers champ. Vol.1, self published, Kindle Version.
Available at
https://www.amazon.com/Computational-T
hinking-computers- decide-limits-ebook/dp/B0B5XZCV
CL.
Pires, J. G. (2022b). Innovating with biomathematics:
the challenge of building user-friendly interfaces for
computational biology, Academia Letters .
https://doi.
org/10.20935/AL5792.
Pires, J. G., da Silva, G. F., Weyssow, T., Conforte,
A. J., Pagnoncelli, D., da Silva, F. A. B. and Carels,
N. (2021). Galaxy and mean stack to create a user-
friendly workflow for the rational optimization of cancer
chemotherapy, Frontiers in Genetics 12.
https://doi.or
g/doi:10.3389/fgene.2021.624259.
Poulton, E. B. (1890). The colours of animals: their meaning
and use, especially considered in the case of insects, D.
Appleton and Company.
Rajabizadeh, M. and Rezghi, M. (2021). A comparative
study on image-based snake identification using
machine learning, Scientific Reports 11(1).
https://doi.
org/10.1038/s41598-021-96031-1.
Rivera, J. D. D. S. (2020). Practical TensorFlow.js: Deep
Learning in Web App Development, Apress. Available at
https://www.amazon.com/Practical-TensorFlow-js-D
eep-Learning-Development/dp/1484262727.
Rowley, J., Callaghan, C., Cutajar, T., Portway, C., P. K.,
Mahony, S., Trembath, D., Flemons, P. and Woods, A.
(2023). Frogid: Citizen scientists provide validated
biodiversity data on frogs of australia, Herpetological
Conservation and Biology 14(1): 155–170. Available at
https://www.herpconbio.org/Volume_14/Issue_1/Row
ley_etal_2019.pdf.
Roze, J. A. (1996). Coral snakes of the Americas: biology,
identification, and venoms, Krieger.
Pires & Dias Braga |Revista Brasileira de Computação Aplicada (2023), v.15, n.3, pp.80–95 95
Ruxton, G. D., Sherratt, T. N. and Speed, M. P. (2004).
Avoiding Attack: The Evolutionary Ecology of Crypsis,
Warning Signals and Mimicry, Oxford University Press.
https://doi.org/10.1093/acprof:oso/9780198528609.
001.0001.
Santos, F. G. D., Machiaveli, E. A. and Marciano, J.
M. V. (2021). Snake classifier: Aplicativo mobile para
classificação de serpentes peçonhentas. Available at
https://www.academia.edu/95702218/Snake_Classi
fier_Aplicativo_mobile_para_classifica%C3%A7%C3%
A3o_de_serpentes_pe%C3%A7onhentas.
Savage, J. M. and Slowinski, J. B. (1992). The colouration of
the venomous coral snakes (family elapidae) and their
mimics (families aniliidae and colubridae), Biological
Journal of the Linnean Society 45: 235–254.
https://doi.
org/10.1111/j.1095-8312.1992.tb00642.x.
Secretaria de Saúde (2022). AcidentesporSerpentes, Paraná,
governo do estado, secretaria de saúde. Available at
ht
tps://www.saude.pr.gov.br/Pagina/Acidentes-por- S
erpentes.
Shanqing Cai, S. B., Nielsen, E. D. and Chollet, F.
(2020). Deep Learning with JavaScript: Neural networks
in TensorFlow.js, Manning. Available at
https://www.am
azon.com/Deep-Learning-JavaScript-networks-Tenso
rFlow-js/dp/1617296171.
Stack Overflow (2023). 2023 Developer Survey:
Programming, scripting, and markup languages, Stack
Overflow. Available at
https://survey.stackoverflow
.co/2023/#programming-scripting-and-markup-langu
ages.
Uetz, P., Freed, P., Aguilar, R. and Hošek, J. (2022). The
reptile database. Available at
http://www.reptile-dat
abase.org.
Wickler, W. (1968). Mimicry in Plants and Animals,
McGraw-Hill.
Wolfram, S. (2023). What Is ChatGPT Doing
... and Why Does It Work?, Wolfram Research,
https://www.amazon.com/What-ChatGPT-
Doing-Does-Work/dp/1579550819. Available at
https://www.amazon.com/What-ChatGPT-Doing-Does-W
ork/dp/1579550819.
Yang, Z. and Sinnott, R. O. (2021). Snake detection
and classification using deep learning, Proceedings
of the 54th Hawaii International Conference on System
Sciences, Proceedings of the 54th Hawaii International
Conference on System Sciences, pp. 1212–1222.
Available at
https://scholarspace.manoa.hawaii.edu/
server/api/core/bitstreams/f9f2de4d-ddaa-4984-9fc
3-9bbca4e27247/content.
