Artificial Intelligence

Author: Miguel Angel Granados – Data Scientist

Computer vision has rapidly evolved over the years, offering transformative solutions to a wide range of applications. However, the field still faces two significant challenges: the extraction of meaningful information from visual data due to the inherent complexities of images and videos, and the omnipresent machine learning-focused approach that, primarily relying on massive datasets, powerful neural network architectures, and immense computational resources with great results, it falls short in interpretability and dealing with unseen adversarial scenarios. Thus, the exploration of novel mathematical frameworks and techniques in computer vision is a must to further advance in the field and push its boundaries. In this article we will delve into some of the most exciting approaches, based on topology and fractal geometry.

Instead of solely relying on machine learning and deep learning processes, this article will shift the focus towards theory and algorithms that directly address the core challenges in computer vision tasks such as object recognition, shape analysis, and image segmentation. By incorporating these novel mathematical frameworks and techniques in conjunction with classic machine learning algorithms, we can construct a more comprehensive and robust computer vision pipeline. It is important to remark that appropriate pre-processing of images and a firm understanding of the problem are a must to succeed in merging these novel frameworks into a computer vision pipeline, whether in conjunction with neural networks or other architectures.

Skeletonization

The first technique that we will be exploring is called skeletonization. It consists in reducing a shape to its essential structure or topological features, that is, finding its skeleton, which is constituted by a set of curves or points that capture the shape’s connectivity and topology.

Skeletonize — skimage 0.21.0 documentation (scikit-image.org)

In computer vision

Let´s first look at how one might implement skeletonization for a computer vision project before we dig into the algorithmic and mathematical details: Suppose we want to segment an image of a tree into its different components, such as the trunk, branches, and leaves. We can first apply an edge detection algorithm to the image to obtain a binary mask of the tree, and after that we then apply skeletonization to the mask to obtain the skeleton of the tree. The skeleton can be used to identify the different components of the tree based on their connectivity. For example, the trunk can be identified as the longest branch of the skeleton, and the leaves can be identified as the shortest branches that are connected to the tree’s branches.

Mathematically, skeletonization involves defining a set of measures that capture the shape’s connectivity and topology. These measures are used to identify the points or curves that lie on the skeleton. For skeletonization, we shall work with an input image where the foreground pixels represent the object of interest, and the background pixels represent the rest of the image.

With distance transform

One commonly used measure is the distance transform, for binary images, which assigns to each point in the shape (pixels or foreground pixel) the distance to the nearest boundary point (obstacle or background pixel). The distance transform can be used to identify the points that lie on the skeleton, which are usually the points that have multiple closest boundary points. This measure can be defined on different metrics like Euclidean or Chebysev. We will see its mathematical definition later on.

After computing the distance transform, the next step is the obtention of the skeleton (thinning): here we aim to reduce the object’s boundary to a one-pixel width skeleton while preserving its topological features. Popular thinning algorithms include Zhang-Suen, Guo-Hall, or iterative morphological thinning algorithms. Let us explore a Voronoi-based algorithm:

a. Computation of the Voronoi diagram of the foreground pixels in the binary image with a distance transform. The Voronoi diagram divides the image space into regions, where each region represents the set of points that are closest to a specific foreground pixel. More explicitly, given a set of points (seeds or sites), the Voronoi diagram divides the space into regions such that each region contains all points that are closer to a particular seed than any other seed. In the context of skeletonization, the Voronoi diagram is computed for the foreground pixels in the binary image, where each foreground pixel serves as a seed. Let’s look at the mathematical statement:

Let D(x,y) represent the distance transform of the binary image, which contains the distance of each pixel (x,y) in the image to the nearest background pixel. Then, the Voronoi diagram for the foreground pixels (x,y) $(x, y)$ can be thought as finding the nearest neighbour (xn,yn) among all foreground pixels. Thus, The Voronoi region for the foreground pixel (x,y) is the set of points (x,y) such that the distance to (xn,yn) is smaller than the distance to any other foreground pixel. Mathematically, the Voronoi region $V (x, y)$ $V (x, y)$ for the foreground pixel $(x, y)$ can be defined as:

b. Extraction of the skeleton: The skeleton can be extracted from the Voronoi Diagram by considering the medial axis or centerlines of the Voronoi regions, that is, the set of points that is equidistant to the object’s boundary. The points on the centerlines of the Voronoi regions typically represent the one-pixel width representation of the object’s shape while preserving its connectivity. Mathematically, the skeleton S can be represented as:

For the computation of the centerline, for each boundary pixel, we need to calculate the shortest distance to both the foreground pixels (inside the Voronoi region) and the background pixels (outside the Voronoi region), and those boundary pixels that have equidistant distances to both the foreground and background pixels are the skeleton points. It’s important to note that finding the exact medial axis is a challenging computational problem, and the approach described above provides an approximation of the centerline by considering skeleton points based on the equidistant property. The resulting skeleton may not be continuous or smooth in complex cases, but it serves as a useful one-pixel width representation of the object’s shape for many practical applications.

Giraffe voronoi 3D model 3D printable | CGTrader

With curvature

Another commonly used measure is the curvature, which is used to identify the points where the shape changes direction. Curvature represents the rate of change of the object’s tangent direction along the boundary (contour). Let’s explore an approach of the Curvature-Pruning Skeletonization algorithm:

1.Curve approximation and curvature computing: we need to fit a curve on the boundary and compute the curvature of the object’s boundary or contour in the input image. For fitting, sciPy and numPy include several functions for polynomial or elliptical fitting, such as numpy.polifit(). Various methods can be used to estimate curvature, such as local fitting of curves or derivative-based approaches. The general curvature formula is as follows:

Firstly, we need to estimate the tangent direction at each point along the fitted curve, using numerical differentiation techniques like finite differences, and calculate the derivatives of the fitted curve using finite increments around each point. numPy provides array operations that facilitate the differentiation process. The curvature formula involves dividing expressions containing the tangent components and their second derivatives. Again, array operations from numPy and the mathematical functions from sciPy allow us to perform these calculations efficiently.

Skeleton pruning by contour approximation and the integer medial axis transform – Andres Solis Montero, Jochen Lang. Computers & Graphics

2. Thresholding: this step involves identifying the regions with high curvature. This threshold determines which points along the boundary are considered significant curvature changes and classifies them: High-curvature points are retained, while low-curvature points are marked for removal. In this step we get a representation of a connected network of points with integer coordinates. These points typically lie along the centerlines or medial axes of the object’s high-curvature regions. This whole representation is called Integer Medial Axis.

3.Pruning (thinning): Intuitively, starting with the initial boundary points, we iteratively remove (or prune) those that are of low curvature, while checking on the object’s connectedness. At this point we might consider the skeleton as a graph, where each boundary point is a node, so checking for connectedness becomes a matter of checking if the graph is connected or not. We repeat this process until no further low-curvature points can be removed without breaking the connectivity of the skeleton.

Skeleton pruning by contour approximation and the integer medial axis transform – Andres Solis Montero, Jochen Lang. Computers & Graphics

To wrap up, the skeletonization procedure is a novel mathematical framework in computer vision that allows us to create a simplified representation of the object’s shape based on distance measures or curvature information. An important note is that specific implementation details and algorithms used within each step may vary depending on the requirements of the application and the available tools or libraries. Furthermore, additional pre-processing or post-processing steps may be required depending on the specific use case, since we can encounter complex shapes or unique curve distributions. For instance, smoothing the skeleton may prove beneficial in achieving a refined final product. Additionally, depending on the specific application, we can use the skeleton for further analysis, such as shape recognition, feature extraction, or object tracking.

Fractal Dimension

A recent novel mathematical framework for computer vision uses fractals as the main object for analysis. Fractals are fascinating and complex geometric patterns that exhibit self-similarity at different scales. These patterns are generated through iterative processes, where a simple shape or equation is repeated over and over again, often using recursive formulas. Here is where the Fractal Dimension comes into play. It is a measure of the complexity of the irregularity of a shape, and it is precisely this way of quantifying irregularity that allows us to apply the concept to computer vision.

Archivo:Mandel zoom 04 seehorse tail.jpg – Wikipedia, la enciclopedia libre

Fractals in Computer Vision

Fractals have applications, particularly in image analysis and pattern recognition. The basic idea behind using fractal dimension in computer vision is that it can capture the intricate details and self-similar structures that traditional methods might overlook. This is particularly useful when dealing with complex natural scenes or patterns that exhibit irregular and self-replicating structures. Here are some examples of its applications:

Texture analysis: Fractal dimension has been used to characterize textures in images, such as the surface of a natural stone or the bark of a tree. By calculating the fractal dimension of different regions in the image, it is possible to extract features that capture the texture’s complexity and use these features for classification or recognition tasks.
Medical imaging: it has been used to analyze medical images, such as X-rays or MRIs. By calculating the fractal dimension of different regions of an image, it is possible to detect irregularities or abnormalities in the image, such as the shape of a tumour or the density of a bone.

Fabric Texture Analysis Using Computer Vision Techniques |
Semantic Scholar Fig 12

An example of fractal texture analysis for mammography of
breast… | Download Scientific Diagram (researchgate.net)

The box-counting method

Fractal objects are self-similar, meaning that they exhibit the same patterns and structures at different scales, and the Fractal dimension is a way of quantifying this self-similarity. With this in mind, the first step of a general pipeline we can follow to apply this concept to a computer vision problem is computing the Fractal Dimension itself of the objects or regions of interest within an image. There are several methods to compute it, including correlation dimension and Hausdorff dimension. On this occasion, let us explore and understand the box-counting method, which is programmed in libraries such as scikit-image and in pyfrac.

Intuitively, this method consists of covering the fractal object with boxes of different sizes and counting the number of boxes required to cover the object at each level of size reduction. The process of successively decreasing the size of the boxes used to cover the object is a key aspect of this method. This reduction in box size allows us to explore the fractal at different scales or resolutions. As we progress to smaller boxes, we delve deeper into the fractal’s self-replicating patterns, enabling us to observe more intricate details. Each level of size reduction provides a finer view of the fractal, enhancing our understanding of its complexity and self-similarity.

Applying the box-counting method to the Koch curve. The number of boxes… | Download Scientific Diagram (researchgate.net)

Why does it make sense to compute self-similarity/irregularity like this? This method allows us to look at how much the number of boxes and the box size follow a power-law relationship, that is, how much one quantity changes to a relative change of the other. The slope of this power-law curve is used to estimate the fractal dimension. In the case of a regular fractal, the number of boxes needed to cover the fractal does not decrease linearly with the size of the boxes, following a stable power-law relationship. On the other hand, Irregular fractals often have a fractional fractal dimension.

Mathematical and Algorithmic Process

Mathematically speaking, this is the procedure:

1.Covering the Object with Boxes: The first step is to cover the fractal object (e.g., points or an image) with boxes of a fixed size ε. The size of ε determines the level of detail or resolution at which we are observing the fractal.

2.Counting Boxes: Next, we count the number of boxes $N (𝜀)$ required to cover the fractal object at that specific scale $𝜀$ In some cases, partially covered boxes may be counted as well.

3.Decreasing Box Size: The process is then repeated for smaller box sizes ( $𝜀 / 2, 𝜀 / 4, 𝜀 / 8$ and so on), and the number of boxes needed to cover the fractal object at each level is recorded.

Then we can express the power-law relationship as:

N (ε) \propto 1 / ε^D

It states that the number of boxes required to cover the fractal decreases at a rate proportional to the inverse of the box size raised to the power of the fractal dimension. This means that as we decrease the box size, the number of smaller boxes needed increases exponentially.

The fractal dimension D is estimated from the slope of the log-log plot of the number of boxes N versus the box size ε. Taking the logarithm of both sides of the power-law relationship equation gives:

log(N(ε))\approx-D*log(ε)

Since the relation is inverse, objects with higher fractal dimensions are more irregular and complex, while objects with lower fractal dimensions are smoother and less complex, that is, the object is self-similar and has a constant fractal dimension across scales. Thus, the log-log plot will appear as a straight line.

Design and implementation of an estimator of fractal dimension using fuzzy techniques (researchgate.net)

Algorithmically, the process above is followed, receiving an object (points, or an image), and the initial box size ε. It is important to set the counter variable $N$ to zero for each box size reduction iteration to keep track of the number of boxes covering the fractal object. The following is a general pipeline we could take as a guide to apply this method to a computer vision problem:

1.Fractal Dimension Calculation: scikit-image and mahotas libraries provide functions to compute it with the box-counting method.

2.Feature Extraction: Once the fractal dimension is calculated for different regions or objects in the image, it can be used as a feature descriptor. These features capture the intricacies and self-similar structures present in the visual data, which might be challenging to represent using traditional methods.

3.Classification and Recognition: The extracted fractal dimension features can then be fed into machine learning algorithms for classification and recognition tasks. For example, in texture analysis, the fractal dimension features can differentiate between various types of textures, enabling accurate classification of different surfaces, such as stones or tree barks. Other applications include:

a) Medical Image Analysis: In medical imaging, the computed fractal dimensions can be utilized to detect irregularities or abnormalities. For instance, in X-rays or MRIs, variations in fractal dimension within specific regions might indicate the presence of tumours or abnormalities in the examined tissues.

b) Segmentation and Region of Interest (ROI) Identification: The Fractal dimensions can also be used for segmentation tasks, where it helps in identifying regions of interest within an image. By setting a threshold on the fractal dimension values, certain areas exhibiting desired complexities or irregularities can be highlighted, aiding in further analysis and decision-making.

c) Noise Reduction and Image Enhancement: Fractal dimension can contribute to image denoising and enhancement. By comparing the fractal dimension of different regions, noise can be distinguished from important image features, allowing for targeted denoising and preservation of critical details.

Final remarks

In summary, exploring and comprehending the latest mathematical frameworks in computer vision is crucial to advancing techniques and algorithms, enabling more efficient and creative solutions to the core challenges. Rather than solely relying on brute force or deep learning algorithms, embracing concepts like skeletonization and fractal dimension showcases how mathematics can enrich the computer vision discipline. Integrating these mathematical principles with appropriate machine and deep learning algorithms empowers us to tackle complex problems effectively.

References

Morphology – Distance Transform. University of Edinburgh https://homepages.inf.ed.ac.uk/rbf/HIPR2/distance.htm#:~:text=The distance transform is an,closest boundary from each point.
Skeleton pruning by contour approximation and the integer medial axis transform – Andres Solis Montero, Jochen Lang. Computers & Graphics Volume 36, Issue 5, August 2012, Pages 477-487:
Computing Dirichlet Tessellations in the Plane – P. J. Green, R. Sibson. The Computer Journal, Volume 21, Issue 2, May 1978, Pages 168–173:
The Fascinating World of Voronoi Diagrams – Francesco Bellelli. https://builtin.com/data-science/voronoi-diagram P. J. Green, R. Sibson
Design and implementation of an estimator of fractal dimension using fuzzy techniques – Xianyi Zeng, Ludovic Koehl, Christian Vasseur. January 2001

Miguel Granados – Data Scientist

EQUINOX

what’s ai?

Discover what is AI and how it will become revolutonary in the industry

Author: Maria Alejandra Ruiz – Data Engineer

AI and creativity: Friends or foes?

Recent advances in the field of Artificial Intelligence have sparked a great deal of interest in “creative AI,” as it is called. While it is well known that AI technology is revolutionising various fields of work, at this time, it has been growing in its influence on creative activities and artistic creation. This article aims to show how AI can enhance creativity and whether partnering with algorithms can help harness the full potential of human creativity through a case study on the application of machine learning techniques in art.

“More help than hurt” news in 1962 about human and computers interaction taken from https://twitter.com/PessimistsArc/status/1637793251720675329/photo/1

The case study

The case examines how incorporating machine learning into an artist’s toolkit affects their creative processes and highlights inventive interactions between artists and their machine-learning tools. It was found that to utilise the generative potential of AI and produce art; artists had to learn new techniques and modify their creative processes. In addition, these new activities and skills involved a change in conventional norms and creative procedures in both agents.

Machine learning models are made to recognise links and patterns in data and use knowledge to produce new and distinctive content. These models can be used to create new visual artworks, as well as new musical compositions, fashion collections, and other types of art.

Many artists’ creative processes are significantly impacted by the use of automatic learning models, which can disrupt or continue previous creative processes. These changes are more apparent in restructuring the creative flow around the generative process, the conceptual shifts around the inherent qualities of the outcomes of automatic learning, and the development of artists’ experience.

By analysing how artists, designers, and musicians use machine learning, we can better understand the various implications of Artificial Intelligence. These implications range from automation to complementarity and relate to a fundamental aspect of human life: creativity. By exploring these professionals’ engagement with AI, it is hoped to gain a broader view of how this technology impacts the creative realm and how it can be used beneficially in the future.

The synergy between art, design and AI

“AI art” refers to images and artistic works produced with Artificial Intelligence. AI analyses these compositions and reinterprets them to create new works based on established rules and cues provided by users from vast amounts of artistic data and information that have been created throughout human history. The production of art through AI has become a powerful and persuasive tool for anyone interested in bringing their wildest ideas to life thanks to its processing capabilities, ability to create variations and the complexity of the details of the images. (1)

This is an exciting space because there is a wide field of people within it. We can find scientists and engineers who work with AI, there are engineers with artistic training who work as scientists, and finally, we have artists who also have some technical training, but their primary means of expression is art.

AI creativity depends on the human being?

As humans must design algorithms and programs in order to create and train AI models, the creativity and artistry of Artificial Intelligence are greatly dependent on human intervention. Also, in order for the AI to produce artistic works, precise data and instructions must be provided. In conclusion, AI cannot produce art on its own; rather, it depends on human contribution to the development of tools and the provision of information for the production of new artistic works.

According to April Youngok Kim, artist and associate professor at Myongji University, in a session on “The Future of Art and Design with AI” at Nvidia’s GTC 2023 event, when she first used generative AI like Meet Journey and Dali, she was blown away. It was no longer just a tool but the best collaboration she had ever had.

-Read our article: AI creative implementations: Nvidia GTC 2023 –

The most amazing (for her) was that the final lines of the images the AI created were very similar to the images in her original existing paintings.

It was then that she realised that the images she was thinking of were her creations or a combination of everything that inspires her, including her favourite artists and art forms.
She felt more like a producer than a creator. She is also very grateful that generative AI has pushed the boundaries of what we can create, especially when one feels stuck and opens up new possibilities. (2)

April’s work : http://cocolove1105.dothome.co.kr/index.php/nft_gif-animated-works-_ongoing-project/

“Non-artists can also create art with their imagination. It changed the concept and boundaries of contemporary art”
April Youngok kim

Generative AI has created a new visual language, but traditional artwork and artists will always be appreciated. Handmade objects will become even more prominent as the combination of Generative AI and drawn animation shows.

Conclusion

So, are AI and creativity: Friends or foes? Ultimately, we conclude that Artificial Intelligence (AI) has the potential to revolutionise the way we conceive and consume cultural products radically. This technology allows us to personalise our experiences, which could lead to a world where the line between real images and those generated by AI becomes blurred.
However, we must be aware of the potential benefits and challenges that this technology brings, as well as its impact globally. AI enables people to tackle previously unattainable challenges, such as creating personalised stories and picture books. At the same time, it also fosters an artistic culture that is more diverse and respectful of minority tastes.
It is critical to consider how AI can change the way we consume and produce art and culture. We must be prepared to embrace the opportunities it offers while ensuring that we adequately address the ethical and social challenges that may arise.
AI ultimately has the power to transform our cultural world in significant ways, and it is everyone’s responsibility to understand and guide this change responsibly and ethically.

Stan Lee homage: https://giantmonster.co/2019/02/28/stan-lee-homage/

References

Zureikat, S. (2023, January 30). AI art vs. human art: What AI reveals about creativity. FromLight2Art. https://fromlight2art.com/ai-art-vs-human-art/?cookie-state-change=1680119854535
Attendee portal (n/d). Nvidia.com. Retrieved March 29, 2023, from https://register.nvidia.com/flow/nvidia/gtcspring2023/attendeeportal/page/sessioncatalog/session/1669929697102001pAKQ.
Vetrov, Y. (2017, January 3). Algorithm-driven design: How artificial intelligence is changing design. Smashing Magazine. https://www.smashingmagazine.com/2017/01/algorithm-driven-design-how-artificial-intelligence-changing-design/

M. Alejandra Ruiz – Data Engineer

EQUINOX

what’s ai?

Discover what is AI and how it will become revolutonary in the industry

Author: Carla Acosta – Lead Designer

TechSauce Summit 2023: Our experience in Southeast Asia

From August 15-18, 2023, the first Artificial Intelligence Week was held in Bangkok with the support of the UK Embassy in Thailand. Equinox AI Lab represented the UK, with eight other leading AI companies, and was an exhibitor at the innovation and technology event called TechSauce Summit. It was our first experience in Southeast Asia!

“It was a unique opportunity to build valuable relationships and learn about the technology market in Southeast Asia, which is definitely on the rise. Thailand is ready to invest and become a benchmark in 5.0 technologies.”

Alejandro Salamanca – Head of AI and Data Science

The event was held on August 16-17 at the Queen Sirikit National Convention Center (QSNCC) in Bangkok’s famous and busy Sukhumvit neighbourhood. The venue was one of the many amazing things we witnessed. It is a space that was renovated from 2019 to 2022, and it stole the attendees’ breath. Its name honours the current king Rama X’s mother, Sirikit. Its areas are spacious, bright and with white and golden sparkles. In addition, it has authentic Thai handicrafts as ornaments on the walls. From the first day, we felt in a completely different world, coming from London.

Pictures taken by the Equinox AI Lab team

Organisation was another virtue worth mentioning. Admission was agile and never felt over-scheduled. Moving on to what concerns us most, technology, we were able to share with startups from Thailand, other companies representing countries such as Poland, Japan, Taiwan and France, and more experienced companies that have carved their way into this industry decades ago, such as CMC from Vietnam.

As we met the startups, we realised that many young entrepreneurs are studying and putting their efforts into building competitive ideas that stand out in the Asian market. One such example is Aigen, which seeks to make AI affordable and accessible to all businesses in Thailand to increase the country’s competitive capabilities. Likewise, we saw recurring themes among these rising stars, such as sustainability, innovation in edible materials, and AI in marketing, healthcare and retail industries.

The digital (and non-digital) experiences took the participants’ attention, including us. Within the app of the event, a Treasure Hunt was designed to collect points throughout the event space. These points could be scanned with a QR or could be seen with AR; furthermore, each participant was given an NFT avatar (in fact, it was our first NFT).

Techsauce summit app screeshot taken by us

Also, experiences such as VR stole the attention of the attendees, zombie games, and balancing experiences. Finally, quadruped robots and drones constantly circulated around the venue, making it a space that seemed out of the planet.

The non-digital experiences also gave us a great surprise. As it is common knowledge, Thai massages are recognised worldwide, and what better way to relax after an intense business talk than to attend a massage on the neck, shoulders and head at the TechSauce Summit?

In addition, they had experiences such as analysing the signature, the wheel of fortune and the lucky number by a specialised Feng Shui consultant; these were definitely new experiences for us.

Pictures taken by the Equinox AI Lab team

In general, we consider this experience completely enriching. We can attest that attending an event like this with a booth is an excellent opportunity to make a lot of contacts that can end in conversions. The event created a superb balance between large, medium and small companies, generating a very promising environment for all. Everyone had opportunities regardless of their size or expertise.

Picture taken by the UK embassy in Thailand

We realised that many companies of diverse industries are at the right maturity level to implement AI in their business processes and strategy, making it a desirable market for our AI and Data Science solutions.

To conclude, we thank everyone who made this event a more enjoyable space for Equinox AI Lab. Thanks to the UK Embassy in Thailand, our booth partners, the people who came with so much enthusiasm to listen to us, and all those who also shared their solutions as we went to their stands. Let’s keep building out-of-this-world solutions. This is just the beginning!

Pd: Thai people are the kindest, and the food is fantastic. We highly recommend visiting this lovely country <3.

Image composed by AI generated images and the magic wand of our designers

Carla Acosta – Lead Designer

EQUINOX

what’s ai?

Discover what is AI and how it will become revolutonary in the industry

Author: Arturo Gutierrez – RPA Engineer

INTRO

Indeed many people wonder how industry giants such as Netflix, Unilever and ANZ Bank have been able to improve their employees’ professional growth, increase their profits and optimise their internal processes. Many point out that their development has been due to their financial muscle or the ideas of someone who has knowledge from other planets.

Nothing could be further from the truth. The truth can be found in a much simpler and devastatingly effective point, such as managing daily activities effectively and improving employees’ skills. This is known worldwide as “profit per employee”, a new measure that leaves behind obsolete topics such as ROI and begins to qualify a company by the value that an employee provides over time.

Now, the question is, how can we improve these rates so that the company is more like the current tycoons that are generating millions of dollars a month?

What can we invest in to obtain the desired “profit per employee” that everyone is talking about, or better yet, as employees complain about the repetitive tasks they are assigned, how can we prevent them from getting bored with their work and ending up leaving the company?

The answer to these questions is found in the term Intelligent Process Automation or IPA: the digital transformation enabler, also known as IA (Intelligent Automation) or Cognitive Automation, which is based on the integration of technologies such as Artificial Intelligence and RPA robotics for the automation of organisational processes.

TACKLING THE MYTHS OF IPA

MYTH 1: IPA is going to kill jobs

Today’s companies need fewer employees, which should no longer be seen as negative. Several studies indicate that more people die of stress or exploitation at work than in wars and terminal illnesses. More than 75% of people are dissatisfied with their jobs because they are repetitive and tedious. It is essential to see an alternative to these jobs that are destroying employees’ professional careers and to see new business strategies that drive “profit per employee” in acquiring new skills and technologies.

69% of workers expect that through automation, they can be more productive in their main professional activities, and 86% of employees think that the use of automation in the workspace drives them to learn about issues of value to the organisation, the fact that fewer employees are needed should be seen as an opportunity for professional growth in the collaborators.

Stable Diffusion generated image of robots working

What IPA can offer in the long term is known as the “Triple A artefact”, which is all the benefits that can be achieved through the implementation of technologies in the processes, divided into three critical factors:

Automation: Repetitive tasks in the processes must be recognised and automated.
Augment: Understand the processes’ behaviour over time and their benefit to the company through data analysis tools.
Abandon: eliminate activities that do not add value to the companies and reflect an unnecessary expenditure of time and effort.

The following image shows how IPA can improve the performance of those repetitive activities that occur in companies daily and increase the “Profit per employee”.

Profit per employee improvement with automation

As can be seen, many repetitive and low-value activities can be automated through IPA, and tools can be generated to help workers improve their performance and professional growth.

MYTH 2: There is insufficient capital or knowledge to implement this technology

To carry out a good development of IPA: the digital transformation enabler in companies is necessary to know the stages of transition and digital transformation that must be adopted to have an exponential value in the business over time. If these three phases are followed with commitment, patience and discipline, it is possible to reach the technological development that business giants have so far.

RPA strategic value phases

Phase 1: Efficiency

Efficiency is needed for automating and identifying through RPA all those repetitive and routine processes currently in the company.

The potential of this technology is not being exploited. RPA generates an ROI between 30-200% in the first 18 months, and in terms of progress in 2021, the RPA market was 3.5 billion USD; it is expected to grow by 40% per year.

But the integration of RPA at this point is not enough; there are many challenges of scalability, strategic investments and expected benefits that a second phase must handle.

Phase 2: Effectiveness

Strategies for scaling this technology over time are being developed. Customers own between 1 and 50 robots. Few have scaled from 51 – 100, and methods are expected to improve the transition from phase 1 to phase 2. However, there are still processes that RPA cannot carry out due to their complexity and data types. For them, it is crucial to scale RPA to an enterprise level and create intelligent automation through a third phase.

Phase 3: Enabling

This is the last phase and where IPA generates its greatest expected benefits. For this phase to bear fruits, it must consider the automation of mid-office processes, back office and customer-facing activities, handle unstructured data, analytics and probabilistic decisions and take into account that most of the challenges are organisational and managerial.

Companies must become digital businesses, where the great value lies in creating a platform of digital options that provide businesses with flexibility, adaptability, strategic choices and resilience at low cost.

As reflected in the explanation of each phase, it is necessary to go from less to more in these integrations. It is advisable to start by identifying through a diagnosis the repetitive processes that can be automated, see how more automation can be scaled at the enterprise level (having more software robots) and finally, carry out a phase of digital transformation where the application of technologies such as artificial intelligence and RPA (IPA) go hand in hand in complex processes that allow obtaining statistics and generate tools to complement the work of employees.

MYTH 3: IPA and Digital Transformation does not apply to my current area and type of business

There are different methods of applying IPA in the processes; the most relevant and which are in a phase of study and application in Equinox AI Lab are the following:

-Document Understanding:

It is known as intelligent document processing. Its main objective is to process documents of different formats and obtain relevant information by integrating two technologies, artificial intelligence and RPA.

This technology is applicable to different industry sectors processing documents such as invoices, receipts, purchase orders, utility bills, landing invoices, passports, and licenses, among other documents that RPA alone cannot extract due to the complexity of their structure and context.

Some examples applicable to specific business areas may include the following:

Financial services and insurance: invoices, IRS Forms, loan applications, mortgage processing, account opening and customer onboarding, claims, vendor onboarding, and processes related to regulatory compliance.
Human Resources: employee onboarding, resume screening and HR file processing.
Manufacturing: sales order processing, customer orders and shipment processing.
Public sector: migrations, school applications and passport applications.
Health: medical forms, medical invoices, medical records and medication prescriptions.

-Here we have some tips for good intelligent document processing:

Have a proper flow for the different processes. It is suggested to use the following.

RPA workflow for company’s processes

Taxonomy is the understanding of the document. In this phase, it is important to define the types of documents and the fields to be obtained through the extraction.
Digitalise: convert the documents through OCR into machine-readable data.
Classify: classify and separate the files into document types.
Extract: extract the information from the documents.
Export: export the data for later use in other processes.

Acquire a balanced variability, and obtain the appropriate number of documents for training the model. In this case, multiplying the number of fields to be extracted by 25 is essential.
Generate a clustering algorithm; choose ten copies of each format and no more than ten if there are many formats.

NLP + RPA

Natural Language Processing (NLP) is a technology that allows the analysis of large amounts of data in text or audio, allowing the computer to interpret and understand human language based on Machine Learning.

For the correct implementation of NLP models, especially in the case of IPA, it is vital to establish the objective, quantity and quality of data you have. The training and evaluation files will be generated according to the model and the objective.

Usually, the input data can be entered with CSV or JSON files, and the output will be in JSON format. These models use different pattern identification techniques and can be used in the following:

Classifying individual data or text: for example, to identify whether a text has a positive or negative meaning.
Extracting entities by category: as in the case of scientific reports from which, it is necessary to extract the chemicals that are mentioned and classify them by type.

CONCLUSIONS

Working with IPA: the digital transformation enabler allows us to demystify certain beliefs that do not allow companies to reach the digital transformation they want to have.

It can be observed that increasing the “Profit per employee” through a triple A artefact benefits not only the professional career of workers but the company’s progress, abandoning those activities that represent a waste of time and focusing on the development of technological tools for improving processes.

It is not necessary to have significant capital and knowledge of another world to apply this technology, just be patient and rely on the three phases of digital transformation “efficiency”, “effectiveness”, and “enablement”. IPA is scalable and applicable to different business areas, and it is hoped that, through this article, the readers will be able to understand the impact that IPA is generating in companies and how it can be helpful to benefit them positively.

REFERENCES

Intelligent Automation – Learn How to Harness Artificial Intelligence to Boost Business & Make Our World More Human (Libro)

Becoming strategic with Intelligent Automation – leslie willcoks

UiPath document Understanding – Documentacion Uipath

Document Understanding: Cómo prepararse para una implementación exitosa – See Document Understanding: Cómo prepararse para una implementación exitosa at UiPath Colombia

NLP – ¿Qué es el procesamiento de lenguaje natural? – Explicación del procesamiento de lenguaje natural – AWS (amazon.com)

TextClassification – Documentación de UiPath

Arturo Gutierrez – RPA Engineer

EQUINOX

what’s ai?

Discover what is AI and how it will become revolutonary in the industry

Author: Oscar Sanchez – Data Architect

ABSTRACT

AI models for predicting cardiac diseases [The Colombian example]

Around the world, heart diseases are the most common cause of death, generally because people don’t pay attention to them. At the same time, there are not enough heart specialists for many people who have or will have heart problems in the future.

For this reason, this article proposes using Artificial Intelligence to predict heart diseases before patients have delicate issues like heart attacks.

Also, to establish how many specialists are needed in different areas to be effective and lose fewer lives because of heart diseases.

Take from Freepik

Cardiac diseases around the world

Around the world, cardiovascular diseases are the first cause of death. The percentage of global deaths caused by cardiovascular problems is expected to be 45% (WHO, 2021) . Besides, out of the 17 million premature deaths (under 70 y.o) due to noncommunicable diseases in 2019, 38% were caused by CVDs.

As a result of the pandemic caused by COVID-19, a significant sample of patients who had complications have presented an increased risk of sudden death, acute myocardial infarction, and arrhythmias, among others (Triana).

Under these circumstances, the percentage of deaths from cardiovascular problems, far from decreasing, will increase significantly after 2020, and it is already a global public health problem. Also, the world is critically concerned about the lack of heart specialists. There are not enough heart specialists to cover the high demand.

Despite being the major cause of mortality, preventive campaigns have not been effective, and many people have died without knowing they had a heart complication to treat.

Cardiac diseases in Colombia and their treatment

Not only around the world, but also in Colombia, cardiac diseases are the most popular cause of dead (minsalud, s.f.).

Figure 1: causes of dead in Colombia source: DANE

According to the image, the main cause of death in Colombia is heart issues, with a total of 13.926 deaths only in the first quarter of 2022.

Despite many prevention campaigns, the number of deaths keeps growing, and lacking specialists doesn’t facilitate the situation.

Also, patients or their health provider entities cannot cover many treatment costs. Because of this, this issue is a main problem for Colombia.

The Colombian government invested in 2017 around 6.4 billion pesos (1.5 million dollars) to treat cardiac diseases (cost, 2017) , but the number of deaths continued to grow, so we have to ask ourselves if investing more solves the problem.

Prediction of cardiac disease with AI

With the use of Artificial Intelligence, we can support and diagnose cardiovascular diseases early, and we wanted to experiment to create a reliable solution.

First, we took into account significant variables such as:

BMI (body mass index)
Bad habits (tobacco, alcohol) (information provided by the patient)
High blood sugar levels (diagnostic tests)
High blood cholesterol levels (diagnostic tests)
Fruit consumption (information provided by the patient)
Vegetable consumption (information provided by the patient)
Having hypertension (diagnostic tests)
Percentage of body fat (measurable values)
Percentage of visceral fat (measurable values)

As you can see, these types of variables are categorised into two types:

Diagnostic tests or measurable values
Subjective information provided by the patient.

The solution to the problem arises from the extraction of the information of the patients from their clinical history, searching for the variables described above to build a training dataset with the highest possible reliability.

For this exercise, we used a dataset provided by Kaggle (kaggle, 2023) , which greatly approximates global behaviour.

After this, we proceed with the training of the adjusted model and, in the same way, the exposure of an API to evaluate new patients.

The response to this evaluation will categorise the patient as a patient without risk or with a potential heart problem. So, the number of cardiologists needed can also be predicted according to the number of patients at risk of developing heart disease.

Applicable AI models

The applicable AI models according to the behaviour of data are:

Decision tree: a type of supervised learning algorithm used in machine learning and data mining that uses a flowchart-like structure to visualise the decisions made by the algorithm based on the input features of the data (Liberman, 2017) .

Random forest: is a supervised learning algorithm used in machine learning that combines multiple decision trees to improve the accuracy and robustness of the model.

The dataset we chose has 253.680 rows for data training and was tested (80/20) with the variables mentioned before. The best accuracy it reached was with random forest, with 70%.

It means that of 100 people, the model could hit 70. It is a good number for saving lives without the cost of corrective treatments in a country like Colombia.

Application and utilities of an accurate prediction

Accurately predicting cardiac disease may help apply preventive treatments to save lives cheaply. In the same way, it might reduce patient risk because preventive treatments are cheaper than corrective treatments and more effective (cost, 2017) .

In a country like Colombia, it may make a difference and increase the coverage around vulnerable populations.

Also, with the classification of people, you may determine the need for a heart specialist in a specific area in Colombia to distribute such specialists better, as they are very short-staffed.

CONCLUSION

AI may be a great supporting tool not only for patients who should improve their health habits and have a good preventive treatment to save their lives but also for getting good use of the heart specialist, knowing that it is a scarce resource in Colombia and worldwide.

Around the world, many initiatives exist for using AI in different health issues, mainly for diagnostics. An early diagnosis of any disease makes the difference between life and death.

According to data protection, health entities should participate in open data initiatives in Colombia to provide good input for research on health topics and other issues for the country’s growth.

REFERENCES

bayer. (2020). Obtenido de https://www.bayer.com/es/co/las-enfermedades-cardiovasculares-son-la-primera-causa-de-muerte-en-colombia-y-el-mundo#:~:text=septiembre%2001%2C%202020-,Las%20enfermedades%20cardiovasculares%20son%20la%20primera%20causa%20de%20muerte%20en,15.543%20a%20enfer

cost. (2017). Obtenido de https://consultorsalud.com/colombia-invierte-64-billones-al-ano-en-tratar-enfermedades-cardiacas/

kaggle. (2023). Obtenido de www.kaggle.com

Liberman, N. (1 de 2017). Obtenido de https://towardsdatascience.com/decision-trees-and-random-forests-df0c3123f991

minsalud. (s.f.). Obtenido de https://www.minsalud.gov.co/salud/publica/PENT/Paginas/enfermedades-cardiovasculares.aspx

WHO. (2021). Obtenido de https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds)

Oscar Sanchez – Data Architect

EQUINOX

what’s ai?

Discover what is AI and how it will become revolutonary in the industry

Author: Ivan Caballero – AI Designer

INTRODUCTION

If you were asked which animal is the world’s deadliest to humans, what animal would you say? Or if A bat and a ball cost $1.10, but the bat costs one dollar more than the ball. Do you know how much the ball costs? Or one that my father asked me once when I was a 10-year-old: what weighs more, a kilo of feathers or a kilo of potatoes?

Perhaps you have heard or faced this type of tricky riddle that is usually used to make you fail in the answer and realise something: we struggle to be critical. Don’t worry; it is natural and inevitable in human beings. In trying to save energy and reduce cognitive load, our mind uses previous experiences to interpret information, make quick decisions, and have judgments about something. However, this process makes mistakes because many variables are overlooked, leaving us short of data and making incorrect judgments. This effect is known as cognitive bias.

In Data Science projects, it is crucial to consider the possible cognitive biases that teams face daily to know how we can avoid them because when understanding and processing a data set, building and designing artificial intelligence models or presenting information in graphs, these biases can affect the performance of what is built or present a completely distorted reality. That is why with this blog, we want to offer some online resources that can help you learn about the different varieties of biases and how to avoid them.

Wikipedia’s Cognitive Bias Codex

Confirmation bias diagram by Equinox AI Lab

https://commons.wikimedia.org/wiki/File:Cognitive_bias_codex_en.svg

This is a compilation of 188 cognitive biases categorised by Buster Benson, a marketing manager at Slack, and illustrated by user John Manoogian III (jm3) and linked to the articles written by user TilmannR on Wikipedia. Although for many, Wikipedia may not be a platform to trust; this graphic representation can help us have a quick reference guide to learn and keep in mind each of the biases and their leading causes.

Decision Lab list

Framing Effect by Equinox AI Lab

https://thedecisionlab.com/

This behavioural science firm has a list of biases categorised into ambiguity, information overload, memory, and speed. Each bias entry describes the bias, when and why it occurs, its effects when making decisions, how to prevent them, and daily life examples. This resource accurately explains biases, and its examples are easy to understand.

On the other hand, it is important for Data Science projects to present and analyse data effectively and realistically. Data visualisation should not be taken lightly. So here we also give you two key resources when it comes to avoiding biases in data visualisation.

From Data to Viz

Disposition Effect by Equinox AI Lab

https://www.data-to-viz.com/caveats.html

This website is intended to be a guide for designing accurate graphs according to what you need to represent in visualisation. Each bias entry is titled with the intention we have when plotting, the good and bad examples when using it and the considerations we must keep in mind to decide if the graph fits our intention.

Geckoboard

Availability bias by Equinox AI Lab

https://www.geckoboard.com/best-practice/statistical-fallacies/#.WsXXTmrFLIU

This company has a section where it lists 15 Data fallacies that may occur when interpreting and analysing data. There you will find what they are about, how to avoid them, and resources related to each fallacy.

Finally, you now have some resources you can take advantage of to avoid biases in Data Science projects. And if you were wondering what the answers to each of the riddles are, here are their solution: the deadliest animal for humans is the mosquito, the ball costs 5 cents, and to my 10-year-old self, both the feathers and the potatoes weigh a kilo :’).

Ivan Caballero – AI Designer

EQUINOX

what’s ai?

Discover what is AI and how it will become revolutonary in the industry