Logo detection in sports sponsorship

Logo detection in sports sponsorship

/in /by and

Consumers love brands that bring them closer to sporting events. This has compelled the largest brands to jump headlong into sports sponsorship. While the benefits of sponsorship are undeniable, measuring ROI precisely remains a challenge.

With machine learning-powered tools, brands can evaluate a campaign’s ROI by analyzing video coverage of sponsored events. Thanks to computer vision solutions, it is possible to precisely determine the time and place a brand was positioned during a sporting event. Image analysis also allows companies to see the branding activities of competitors and compare their brands’ visibility. Having precise information on brand positioning also enables advertising equivalent value to be calculated, and the most impactful events to be determined and the activities of competitors to be monitored. Such analyses would be extremely time consuming and far less accurate if performed manually. Automated analysis based on advanced machine learning algorithms allows brands to quickly provide new valuable insights and boost the effectiveness of marketing campaigns. To address these needs, deepsense.ai developed an automated tool for logo detection and visibility analysis that provides both raw detection and a rich set of statistics.

Solution overview

deepsense.ai’s solution is based on a combination of two approaches – supervised object detection and one-shot learning.

Supervised object detection approach

In this approach, to train models on a labelled data set, one of the many well-tested architectures can be used. They include fully convolutional YOLO / SSD or the R-CNN family (regional convolutional neural networks). Since video streaming is essential to logo detection during sports broadcasts, a fully convolutional model is the best choice. In this case, the model does not have to process each proposal region separately and has a constant inference time, independent of the number of objects detected. This enables the model to run in real time.

The advantages of this approach include:

  • Simplicity of operation, well-developed use cases (many open, tested and ready-to-use implementations).

But there are also disadvantages:

  • It’s impossible to quickly add a new version of a logo without obtaining a large amount of training data;
  • The system is very sensitive to changes in the appearance of the logo, and updating a model trained as described above would require a large amount of new data.

Therefore, in order to increase the efficiency of the system, the approaches were combined (supervised object detection + one-shot learning).

The One-shot learning approach

This approach effectively solves the problem of dynamic logos and allows us to add new logos to the database without the need to collect large amounts of data. All we need is to have a reference set of template vectors for each supported logo and model detecting logo region proposals without performing classification on them. The model is trained by a process known as triplet loss.

During the training process, each mini-batch consists of a three element tuple:

  1. Example of company A logo (anchor),
  2. Photo with regions with the company’s logo (positive),
  3. Image with regions with a different brand’s logo (negative).

For architecture, we use the fully convolutional YOLOv3 model, which will both embed the template set of logos into a certain relatively low-dimensional vector space and detect objects in photos (but without assigning them specific classes).

During training, the “template logo” (anchor) is encoded by the same model that is used for object detection. The one difference is that on the map of features extracted for anchor, we calculate average pooling to obtain a single feature vector-anchor vector.

An approximate diagram of this process is presented in the figure below.
Logo detection in sports sponsorship 1

The optimized target function in this case is triplet margin loss – a differentiable function that yields small values if the vector representing the positive region is close to the anchor pattern vector (the logos are similar) and the negative region vector is far away from it (they are not similar).

After training, the model will process the available template logos to create a template vector database for each of the classes supported. In the inference after the logo region is detected, we extract the vector representing this region and compare it with the patterns. The label of the most similar is selected as the class of the given detection.

Updating a model trained in this way requires only that new elements be added to the reference vector database, or the “old” logo be replaced with “new” one, without having to train the model for optimization issues.

Parallelization of the stream

To speed up the system’s performance, we parallelized the stream without processing it “frame by frame”. As the streaming data flowed in gradually over time, we opted not to use “batch” inference with one instance of the model. In this context, it is also important to synchronize the processes in order to return the processed stream elements in chronological order.

  1. We initialize n workers (parameter, natural number). Each of them is simply a pair of YOLOv3 detection networks (one trained with a supervised method and the other with one-shot).
  2. We create a FIFO queue to which we throw the incoming data from the stream and from which workers collect frames for processing.
  3. To ensure the chronology, the processed frames are thrown by workers into the heap,
  4. A separate, looped process checks that the heap is not empty. If it isn’t, then the element with the smallest id is taken from it; if the value is 1 greater than the id of the last processed frame, then we update the value of this variable and return the processed frame, alternatively the frame and id are thrown back onto the heap.

The diagram below presents an approximate scheme of the system.
Logo detection in sports sponsorship 2

Enabling live processing, this approach significantly improves the performance.

Logo detection analytics

Automated logo detection analytics helps advertisers evaluate the results of sponsorships by providing a series of statistics, charts, rankings and visualizations that can be assembled into a concise report. The statistics can be calculated globally and per brand. Some features include brand exposure size in time, heatmaps of a logo’s position on the screen and bar charts to allow you to easily compare various statistics across the brands. Last but not least, we have a module for creating highlights – visualizations of the bounding boxes detected by the model. This module serves a double purpose: in addition to making the analysis easy to track, such visualizations are also a source of valuable information for data scientists tweaking the model.

Paramount factors in successful machine learning projects

Paramount factors in successful machine learning projects. Part 1/2.

/in /by

Much has been said about the effective running of machine learning projects. However, the topic keeps coming up. Data Scientists spend a lot of time discussing modeling methods, while – in my opinion – the overarching goal of running machine learning projects in companies fades into the background. It is vital to remember that the purpose of ML projects is not modeling itself, but achieving defined business goals.

While for data scientists modeling is often the most exciting part of the job, the other steps of an ML project should not be neglected, as doing so may imperil the valuable business results we set out to achieve in the first place. Properly executing the entire process of building and implementing machine learning models is essential.

PROCESS. As you make your bed, so you must lie upon it.

The entire ML project process can be described as a five-point checklist.

  1. FRAMING – the main goal here is to determine the essence of a business problem and phrase it in Data Science lingo.
  2. DATA – fuels the whole solution, so we need to painstakingly examine and understand it.
  3. MODELING – building the model is the core activity and often is viewed as the most exciting part.
  4. PRESENTATION & CONTINUATION – for our efforts to be truly fully appreciated, both the results and the solution must be described in a way that it is understandable and useful for business stakeholders.
  5. PRODUCTION & MAINTENANCE – because data science projects require so much experimentation, how they are to be brought to the production stage may end up being neglected. Is the code up to snuff? Do my timelines need to be adjusted? Will the solution we produce ultimately solve the problem stakeholders need solved? Make sure that you think about this at some point.
FRAMING. All that glitters is not gold.

Often, before we start a project, it seems that everyone has the same understanding of the basic concepts and purpose. It is worth double-checking to ensure they do. A project deepsense.ai did for a client in the banking sector may serve as a good example of why. Our job was to predict churn, which seemed straightforward enough at first glance. The problem was that everyone’s underlying interpretation of “churn” was slightly different. Only a series of detailed questions allowed us to agree on what exactly we define as “churn,” taking into account the time horizon, specific customer activities and eventual net profit from this customer. The more thoroughly we analyze and define a business problem, the more precisely we will be able to transfer it into metrics.

These are the questions that are worth answering during this phase:

How will the solution be used?

Determine at the outset in what context the results of our modeling will be used. Will it be input for business decisions, support for process automation or just some improvements within the system.

How should performance be measured?

Discuss the KPIs and establish how project success will be measured. The importance of this point shouldn’t be underestimated.

How will the solution be tested and validated?

Consider how your validation pipeline should look like during the development and what additional testing should be done during and after it. Remember that business stakeholders must be able to confirm that the solution works well.

Will additional requirements or limitations arise?

Last but not least, analyze technical capabilities and possible limitations related to technical constraints, data extraction or tolerable latency.

DATA. A bad workman blames his tools.

The garbage in / garbage out principle is well known. I would encourage you to look at it not only from the perspective of data, but also of the model development.

Key aspects at this stage include:

Understand what data can be available, and request it early

Due to the complexity of business processes, we almost always encounter difficulties in obtaining properly prepared data. Take this into account and plan more time to request and validate data – possibly even before we officially kick off the project.

Understand how the data was extracted and preprocessed

Surprises are rare here, but when they occur, they tend to be big ones. The person responsible for data extraction may make certain decisions or errors that can significantly distort the results – either by accident, miscommunication or simply out of a lack of knowledge of modelling practices. It is always a good idea to understand the data extraction process and later thoroughly review the data with the business and data owners to confirm that you are all on the same page.

Explore the data

First of all, double-check that you have all the data you requested. Then perform the critical step of pre-modelling Exploratory Data Analysis. It should help you understand the data, the problem itself, generate insights, discover patterns, spot anomalies and test hypotheses. You can approach the data exploration as if it were up to you, rather than the model, to generate predictions. Don’t forget the simple stuff at this stage: ask basic questions, compute statistics, check features and labels distribution.

Confront your findings with the data owners

With all the findings from the EDA, conclusions, insights, and new hypotheses, talk to the data owners and business stakeholders. Confronting all of these early and with people that know the data and problem (hopefully better and with more intuition than we have) will provide a better foundation for modelling.

MODEL. It is never late to learn.

I hope I won’t disappoint you here, but given the importance of modelling and people’s eagerness to do it, I’m going to cover it in detail in the second part. As for now, just for the sake of order, let’s just cover a few key aspects.

Try a number of different approaches

Modeling is a process of constant experimentation and it is always worth trying a number of different approaches to models, features and hyperparameters. The key thing here is to do this in the proper order – starting from basic benchmarks and standard or off-the-shelf approaches before rolling up your sleeves and unleashing your creativity.

Understand errors, important variables, predictions, …

Understanding those will give us ideas for improvement as well as allow us to catch potential issues or anomalies in time.

Confront your findings and predictions with business stakeholders

Whatever your findings, make sure they are either expected or you work on explaining them. Another good educational exercise is asking business stakeholders what they expect the data to reveal and cross-checking this with what you discover.

Do at least a couple of iterations

Don’t forget the experimental nature of modeling. Developing a good solution usually takes time and requires multiple iterations. Don’t be afraid to revisit earlier steps if necessary – especially after discovering something new about the problem/data or stalling.

PRESENTATION & CONTINUATION. All’s well that ends well.

How and why does your solution achieve your business objective?

A well-thought-out presentation of the project’s results may determine its success or failure. That is why we need to pay particular attention to emphasizing the business value of the project and the impact of the modeling on specific business processes. Otherwise business stakeholders may not understand the solution we deliver or its actual value. Doing the work is one thing but solving the problem and convincing others that we have done so is quite another.

What steps are necessary towards full-deployment/productionization?

More often than not the PoC doesn’t fully reflect the project’s business value. Therefore, at this stage, ask yourself “what’s next?” in order to determine what further developments are needed to achieve the desired results.

Optional: Prepare for the handover

If you are aiming for a full handover, be sure that both sides are on the same page – no one likes surprises here.

PRODUCTION & MAINTENANCE. Don’t count your chickens before they hatch.

Make sure your code is production-ready

Because of the experimental nature of data science, code quality and general software engineering principles may take a back seat to modelling. Going “live” is the last call to account for this.

Monitor, measure, and retrain only if necessary

Make sure that you monitor both inputs (feature space) and outputs (your predictions and actual labels if possible). Detecting any data shifts late will surely make many people unhappy. As for retraining, understand how often it has to be done and figure out the right degree of automation.

THE DEEPSENSE.AI TAKEAWAY

Many businesses have already learned the value of putting machine learning to use. The role of Data Science or ML teams is growing and advanced data analysis is becoming a key factor to support strategic imperatives. It is therefore crucial that data scientists keep in mind the main goal of the project – solving or improving a specific business problem (as opposed to just playing with ML). The modelling itself may well be the key or core ML activity, yet if unaccompanied by all the other steps it won’t achieve the ultimate goal. Hopefully, the above checklist will improve your collaboration with business stakeholders and bring greater success to your projects.

Machine Learning for Applications in Retail and Manufacturing

Machine learning for applications in retail and manufacturing

/in /by

Machine learning is still perceived as an innovative approach in business. The technological progress and the use of Big Data in business make ML-based solutions increasingly important. As Forbes magazine indicates, 76% of enterprises today prioritize artificial intelligence and machine learning over other IT initiatives [1].

The concept of machine learning is derived from advanced data analysis of pattern and dependency recognition. ML assumes that when models analyze new data, they can adapt independently and use this new knowledge to develop by learning from previous experiences. Machine learning models can enhance nearly every aspect of a business, from marketing and sales to maintenance.

Predictive maintenance

Machine learning enables predictive monitoring, with algorithms anticipating equipment breakdowns before they occur and scheduling timely maintenance. With the work it did on predictive maintenance in medical devices, deepsense.ai reduced one client’s downtime by 15%.

But it isn’t just in straightforward failure prediction where machine learning supports maintenance. In another recent application, our team delivered a system that automates industrial documentation digitization, effectively reducing workflow time by up to 90%. We developed a model that recognizes and adds descriptions for all symbols used in the installation documentation. The schematics, including the technical descriptions of all components, are fully digitalized. The model reduces the work to a 30-minute review by a specialist. It also handles the most tedious tasks, thus reducing the effort required of human specialists and the number of errors they make in performing them.

An international manufacturer of medical devices was looking for a solution that would reduce device downtime. Our experts built a predictive maintenance model that pores over historical data, searching for anomalies and signs of a breakdown before one occurs. The model reduced breakdown-related downtime by more than 15%. Such a  solution can be applied in machine-reliant industries, where breakdowns bring operations to a halt and hamper overall company performance.

Quality control

Machine learning is also being adopted for product inspection and quality control. ML-based computer vision algorithms can learn from a set of samples to distinguish the “good” from the flawed. In particular, semi-supervised anomaly detection algorithms require only “good” samples in their training set, making a library of possible defects unnecessary. Alternatively, a solution can be developed that compares samples to typical cases of defects.

One of our clients asked us to tackle two of their visualization problems on a food production line – to detect sauce smears on the product’s inner packaging, and to identify correct positioning of the product’s topping. The system deepsense.ai delivered was able to identify 99% of faulty products with topping defects and while raising the alarm with a 99% accuracy rate for sauce smears. The model significantly reduced the need for manual quality control, hence lowering costs.

ML-based computer vision solutions can also be an essential component in the monitoring of hazardous areas in factories, tracking whether every worker is following safety requirements (including wearing helmets, glasses, vests, earmuffs  etc.) and, if not, sending an instant alert to the supervisor with a detailed description of the event that has occurred.

Demand forecasting

In the field of predictive analytics demand forecasting can be used to predict consumer demand. Such forecasting is done by analyzing statistical data and looking for patterns and correlations. With machine learning taking the practice to a higher level, modern demand forecasting techniques go far beyond simple historical data analysis.

More recent techniques combine intuition with historical data. Modern merchants can dig into their data in a search for trends and patterns. At the pinnacle of these techniques sit demand forecasting machine learning models, including gradient boosting and neural networks, which are currently the most popular types and outperform classic statistics-based methods. Historical data from transactions form the basis of more recent demand forecasting techniques. These are data that sellers collect and store for fiscal and legal reasons. Because they are also searchable, these data are the easiest to use.

This modern approach is extremely effective. One of our clients from the retail industry was losing millions of euros a year due to out-of-stocks. There was a daily cap on how many new items its warehouse could receive. Our team built a demand forecasting model for products that were new to market. It enables the company to use the cap more efficiently by ordering more hot products and fewer of those that are less in demand. We used Gradient Boosting, Random Forest and Neural Networks to build the model, and the trifecta reduced out-of-stocks by 30%.

Marketing optimization

Companies can maximize ROI on their marketing activities by implementing machine learning into their customer analysis. Sophisticated data analysis helps identify customers with the highest ROI on ads to make the most of marketing campaigns. It also optimizes channel mix with advanced attribution models.

deepsense.ai designed a model for a leading mobile advertising platform that predicts the click-through rate of internet advertisements. The model analyzes historical data on site user behavior to spot patterns and uncover anomalies. It enabled clients to identify an abnormal pattern among users, which turned out to be bots engaging in fraudulent clicking. The solution effectively identified internet bots that click ads, significantly boosting CTR predictions – up to 90% of bots were spotted and CTR predictions were improved by up to 35% over existing heuristics.

Summary

Analyzing large amounts of data has become a crucial part of the retail and manufacturing business landscape. It has traditionally been done by experts, based on know-how honed through experience. With the power of machine learning, however, it is now possible to combine the astonishing scale of big data with the precision and intelligence of a machine-learning model. While the business community must remain aware of the multiple pitfalls it will face when employing machine learning, that it endows business processes with awesome power and flexibility is now beyond question.

[1] https://www.forbes.com/sites/louiscolumbus/2021/01/17/76-of-enterprises-prioritize-ai–machine-learning-in-2021-it-budgets/?sh=6d24c5e3618a

AI solutions that are designated to bring value in banking

AI solutions are boosting value in banking

/in /by and

According to research done by Business Insider Intelligence, banks will make some $450 billion by 2023 by applying Artificial Intelligence. No wonder, then, that AI is playing an increasingly important role in financial institutions’ roadmap for the coming years – 75% of banks with over $100 billion in assets already have AI strategies in place.

Banks don’t just create AI strategies, but are increasingly using AI and Machine Learning in their day-to-day business. We often work with them on ideation workshops, PoC and solution implementation. As an example, we recently had the opportunity to work with Santander Consumer Bank, running workshops and researching how to use ML to boost the sustainability of loan portfolios. We were able to significantly reduce risk while maintaining the same acceptance rate for extending loans.

Apart from credit risk modelling, there is already an impressive range of use cases for AI in banking, covering everything from customer service to back-office operations. Here is a list of the most common AI solutions in the banking sector: AI solutions in banking sector

Customer service automation

  • Chatbots– applying chatbots to automate customer service increases customer satisfaction. In fact, most simple issues can be solved entirely without human interference. Behind the scenes, automation significantly reduces customer service workloads.
  • Biometric identification enables explicit or unnoticed identity verification within remote channels. This can include voice identity verification in call centers or typing manner verification in online banking.

Customer insights:

  • Customer 360 view – applying deep learning to customer analytics makes it easier to combine insights from various data sources (e.g. transactions, online banking logs, call center interactions). This helps us better understand a bank’s customers and build personalized recommendations and Intelligent customer assistants, making the business more responsive and efficient.
  • Churn prediction – Thanks to accurate AI algorithms, churn probability predictions improve customer retention. This is important as customers often churn without obvious warning signs. Thus, it is difficult to run precisely targeted anti-churn campaigns. On the other hand, retention activities can be expensive, sometimes much more so than the value a potential customer may bring.
  • Customer life-time value is often used to understand how valuable a particular relation is and to optimize other activities – for example, by integrating Customer Lifetime Value with a probability-of-churn function to focus retention activities on the most valuable clients.

Boosting Sales

  • New client acquisition – Deep learning is particularly suited to improving remarketing. As with the customer 360 view, it promotes the use of all possible information about a prospective customer. This includes cookies and how the individual has interacted with a website – from time spent to what they hovered over and how far into the site they went.  Understanding customer behaviour enables a bank to focus marketing activities on potential customers and show them personalized ads, translating into even 2,5x uplift from advertising activities.
  • X-sell – ML techniques can be used to improve the selection of customers targeted for outbound CRM campaigns. They combine the benefits from both the Customer 360 view and advanced probability of purchase predictions. This allows a bank to choose the right customer and the right product to cross-sell. As an example, ML has been shown to improve credit card x-sell by 12,5%.

Credit risk management

  • Loan application assessment – Machine Learning can analyze unstructured data (e.g. transaction descriptions) more thoroughly than other techniques and find non-obvious dependencies. ML techniques can also be combined with traditional scoring models to get even better results.
  • Fraud detection – ML enables nearly fully automated fraud detection, adapting to individual patterns and changing behaviors. It can be applied in areas where a high volume of events needs to be analyzed in real time, e.g. in card payments. AI can find complex correlations, so even the wildest purchase will make sense to AI. Those it can’t wrap its algorithms around will lead to the detection of fraud.
  • Debt collection strategies – AI algorithms can generate a customized communication strategy for each customer. It will adjust the contact channel, recommending script for CC, or propose a schedule.
  • Continuous portfolio evaluation – detecting SME clients, whose risk of default has risen.  This enables banks to react rapidly and start the recovery process before other creditors do.

Back-office optimization

  • Workflow documentation – classifying incoming emails to go to the appropriate department (sales, complaints, support) and customer segmentation (individual, SME, Corporate) reduces the manual work involved with organizing customer service departments.
  • Process automation – including for cash operations, trade finance, credit application processing, accounting processes.

Summary

A wide range of ML and AI applications is increasingly being used to solve real business problems in banking. As AI becomes more popular, those applications will become the market standard.

 

This article was prepared in cooperation with Santander Consumer Bank

Santander Consumer Bank - Logo

Using Machine Learning in credit risk modelling to reduce risk costs

Using machine learning in credit risk modelling

/in , /by

Cost of risk is one of the biggest components in banks’ cost structure. Thus, even a slight improvement in credit risk modelling can translate in huge savings. That’s why machine learning is often implemented in this area.

We would like to share with you some insights from one of our projects, where we applied machine learning to increase credit scoring performance.To illustrate our insights we selected a random pool of 10 000 applications.

How the regular process works

Loan applications are usually assessed through a credit score model, which is most often based on a logistic regression (LR). It is trained on historical data, such as credit history. The model assesses the importance of every attribute provided and translates them into a prediction.

The main limitation of such a model is that it can take into account only linear dependencies between input variables and the predicted variable. On the other hand, it is this very property that makes logistic regression so interpretable.  LR is in widespread used in credit risk modelling.

Credit scoring from a logistic regression model
Credit scoring from a logistic regression model

What machine learning brings to the table

Machine learning enables the utilization of more advanced modeling techniques, such as decision trees and neural networks. This introduces non-linearities to the model and allows to detect more complex dependencies between the attributes. We decided to use an XGBoost model fed with features selected with the use of a method called permutation importance.

Credit scoring from tree-based model
Credit scoring from tree-based model

However, ML models are usually so sophisticated that they are hard to interpret. Since a lack of interpretability would be a serious issue in such a highly regulated field as credit risk assessment, we opted to combine XGBoost and logistic regression.

Combining the models

We used both scoring engines – logistic regression and the ML based one – to assess all of the loan applications

With a clear correlation between the two assessment approaches, a high score in one model would likely mean a high score in the other.

Loan applications assessed by 2 models
Loan applications assessed by 2 models

In the original approach, logistic regression was used to assess applications. The acceptance level was set around 60% and the risk resulted at 1%

Initial credit application split (acceptance to portfolio risk)
Initial credit application split (acceptance to portfolio risk)

If we decrease the threshold by a couple of points, the acceptance level hits 70% while the risk jumps to 1,5%

Credit applications' split after lowering the threshold
Credit applications’ split after lowering the threshold

We next applied a threshold for an ML model, allowing us to get an acceptance percentage to the original level (60%) while bringing the risk down to 0,75% that is by 25% lower than the risk level resulting from only traditional approach.

Credit applications' split after applying Machine Learning
Credit applications’ split after applying Machine Learning

Summary

Machine learning is often seen as difficult to apply in banking due to the sheer amount of regulation the industry faces. The facts don’t necessarily back this up. ML is successfully used in numerous, heavily regulated industries. The example above is just one more example of how. Thanks to this innovative approach it is possible to increase the sustainability of the loans sector and make loans even more affordable to bank customers. There’s nothing artificial about that kind of intelligence.

3D meets AI - an unexplored world of new business opportunities

3D meets AI – an unexplored world of new business opportunities

/in , , /by , and

AI has become a powerful force in computer vision and it has unleashed tangible business opportunities for 2D visual data such as images and videos. Applying AI can bring tremendous results in a number of fields. To learn more about this exciting area, read our overview of 2D computer vision algorithms and applications.

Despite its popularity, there is nothing inherent to 2D imagery that makes it uniquely suitable for AI application. In fact, artificial intelligence systems can analyze various forms of information, including volumetric data. In spite of the increasing number of companies already using 3D data gathered by lidar or 3D cameras, AI applications aren’t the mainstream in their industries.

In this post, we describe how to leverage 3D data across multiple industries with the use of AI. Later in the article we’ll have a closer look at the nuts and bolts of the technology and we’ll aslo show what it takes to apply AI to 3D data. At the end of the post, you’ll also find an interactive demo to play with.

In the 3D world, there is no Swiss Army Knife

3D data is what we call volumetric information. The most common types include:

  • 2.5D data, including information on depth or the distance to visible objects, but no volumetric information of what’s hidden behind them. Lidar data is an example.
  • 3D data, with full volumetric information. Examples include MRI scans or objects rendered with computer graphics.
  • 4D data, where volumetric information is captured as a sequence, and the outcome is a recording where one can go back and forth in time to see the changes occurring in the volume. We refer to this as 3D + time, which we can treat as the 4th dimension. Such representation enables us to visualize and model dynamic 3D processes, which is especially useful in medical applications such as respiratory or cardiac monitoring.

There are also multiple data representations. These include a compound of 2D images along the normal axis, sparse Point Cloud representation and voxelized representation. Such data could have additional channels, like reflectance in every point of a lidar’s view.

Depending on the business need, there can be different objectives for using AI: object detection and classification, semantic segmentation, instance segmentation and movement parameterization, to name a few. Moreover, every setup has its own characteristics and limitations that should be addressed with a dedicated approach (or, in the case of artificial neural networks, with a sophisticated and thoroughly designed architecture). These are the main reasons our clients come to us, and to take advantage of our experience in the field. We are responsible for delivering the AI part of specific projects, even though the majority of their competencies are built in-house.

Let us have a closer look at a few examples

1. Autonomous driving

  • Task: 3D object detection and classification,
  • Data: 2.5 Point clouds captured with a lidar: sparse data, big distances between points

Autonomous driving data are very sparse because:

  • the distances between objects in outdoor environments are significant
  • In the majority of cases lidar rays from the front and rear of the car don’t return to lidar, since there are no objects to reflect them.
  • The resolution of objects gets worse the further they are from the laser scanner. Due to the angular expansion of the beam it’s impossible to determine the precise shape of objects that are far away.

For autonomous driving, we needed a system that can take advantage of data sparsity to infer 3D bounding boxes around objects. One such network is the part-aware and aggregation neural network i.e. Part-A2 net (https://arxiv.org/abs/1907.03670). This is a two-stage network that uses the high separability of objects, which functions as segmentation information.

In the first stage, the network estimates the position of foreground points of objects inside bounding boxes generated by an anchor-based or anchor-free scheme. Then, in the second stage, the network aggregates local information for box refinement and class estimation. The network output is shown below, with the colors of points in bounding boxes showing their relative location as perceived by the Part-A2 net.

3D meets AI - Autonomous drivingSource of image: From Points to Parts: 3D Object Detection from Point Cloud with Part-aware and Part-aggregation Network

2. Indoor scene mapping

  • Task: Object instance segmentation
  • Data: Point clouds, sparse data, relatively small distances between points

A different setup is called for in mapping indoor environments, such as we do with instance segmentation of objects in office space or shops (see this dataset for better intuition: S3DIS dataset). Here we employ a relatively high-density representation of a point cloud and BoNet architecture.

In this case the space is divided into a 1- x 1- x 1-meter cubic grid. In each cube, a few thousand points are sampled for further processing. In an autonomous driving scenario, such a grid division would make little sense given the sheer number of cubes produced, many of which are empty and only a few of which contain any relevant information.

The network produces semantic segmentation masks as well as bounding boxes. The inference is a two-stage process. The first produces a global feature vector to predict a fixed number of bounding boxes. It also tallies scores to indicate whether some of the predicted classes are inside those boxes. The point-level and global features derived in the first stage are then used to predict a point-level binary mask with the class assignment. The pictures below show a typical scene with the segmentation masks.

3D meets AI - Indoor scene mapping
An example from the S3DIS dataset. From left: input image, semantic segmentation labels, instance segmentation labels

3. Medical diagnosis

  • Task: 3D Semantic segmentation
  • Data: Stacked 2D images, dense data, small distance between images

This is a highly controlled setup, where all 2D images are carefully and densely stacked together. Such a representation can be treated as a natural extension of a 2D setup. In such cases, modifying existing 2D approaches will deliver satisfactory results.

An example of a modified 2D approach is the 3D U-Net (https://arxiv.org/abs/1606.06650), where all 2D operations for a classical U-Net are replaced by their 3D counterparts. If you want to know more about AI in medicine, check out how it can be used to help with COVID-19 diagnosis and other challenges.

3D meets AI - Medical diagnosis
Source: Head CT scan

4. A 3D-enhanced 2D approach

There is also another case, where luckily, it can be relatively straightforward to apply expertise and technology developed for 2D cases in 3D applications. One such scenario is where there are 2D labels available, but the data and the inference products are in 3D. Another is when 3D information can play a supportive role.

In such a case, a depth map produced by 3D cameras can be treated as an additional image channel beyond regular RGB colors. Such additional information increases the sensitivity of neural networks to edge detection and thus yield better object boundaries.

3D meets AI - A 3D-enhanced 2D approach
Source: Azure Kinect DK depth camera

Examples of the projects we have delivered in such a setup include:

  • Defect detection based on 2D and 3D images.

We developed an AI system for a tire manufacturer to detect diverse types of defects. 3D data played a crucial role as it allowed for ultra-precise detection of submillimeter-size bubbles and scratches.

  • Object detection in a factory

We designed a system to detect and segment industrial assets in a chemical facility that had been thoroughly scanned with high resolution laser scanners. Combining 2D and 3D information allowed us to digitize the topology of the installation and its pipe system.

3D data needs a mix of competencies

At deepsense.ai, we have a team of data scientists and software engineers handling the algorithmic, visualization, and integration capabilities. Our teams are set up to flexibly adapt to specific business cases and provide tailor-made AI solutions. The solutions they produce are an alternative to pre-made, off-the-shelf products, which often prove too rigid and constrained; they fail once user expectations deviate from the assumptions of their designers.

Processing and visualizing data in near real time with appropriate user experience is no piece of cake. Doing so requires a tough balancing act, including

combining specific business needs, technical limitations resulting from huge data loads and the need to support multiple platforms.

It is always easier to discuss based on an example. Next section shows what it takes to develop an object detection system for autonomous vehicles with outputs accessible from a web browser. The goal is to predict bounding boxes of 3 different classes: car, pedestrian and cyclist, 360 degrees around the car. Such a project can be divided into 4 interconnected components: data processing, algorithms, visualizations and deployment.

Data preprocessing

In our example, we use the KITTI and A2D2 datasets, two common datasets for autonomous driving, and ones our R&D hub rely on heavily. In both datasets, we use data from spinning lidars for inference and cameras for visualization purposes.

Lidars and cameras work independently, capturing data at different rates. To obtain a full picture, all data have to be mapped to a common coordinate system and adjusted for time. This is no easy task. As lidars are constantly spinning, each point is captured at a different time, while simultaneously the position and rotation of the car in relation to world coordinates is changing. Meanwhile, the precise location and angle of the car is not known perfectly due to limitations of geolocation systems such as GPS. These difficulties make it extremely difficult to precisely and stably determine the absolute positions of objects around you (SLAM can be used to tackle some of the problems).

Fortunately, absolute positioning of objects around the vehicle is not always required.

Algorithms

There are a vast number of approaches when it comes to 3D data. However, factors such as the length to and between objects and high sparsity will play an essential role in which algorithm we ultimately settle on. As in the first example above, we used Part-A2 net.

Deployment

We have relied on a complete, in-house solution for visualization, data handling, and UI. We have used expertise in the Unity engine to develop a cross-platform, graphically rich and fully flexible solution. In terms of a platform, we opted for maximum availability, which can be satisfied by a popular web browser like Chrome or Mozilla and WebGL as Unity’s compilation platform.

Visualization/UI

WebGL, while very comfortable for the user, disables drive access and advanced GPU features, limits available RAM to 2GB and processing to a single thread. Additionally, while standalone solutions in Unity may rely on existing libraries for point cloud visualization, making it possible to visualize hundreds of millions of points (thanks to advanced GPU features), this is not the case in WebGL.

Therefore, we have developed an in-house visualization solution enabling real-time, in-browser visualization of up to 70 mln points. Give it a try!

Such visualization could be tailored to the company’s specific needs. In a recent project, we took a different approach: we used AR glasses in visualizing a factory in all its complexity. This enabled our client to reach next level user experience and see the factory in a whole new light.

Summary

We hope that this post has shed some light on how AI can be used with 3D data. If you have a particular 3D use case in mind or you are just curious about the potential for AI solutions in your field, please reach out to us. We’ll be happy to share our experience and discuss potential ways we can help you apply the power of artificial intelligence in your business. Please drop us an email at contact@deepsense.ai.

Smart alerting with AIOps

Smart alerting with AIOps

/in , /by and

A trigger-based alerting in infrastructure management is currently insufficient to keep operations smooth and uninterrupted. AI-powered IT operations deliver smart alerts that not only improve a team’s ability to respond swiftly but can also deliver significant automation. 

According to Gartner data quoted by the TechBeacon, 40% of organizations are predicted to strategically implement AIOps to enhance performance monitoring by 2022. The reason is obvious: AIOps helps build resilient and reliable IT corporate ecosystems. It also boosts overall company performance, an area that can always be improved. For example, for 34.4% of organizations it takes more than 30 minutes on average to resolve IT incidents impacting consumer-facing digital services.

The standard approach to alerting

In the traditional approach to automating IT operations, sets of previously defined conditions triggered alerts.

A huge drawback of this approach was that problems had to be identified, defined and researched. Rules were then set to react to the problem when it occurred. This approach was effective only in dealing with problems that had already come up, but it was useless when conditions changed. That’s common in IT infrastructure.

Also, manually setting the rules opens the process up to mistakes hidden within the code and rules – the more complicated the system or instruction, the better hidden a malfunction can be. And there it lurks, ready to hit right when the rule should have saved the infrastructure.

This is where AI tools come in handy.

How AIOps makes basic alerting smart alerting

The key difference between trigger-based alerts and AI-powered alerts is that the first comes with a huge amount of manual work, while the second can cut workloads significantly by automating the most tedious tasks (read more about that in our What is AIOps blog post).

Both types of alerts send a constant flow of data in from the system, including logs, events or metrics collected by sensors. However, only in the AI alerts is the data preprocessed and initially analyzed before a human gets it–stripped of the noise and chaos usually seen in the raw data.

Pattern recognition-based alerting

Unlike in less sophisticated systems, AI-powered solutions deliver pattern-based, rather than triggered, alerting. The core difference is in recognizing not a particular, single event or a chain of events that occur within the network, but the overall pattern of a malfunction or attack. Thus, if there is a new channel or technology used, the system will remain vigilant. Read more about that in our AIOps for Network Traffic Analysis (NTA) blog post.

Also, the pattern-based approach not only helps spot known types of attacks but also can be useful when approaching the unknown. Harnessing the power of machine learning or reinforcement learning paradigms of supervised learning and reinforcement learning allows the neural network to learn how IT infrastructure fits into the daily work. This essentially makes it the direct opposite of a  trigger-based system. Instead of spotting the signs of malicious activity, machine learning models ensure operations can detect anomalies within the system.

Policy-driven escalation and remediation

Smart alerting is a great tool for DevOps or administration specialists, though it remains only one step of their work. Any alert is a problem that needs to be fixed or at least a condition that requires supervision. In keeping IT infrastructure up and running as well, notification remains only a single step. Services lost even briefly can be very costly.

With AIOps implemented, ml-armed supervisors will not only spot the signs of a malfunction but also respond with the best policy available, according to their previous experience and information implemented in the network.

A simple example would be delivering a mirrored piece of infrastructure and redirecting the traffic while providing an alert to the supervisory team. AIOps handles this with alert escalation workflows and immediate response. Also, the AIOps infrastructure correlates the dependencies based on incoming data and responds not to a single event, but to address the group of alerts continuously. The approach saves the time usually spent grouping alerts into same-event related ones.

Key benefits of AIOps alerting

AIOps solutions have been adopted for their numerous benefits, including the following four.

  • Savings – any anomalous situation in the network can result in costs skyrocketing, especially in the age of growing cloud workloads. When the company is charged for every byte of information stored or transferred, inefficiencies or instructions add up to unbearable costs, especially in the long term.
  • Security – alerting delivers knowledge about anomalous situations in the system, and any anomaly can be a sign of an ongoing intrusion or malfunction that can end up shutting the system down.
  • Scalability – smart alerting systems deliver more clear information about the overall state of the infrastructure. Repeating patterns in network performance are a sign to up- or downscale the entire system, be it once, with on-demand cloud solutions, or permanently  by adding new components.
  • Smart cloud management – delivering a real-time alert is crucial to initiating next steps; in an AIOps alert platform, these can be automated and launched contextually, as the current situation demands, not only according to predefined conditions.

Summary

Smart alerting is the foundation of next-level infrastructure management. Companies can harness the power of AI-powered solutions to optimize operations and deliver better results.

If you’d like to hear more about deepsense.ai’s AIOps platform and our approach, contact us using the form below or drop us a line at aiops@deepsense.ai.

AIOps Network Traffic Analysis (NTA) - a business guide

AIOps Network Traffic Analysis (NTA) – a business guide

/in , /by

Network Traffic Analysis (NTA) is a key component of modern cybersecurity in companies. With machine learning and artificial intelligence solutions, the sheer amounts of data to analyze is an asset to be used rather than, as was once the case, a challenge to overcome.

This post looks at:

According to Markets and Markets data, the global network traffic monitoring software and traffic analysis tool market is projected to grow from $1.9 billion in 2019 to  $3.2 billion by 2024. The growth is driven mostly by the increasing demand for sophisticated network monitoring tools and advanced network management systems that can handle the growing traffic and increasing flow of information.

The growth in internal traffic is a direct reflection of global trends. According to Cisco data, nearly 66% of the global population will be online by 2023. The increase in traffic is driven not only by users but also by the myriad of connected devices that form the IoT cloud around us.

The share of Machine-to-Machine (M2M) connections is estimated to grow from 33% in 2018 to 50% in 2023, while the consumer segment will rise to 74% of this share and business segment for 26%.

What is network traffic analysis

In its most basic form, Network Traffic Analysis (NTA) is the process of recording and analyzing network traffic patterns in search of suspicious elements and security threats. The term was originally coined by Gartner to describe a growing industry in the computer security ecosystem.

The foundation of NTA is the assumption that there is a “normal” situation in the system that reflects daily operations. Due to seasonal or general trends, operations fluctuate naturally, but overall the system remains stable and thus internal network monitoring can be done with a traffic analyzer. Knowing the “normal” situation is the first step in spotting signs of malicious activities within the system.

In addition to spotting security threats, NTA is also used to optimize the system, spotting inefficiencies as well as the system’s need for additional components when it arises.

Network Traffic Analysis software tools analyze a system’s communication flow, including

  • TCP/UDP packets
  • “Virtual network traffic” done in virtual private networks
  • Traffic to and from cloud environments  (storage, computing power, etc.)
  • API calls to cloud-based apps or SaaS solutions.

This means that nearly all traffic and information flow can be tracked and analyzed by smart network traffic analysis solutions. Modern solutions often use sophisticated techniques like reinforcement learning.

A key component of network analytics tools is the dashboard used to interface with the team, which receives clear information about the network. The dashboard enables easier network performance monitoring and diagnostics and is a convenient way to convey technical knowledge to those who lack it. Reducing complexity to simplicity, the dashboard will

Play its part in convincing your financial director to spring for a powerful new server or another essential component.

NTA solutions are clearly sophisticated and powerful tools. But what are the direct benefits of network traffic analysis?

The benefits of network traffic analysis

There are at least several benefits:

  • Avoiding bandwidth and server performance bottlenecks – Armed with knowledge about how information flows in the system, one can analyze network problems, define problems and start looking for solutions.
  • Discovering apps that gobble up bandwidth – tweaking the system can deliver significant savings when API calls are reduced or information is reused.
  • Proactively reacting to a changing environment – a key feature when it comes to delivering high-quality services for clients and customers. The company can react to increasing demand or spot signs of an approaching peak to harden the network against it. Advanced network traffic analysis tools are often armed with solutions designed to respond in real-time to network changes much faster than any administrator would.
  • Managing devices exclusively – with modern network monitoring applications companies can group devices and network components to manage them, effectively making use of network performance analytics done earlier.
  • Resource usage optimization – With all apps, devices, components, and traffic pinpointed with a dashboard, the company can make more informed decisions about the system’s resources and costs.

The key challenge in computer network management is processing and analyzing the gargantuan amounts of data networks produce. Looking for the proverbial needle in the haystack is an apt metaphor for searching for insights among the data mined from a network.

Using ML tools is the only way to effectively monitor network traffic.

How machine learning can support traffic analysis

The key breakthrough that comes from using machine learning-powered tools in NTA is in automation. The lion’s share of the dull and repetitive yet necessary work is done by machines. Also, in real-time network analysis, time is another component that can be handled only by machines. Machines and neural networks can spot and analyze the hidden patterns in data to deliver a range of advantages for companies. To name just a few:

Intrusion detection

The first and sometimes the only sign of intrusion into a system is suspicious traffic that can be easily overlooked. Intrusions are often detected only after 14 days.

AI-based solutions are tireless, analyzing traffic in real-time. Armed with the knowledge of infrastructure operations, the system can spot any sign of malicious activity.

Reducing false positives

AI-based solutions are less prone to the false-positives that can turn the life of a system administrator into a living hell. AI-based systems significantly enrich ML-supported NTA with false-positive detection and reduction, enabling the team to focus more on real challenges than on verifying every alert.

Workload prediction

With data about ongoing system performance, the solution can deliver information about predicted traffic peaks or downs to optimize spending.

Thus the benefits are twofold. First, the company can manage the costs of infrastructure, be it cloud or on-prem, to handle the real traffic and avoid overpaying. Second, there is much more predictability in the estimated need for resources, so they can be booked in advance or the costs can be optimized in other ways.

Spotting early signs of attack (DDoS)

Distributed Denial of Service attacks attempts to suddenly overload a company’s resources in an effort to take down the website or other online service. The losses are hard to predict – from the company’s reputation being hit as unable to defend itself against cybercrime attacks, to the staggering and quickly accruing losses due to being unavailable for customers.

With the early information about incoming attacks, the company can set up defenses like blocking certain traffic, ports or locations to keep availability on other markets. Also, network traffic reports can be used by various agencies that fight cybercrime and will hunt for those responsible for the attack.

Malicious packet detection

Sometimes it is not about the intrusion and the malicious activity is not aimed directly at the company. A user could have downloaded malware onto a private device connected with an enterprise network via a VPN. With that, the infection can spread or the software itself can leverage the company’s resources, such as computing power, and use it for its own purposes, like mining cryptocurrency without the owner’s consent.

Summary

Network traffic monitoring and analysis is one of the key components of modern enterprise-focuses cybersecurity. The gargantuan amounts of data to process also make it a perfect foundation for ML-based solutions, which thrive on data.

That’s why deepsense.ai delivers a comprehensive AIOps architecture-based platform for network data analytics.

If you have any questions about the AIOps solutions we provide, don’t hesitate to contact Andy Thurai, our Head of US operations via the contact form or aiops@deepsense.ai email address.

Five AI trends 2020 to keep an eye on

Five AI trends 2020 to keep an eye on

/in , , /by

While making predictions may be easy, delivering accurate ones is an altogether different story. That’s why in this column we won’t just be looking at the most important trends of 2020, but we’ll also look at how the ideas we highlighted last year have developed.

In summarizing the trends of 2020, one conclusion we’ve come to is that society is getting increasingly interested in AI technology, in terms of both the threats it poses and common knowledge about other problems that need to be addressed.

AI trends 2019 in review – how accurate were our predictions?

AI trends 2020 - AI trends 2019 in review – how accurate were our predictions?
In our AI Trends 2019 blogpost we chronicled last year’s most important trends and directions of development to watch. It was shortly after launching the AI Monthly Digest, a monthly summary of the most significant and exciting machine learning news. Here’s a short summary of what we were right and wrong about in our predictions.

  • Chatbots and virtual assistants –  powered by a focus on the development of Natural Language Processing (NLP), our prediction was accurate–the growth in this market would be robust. The chatbot market was worth $2.6 billion in 2019 and is predicted to reach up to $9.4 billion by 2024.
  • the time needed for training would fall – the trend gets reflected by larger neural networks being trained in a feasible time, with GPT-2 being the best example.
  • Autonomous vehicles are on the rise – the best proof is in our own contribution to the matter in a joint-venture with Volkswagen.
  • Machine learning and artificial intelligence are being democratized and productionized – According to Gartner, 37% of organizations have implemented AI in some form. That’s a 270% increase over the last four years.
  • AI and ML responsibility and transparency – the trend encompasses the delivering unbiased models and tools. The story of Amazon using an AI-based recruiting tool that turned out to be biased against female applicants made enough waves to highlight the need for further human control and supervision over automated solutions.

Apparently, deepsense.ai’s data science team was up to date and well-informed on these matters.

“It is difficult to make predictions, especially about the future.”
-Niels Bohr

The world is far from slowing down and Artificial Intelligence (AI) appears to be one of the most dominant technologies at work today. The demand for AI talents has doubled in the last two years with technology and the financial sector absorbing 60% of talented employees on the market.

The Artificial Intelligence market itself is predicted to reach $390.9 billion by 2025, mainly by primarily by automating dull and repetitive tasks. It is predicted that AI will resolve around 20% of unmet healthcare demands.

Considering the impact of AI on people’s daily lives, spotting the right trends to follow is even more important. AI is arguably the most important technology trend of 2020, so enjoy our list!

Natural language processing (NLP) – further development

AI trends 2020 - Natural language processing (NLP) – further development
Whether the world was ready for it or not, GPT-2 was released last year, with balance between safety and progress a guiding motif. Initially, OpenAI refused to make the model and dataset public due to the risk of the technology being used for malicious ends.

The organization released versions of the model throughout 2019, with each confirmed to be “hardened against malicious usage”. The model was considered cutting edge, though like most things in tech, another force soon prevailed. At the end of January 2020, Google Brain took the wraps off of Meena, a 2.6-billion parameter end-to-end neural conversational model trained on 341 GB of online text.

The convenience of NLP solutions is enjoyed by users who have embraced virtual assistants like Google Assistant, Alexa or Siri. According to Adroit Market Research, the market of Intelligent Virtual Assistants is predicted to grow at 33% compound annual growth rate between now and 2025. The market was valued at $2.1 billion in 2019. The increasing use of smartphones and other wearable intelligent devices, among other trends, is predicted to be a driver of the growth.

Started with a consumer-centric approach, virtual assistants are predicted to get more involved in business operations, further automating processes as well as tedious and repetitive tasks. According to Computerworld, approximately 40% of business representatives are planning to implement voice technology within 24 months – that is, no later than in 2021. NLP is shaping up to be a major trend not only this year, but well into the future.

Autonomous vehicles

AI trends 2020 - Autonomous vehicles

It is 2020 and driverless cars have yet to hit the streets. In hindsight, the Guardian’s prediction that there would be 10 million self-driving cars on the road by 2020 is all too easy to scoff at now.

On the other hand, tremendous progress has been made and with every month the autonomous car gets closer to rolling out.

deepsense.ai has also contributed to the progress, cooperating with Volkswagen on building a reinforcement learning-based model that, when transferred from a simulated to a real environment, managed to safely drive a car.

But deepsense.ai is far from being the only company bringing significant research about autonomous cars and developing the technology in this field. Also, it is a great difference between seeing an autonomous car on busy city streets and in the slightly less demanding highway environment, where we can expect the automation and semi-automation of driving to first get done.

According to the US Department of Transportation, 63.3% of the $1,139 billion of goods shipped in 2017 were moved on roads. Had autonomous vehicles been enlisted to do the hauling, the transport could have been organized more efficiently, and the need for human effort vastly diminished. Machines can drive for hours without losing concentration. Road freight is globally the largest producer of emissions and consumes more than 70% of all energy used for freight. Every optimization made to fuel usage and routes will improve both energy and time management.

AI getting popular – beneath the surface

AI trends 2020 - AI getting popular – beneath the surface
There is a lot of buzz around how AI-powered solutions impact our daily lives. While the most obvious change may be NLP powering virtual assistants like Google Assistant, Siri or Alexa, the impact on our daily lives runs much deeper, even if it’s not all that visible at first glance. Artificial intelligence-powered solutions have a strong influence on manufacturing, impacting prices and supply chains of goods.

Here are a few applications being used without batting an eye:

  • Demand forecasting – companies collect tremendous amounts of data on their customer relationships and transactional history. Also, with the e-commerce revolution humming along, retail companies have gained access to gargantuan amounts of data about customer service, products and services. deepsense.ai delivers demand forecasting tools that not only process such data but also combines it with external sources to deliver more accurate predictions than standard heuristics.  Helping companies avoid overstocking while continuing to satisfy demand is one essential benefit demand forecasting promises.
  • Quality control – harnessing the power of image recognition enables companies to deliver more accurate and reliable quality control automation tools. Because machines are domain-agnostic, the tools can be applied in various businesses, from fashion to construction to manufacturing. Any product that can be controlled using human sight can also be placed under the supervision of computer vision-powered tools.
  • Manufacturing processes optimization – The big data revolution impacts all businesses, but with IoT and the building of intelligent solutions, companies get access to even more data to process. But it is not about gathering and endless processing in search of insights – the data is also the fuel for optimization, sometimes in surprising ways. Thanks solely to optimization, Google reduced its cooling bill by 40% without adding any new components to its system. Beyond cutting costs, companies also use process optimization to boost employee safety and reduce the number of accidents.
  • Office processes optimization – AI-powered tools can also be used to augment the daily tasks done by various specialists, including lawyers or journalists. Ernst & Young is using an NLP tool to review contracts, enabling their specialists to use their time more efficiently. Reuters, a global media corporation and press agency, uses AI-powered video transcription tools to deliver time-coded speech-to-text tools that are compatible with 11 languages.

Thanks to the versatility and flexibility of such AI-powered solutions, business applications are possible even in the most surprising industries and companies. So even if a person were to completely abandon technology (right…), the services and products delivered to them would still be produced or augmented with AI, be they clothing, food or furniture.

AI getting mainstream in culture and society

AI trends 2020 - AI getting mainstream in culture and society
The motif of AI is prevalent in the arts, though usually not in a good way. Isaac Asimov was among the first writers to hold that autonomous robots would need to follow a moral code in order not to become dangerous to humans. Of course, literature has offered a number of memorable examples of AI run amok, including Terminator and HAL 9000 from Space Odyssey.

The question of moral principles may once have been elusive and abstract, but autonomous cars have necessitated a legal framework ascribing responsibility for accidents. Amazon learned about the need to control AI models the hard way, albeit in a less mobile environment: a recruiting tool the company was using had to be scrapped due to a bias against women.

The impact of AI applications on people’s daily lives, choices and careers is building pressure to deliver legal regulations on model transparency as well as information not only about outcomes, but also the reasons behind them. Delivering AI in a black-box mode is not the most suitable way to operate, especially as the number of decisions made automatically by AI-powered solutions increases.

Automating the development of AI

AI trends 2020 - Automating the development of AI
Making AI mainstream is not only about making AI systems more common, but widening the availability of AI tools and their accessibility to less-skilled individuals. The number of models delivering solutions to power with the machine and deep learning will only increase.It should therefore come as no surprise that the people responsible for automating others’ jobs are keen to support their own jobs with automation.

Google enters the field with AutoML, a tool that simplifies the process of developing AI and making it available for a wider audience, one that, presumably, is not going to use ML algorithms in some especially non-standard ways. AutoML joins IBM’s autoAI, which supports data preparation.

Also, there are targeted cloud offerings for companies seeking to harness ready-to-use components in their daily jobs with a view to augmenting their standard procedures with machine learning.

Summary

While the 2020 AI Trends themselves are similar to those of 2019, the details have changed immensely, thus refreshing our perspective seemed worth our while. The world is changing, ML is advancing, and AI is ever more ubiquitous  in our daily lives.

What is AIOps - AI for IT operations explained

What is AIOps – AI for IT operations explained

/in /by

Every business now depends on IT. Efficient IT Operations is mandatory for all businesses, especially those operating in a hybrid mode – a mix of existing data centers and multi-cloud locations.

As with any business process, IT operations can be augmented with machine learning-based solutions. IT is particularly fertile ground for AI as it is mostly digital, has seemingly endless processes requiring automation and there are gigantic amounts of data to process.

IT Operations are expensive!

According to Research and Markets data, global IT Ops & service management (ITSM) is predicted to reach $35.98 billion by 2025 with an annual growth rate of 7.5% YoY.

As the importance of IT operations has ramped up, so has the pressure on ITOps teams. A range of issues puts pressure on teams: shrinking budget for IT operations, multi cloud-based applications, dynamic scaling of infrastructure, limited availability of experienced ITOps personnel, the constant threat from outsiders given the nature of cloud applications, the extension of applications to edge locations with IoT and mobile devices.

AIOps is here to support the maintenance teams and provide AIOps tools to solve problems once thought unsolvable.

What is AIOps

AIOps supports infrastructure management and operations with AI-based solutions. It is employed mainly to automate tasks, improve process efficiency, quicken reactions — sometimes even to a real-time response rate — and deliver accurate predictions on upcoming events.

The big data revolution and machine learning technology have driven change, making it possible to process the vast amounts of information IT infrastructure generates. AI can solve the following challenges:

  • Anomaly detection – despite fluctuations and the dynamic nature of data, the internal infrastructure ecosystem is a stable environment. Thus, any anomaly can signal the existence of a problem. Also, early detection of an anomaly is usually a sign of a problem that has yet to be fully understood.
  • Event consolidation – An AI model can simplify huge amounts of data, dividing it into multiple layers and finding insights.
  • Service tickets analytics – when fed data on tickets submitted to a service desk, an ML-based model can predict seasonal spikes and requests. This can help the service desk owner deploy help desk personnel s needed.
  • Detecting seasonality and trends – when using an AI-powered solution, any trend can be divided into 3 components – seasonality, trend and residual. That increases the predictability of long-term commitments and makes managing them more effective.
  • Frequent pattern mining – machine-powered analysis delivers insights that are beyond the reach of humans. Machines not only process more data but also , unlike humans, make unbiased decisions. They also find correlations that are impossible for humans to detect.
  • Time series forecasting – AI-based models can forecast future values such as memory load, network throughput ticket count or other values in the future. This enables AIOps solutions to deliver early alert predictions.
  • Noise reduction – AIOps solutions eliminate noise and concentrate on the real underlying problems.

AI helps ITOps run smoother

There are currently several major challenges for IT departments.

Fraud Detection/Security

According to IBM data quoted by CSO, the average time to identify a breach in 2019 was 209 days. Such a sizable delay is caused mainly by security teams being overwhelmed with work and the stealth operations of criminals. Cybercrime is a highly profitable venture, with profits reaching up to $1.5 trillion a year. Cybercriminals don’t play favorites, targeting victims of all stripes, from individuals to international corporations. In May 2019 authorities tracked down a group that had availed itself of an estimated $100 million.

Anomaly detecting AI and machine learning-based AIOps solutions can spot even the slightest signs that unexpected events are occurring in a system. AIOps can be trained to learn what “typical operations” look like and spot anything out of place. It can also send real-time notifications to the team.

Eliminate Downtime

According to an ITIC study, for 86% of companies surveyed, a single hour of downtime costs $300,000. For 34%, the cost comes in at a staggering $1 million.

AIOps comes with various tools to help keep the lights on and operations running smoothly. In addition to anomaly detection, time series prediction serves as a benchmark and a tool for designing maintenance flow. It also supports efficient resource management. Pattern mining spots inefficient components and bottlenecks to be optimized. It also enables the mapping of both seasonality and trends, so resources supporting operations can be assigned efficiently.

Capacity planning

Before the cloud, companies were forced to overpay for servers and computing power because they had to stay on top of fluctuations in their seasonal needs for computer power. Today, despite the access they have to endless power and storage in the cloud, IT teams around the world continue to struggle with capacity planning and delivering scalable infrastructure to meet the irregular demands for infrastructure.

Nearly all AIOps functionalities support the goal of delivering a stable and scalable environment. With capacity planning supported by time series forecasting and ticket analysis, IT teams can manage their infrastructure scaling and maintenance not only to avoid downtime but also to minimize costs and utilize their systems as efficiently as possible.

A great example comes from Google, whose AI-based system delivered new operational efficiency recommendations for data center cooling systems, effectively cutting costs by 40%.

Noise reduction and pattern mining deliver clear insights.  On the other hand, scraping through the data in real-time enables an AIOps platform to deliver insights faster and using those insights more actionably.

Summary

AIOps machine learning-powered solutions can significantly improve today’s data-heavy IT infrastructure management.

A good way to learn more about the AIOps strategy landscape is to meet our team during the upcoming AIOps conference in Fort Lauderdale, Florida. Our specialists will be more than happy to strategize, assist, answer your questions, and share their expertise from the field. If you’d like to meet us there, just drop us a line or stop by our booth (#1104). Contact us at AIOps@deepsense.AI to see how we can help!