Get Started

Let’s face it, though: nobody likes the constraints imposed by on-prem. Having to talk your IT  department into setting an in-house deployment project, the cost of installation, then cost of running it, potential scalability issues... Sounds familiar?

So far our solution has been hybrid hosting: cloud-based Kili with data hosted on-prem. Although this compromise solution does the job for many of our clients, it also comes with a bunch of constraints. And, worse still, doesn’t meet the strictest security and privacy policies.

This is why we decided to offer something completely different: how about deploying Kili directly in your Microsoft Azure premise? You keep all the advantages of on-prem: the platform is not shared with anyone and only you control what goes in and out. What’s more, you get to use all the commodities that cloud marketplaces are offering: automatic deployments, scalability by design, and automatic updates. At the same time, you don’t have any of the disadvantages of the traditional on-prem.

Last but not least, you can be set and ready to go in days, rather than weeks (a well known problem for traditional on-prem). Is it the best of both worlds? We think so.

Contact our Sales team to learn more about purchasing 

Kili is now available on Microsoft Azure Marketplace

Kili Technology is now available on Microsoft Azure Marketplace

To be able to train a model using YOLO v7, you have to create a data.yaml file that indicates the image and label data layout and the classes that you want to detect. You also have to organize your data accordingly. Kili CLI will help you bootstrap this step, and does not require a project-specific setup.

 file, from the repository (example: the COCO 2017 dataset).

# COCO 2017 dataset http://cocodataset.org
 
# download command/URL (optional)
download: bash ./scripts/get_coco.sh
 
# train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/]
train: ./coco/train2017.txt  # 118287 images
val: ./coco/val2017.txt  # 5000 images
test: ./coco/test-dev2017.txt  # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794
 
# number of classes
nc: 80
 
# class names
names: [ 'person', 'bicycle’, …]

Let’s go through the steps of creating a .yaml file so you can use the Kili data to train a YOLO v7 project.

You can integrate the following steps into a kili_to_yolov7.sh file, or play along with this 

Let’s suppose that you have an object detection project in Kili. For example, we loaded these images from the 

First, export the annotations with the following command:

kili project export  --project-id $KILI_PROJECT_ID --verbose --output-format
yolo_v7 --output-file your_export.zip

mkdir my_dataset/

unzip -q your_export.zip -d my_dataset/

This will create a dataset folder with the following subfolders:

my_dataset

├── README.kili.txt

├── data.yaml

├── images

└── labels

 datasets that are needed to make yolo_v7 happy, with a random strategy

The following command randomly dispatches each asset path into one file among 

find `pwd`/my_dataset/images -name "*.jpg" | awk '{v=rand();if(v<0.5) {print >
"train.txt"} else if (v>0.8) {print > "test.txt"} else {print > "val.txt"}}'

And finally, you can add the dataset files to the 

echo "train: /path/to/my_dataset/train.txt" >> /path/to/my_dataset/data.yaml
echo "val: /path/to/my_dataset/val.txt" >> /path/to/my_dataset/data.yaml
echo "test: /path/to/my_dataset/test.txt" >> /path/to/my_dataset/data.yaml


names: ['BICYCLE']

train: /path/to/my_dataset/train.txt

val: /path/to/my_dataset/val.txt

test: /path/to/my_dataset/test.txt

To use YOLO v7, you first need to install the YOLO v7 repository following these 

, and make sure to download the initial model weights with:

wget https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7.pt

We are now ready to use Yolov7! First, you can run Yolo V7 training using:

python train.py --workers 8  --batch-size 16 --data /path/to/your/data.yaml --img 320 320 --cfg cfg/training/yolov7-tiny.yaml --weights yolov7.pt --name yolov7 --hyp data/hyp.scratch.tiny.yaml 

Once done, you can import the predictions back into Kili. This can help you diagnose labeling or model errors, or serve as a source of pre-annotations.

First, run the YOLO v7 detection using the model you have generated in the previous step:

python detect.py --weights path/to/the/trained/model/best.pt --conf 0.25 --img-size 640 --source /content/my_dataset/labels/ --save-txt --project path/where/to/store/labels

Then import the results into Kili. You need to specify a 

 of your choice to be able to identify the predictions, and the 

 in your original project ontology (here, 

kili project label path/where/to/store/labels/exp/labels/ --project-id $KILI_PROJECT_ID --prediction --model-name my-YOLO-v7 --metadata-file my_dataset/data.yaml --target-job JOB_0 --input-format yolo_v7 --api-key=$KILI_API_KEY

Now you can visualize the predictions in the Kili interface!

Note that you can also specify the project id of another project, provided that it has the same ontology as the original project. This way, you can use the model-generated annotations to bootstrap human annotations. A human annotator will only have to validate or correct the model’s predictions. If these predictions are accurate enough, the ratio validation/correction will be high, saving you a great deal of annotation time.

In this tutorial, we showed you how to export the data of an object detection project created in Kili to the Yolo V7 format, train a model with Yolo V7, and then push back the predictions to a Kili project. Note that you can also do the same operations with Yolo V4 and V5 compatible models; simply change the label format in Kili import and export! 

An article by Pierre Leveau, Lead Machine Engineer at Kili Technology

Using Kili Technology to work with YOLO v7

Every year, millions of tons of plastic spill from our rivers to our oceans. A portion of this plastic travels to ocean garbage patches, where it gets caught in circulating currents. These garbage patches are a heavy burden to our oceanic life and if no action is taken, plastic will increasingly impact our ecosystems, health, and economies.

The Ocean Cleanup’s mission is to close the sources of plastic pollution & clean up what has already accumulated in the ocean in order to remove 90 % of floating ocean plastic. To do that, The Ocean Cleanup relies on technology to identify, classify & locate the garbage patches & their content.

The current datasets of The Ocean Cleanup were built using conventional methods (trawls) that are very labor-intensive, or less conventional methods (airplane) that are very costly and complex to organize. The team has worked on an AI object detection software, combined with automated time-lapse image series along GPS-tagged transects. This creates a remote sensing approach to detect and map the dynamic behavior of floating ocean plastic more efficiently. Ultimately, this growing dataset will help The Ocean Cleanup determine where to deploy cleanup in an extensive area with an uneven distribution of plastic debris.

Deep dive about Ocean Cleanup’s AI research in this 

To help The Ocean Cleanup in their mission, we are organising a data-science challenge during which we’re mobilising hundreds of data scientists to annotate a massive dataset of polluted ocean photographs. 

 participants in the challenge will annotate images from The Ocean Cleanup’s on-board cameras, classifying the detected objects. With a mix of human and model annotation to accelerate the labeling process, your goal will be to annotate efficiently to create the highest quality dataset. 

To help you get the best results, we’ve partnered with great data-science tools to help you be as successful as possible:

Weights & Biases helps your ML team unlock their productivity by optimizing, visualizing, collaborating on, and standardizing their model and data pipelines – regardless of framework, environment, or workflow.

Track everything you need to make your models reproducible with Weights & Biases— from hyperparameters and code to model weights and dataset versions.

OVHcloud helps data scientists and ML teams to easily start notebooks or trainings in minutes. Our AI-managed services allow you to get the right tools at each step of your AI projects, so you can focus on your data skills instead of infra complexity, share easily with your team, manage datasets securely and accelerate your model delivery.

Isahit is an Ethical On-demand Workforce Platform for digital tasks, certified B Corp. We offer you to Build, Train & Deploy a custom and diverse workforce on all your wide digital tasks projects.

Thanks to our diverse & skilled workforce, powerful in-house, partners labeling tools, and efficient workflow, we cover issues across many industries and address a wide range of use cases: from skin recognition to food to predictive maintenance and guarantee 100% accuracy for your models.

During the challenge, Kili Technology and our partners put at your disposal:

By participating in this challenge, you’ll be able to:

Winners of the challenge will be determined on the efficiency & quality of the labeling, and will be rewarded with extensive free vouchers for all solutions and swag packs!

Participants can join as teams or as single contributors. Beginners are welcome. 

Data-Scientists, Join The Ocean Cleanup Challenge! 

At Kili, we believe the quality of the data is what differentiates a successful AI project from one that will fail. That is why we were interested in finding a way to automatically detect the images with incorrect annotations.

error detection for image classification projects

, we continued to investigate how to find labeling errors for Object Detection projects.

Contrary to the image classification problem, where we have a single type of error and a clear metric on how good or bad a prediction is (just by taking the probability of the predicted class), for object detection it is not as  obvious. The issue is that there are many types of errors. According to Daniel Bolya [1] there are 6 types : 

This makes the approach that we used for 

 impossible to use in this case as well. Also, another issue is that we would need the probabilities for all classes, but the current state-of-the-art models provide a single probability for the chosen class, since every bounding box is closely tied to a class. 

Finally, we did not find research done on the object detection task and the only solution we found was that of Daniel Bolya [1] that can only classify the types of errors of a set of annotations given the ground truth (for example to see which types of errors does a model produce by comparing its results to the initial data), but in our case we do not have a given ground truth.

The solution we came up with was to train a neural network using cross validation, as we did for the classification problem, but this time we would just compare the model prediction with the initial annotation for the entire image. For that purpose, we would  calculate a metric of how close the two annotations are based on the comparison and rank all the images according to this metric. Having a very low metric would mean that the annotations are not similar and thus that there is a higher chance of the original annotation being wrong. 

We would then be interested in the images with the lowest values of the metric and we prioritize them for the Review process. Contrary to the solution for classification, we do not identify an exact set of images that are likely to be miss-labeled, but we order all the images with decreasing priority. Once the algorithm is finished, it would be very easy for the reviewer to first see the prioritized images in the Kili platform.

We first separate all the images into 4 folds for cross validation. Next, for each fold the model is trained using the YOLOv5 framework with a pre-trained model and by applying transfer learning. Once the training is finished, the model creates predictions and calculates the metric for each image in the test set (the fourth fold that was not used for training and validation).

The most used metric in research currently is 

 (or mean average precision) and it was introduced by the creators of the COCO (

) dataset. The main criterion for computing this metric is IoU (or Intersection over Union) and it measures the overlap of two bounding boxes :

A prediction is considered correct if it has a value of 

 greater than a certain threshold - usually this threshold was 0.5 (for example for the PASCAL VOC metric), but COCO introduced a new metric that calculates the average precision using different thresholds ranging from 0.5 to 0.95 with steps of 0.05 and then returns the average of all these precisions. This is useful to reward models that give predictions closer to the ground truth in terms of localization.

In order to test this algorithm, we created a small dataset consisting of 250 images of forklifts, with the forklift being the single object class. We annotated all the images ourselves and for 15 out of 250, we created erroneous annotations, hoping that they will have the lowest metrics in the end and they will be prioritized for review.

We then launched our algorithm which at the end updated the priority value of each image in the Kili project. Out of the 17 most prioritized images, 13 were with erroneous labels, which proves that this has indeed helped us to faster find the errors in our annotated images. Thanks to this, the reviewer would not need to look at all the images in the dataset, but only at a subset until seeing that the rest of the annotated images are good enough.

This research was a first exploration of the problem and it gave very promising results, but the biggest problem was the time it took for getting results. For the dataset of 250 images, by using Google Colab Pro and thus the best GPUs, the whole process took around 2 hours (training each fold for 100 epochs and taking as pre-trained network 

which is the smallest network provided by YOLOv5). The main issue is the fact that we use cross-validation and thus we train the same network four times to get precise results for each fold. 

Without taking the time into account, we also need to see how this approach would work when having many classes since we would need to decide if the metric should be calculated as the mean between all classes, or if a weighted mean should be used depending on the number of objects of each class.

To conclude, further research is needed to create a solution that would be optimal and that would be possible to use for all projects. We will continue working on the subject and will try to find a simpler and less time consuming solution.

[1] Daniel Bolya, Sean Foley, James Hays, and Judy Hoffman. Tide: A general toolbox for identifying object detection errors, 2020.

Automatic error identification for image object detection in Kili

Ukraine war, high inflation rate, high volatility of oil prices, aggressive rate action by central banks, upcoming recession cycle… Investing in the financial market or not, we have been greatly impacted by the ever-changing economy from different aspects of our life. The question is, therefore, how do we protect ourselves from the uncertainty of market volatility and seize opportunities? That’s exactly why we brought about the August Community Challenge. 

For the longest time, people try to get more insights into what’s happening in the market, so as to make sense of the next action. Information is always the key. During this challenge, we went through a selected dataset of financial articles and annotated the most valuable information. After a whole week of annotation, we are proud to have created a dataset that can help people classify and predict financial trends. 

This challenge was made possible by our community of data scientists & ML engineers. From different parts of the world, our community contributed to our dataset and eventually made it happen.

There are a few members that we’d like to highlight here, and the first one is ​​Isabelle Francis:

“Sometimes small achievements make my day. I was one of the two winners of 

's Financial News Analysis challenge, where thousands of participants annotated thousands of data. It was the first time that I got to use my knowledge in the financial industry, and put it to use to train and annotate data for a Machine Learning (labeling data and sentiments) model. In the newsroom, we call these corporate event articles ‘market movers’ i.e. we anticipate stocks will move up or down… Then there's also macro/microeconomic data. 

Nonetheless, I had fun learning Kili Technology's tool and was humbled to work alongside such a brilliant team. Never knew that I would contribute to training an AI model; and try to figure out how typical investors and financial journalists/analysts would perceive, when it comes to ‘market moving’ reports”

The other winner Khadidiatou Laye shared:

 for this educational opportunity for me. Thanks to 

 for assisting us throughout this adventure. Congratulations to 

 who undoubtedly contributed to my motivation.

We thank everyone who participated and contributed to creating this dataset. We look forward to meeting every one of you in 

, and of course, seeing you during our next challenge! Till next time!

Uniting Data-scientists For Good: How We Annotated A Dataset To Predict Financial Trends

Artificial intelligence (AI) has attracted keen interest amongst companies of all sizes. According to 

, the global AI software market is poised to reach $126 billion by 2025. Various strong cases have painted AI as the technological shift that allows businesses to thrive in terms of productivity, competitiveness, customer experience, and more. 

Yet, in a rush to incorporate AI in business strategies, you must pause and question if your AI app is trustworthy by design. If you assume that the AI app is ready for deployment because it has been trained with large data sets, you risk potential fallouts from latent or undiscovered flaws.  

 caught attention for the wrong reasons months into its launch. Women customers found to their dismay that their male counterparts were accorded a more favorable credit margin. The incident triggered questions on the possibility of gender bias, whether directly or indirectly, in the machine learning algorithm used for issuing the cards. 

As a key stakeholder in profit-driven organizations, ensuring your AI app remains consistently reliable is pivotal. Conventional AI development methods no longer suffice for increasingly complex and at-scale use cases. Instead, companies must base their applications on a reliable AI framework to add value to customers. 

Reliable AI considers the implications of machine learning programs on customers' safety, fairness, and other sensitivities. Developing a reliable AI application requires revisiting and transforming from the proverbial 'black box' AI concept, which hinders data scientists from explaining how certain decisions were made. 

To make AI more predictable and interpretable, your data engineers must design AI to exhibit the following characteristics.

The AI app must be reliable across diverse types of data and noises in real-life applications. Not only does the app need to deduce accurately from the data inputs, but it must do so within a specific time limit. This is crucial if you apply the AI app to mission-critical applications. 

Instead of being shrouded in secrecy, data movements and decision criteria are conveyed clearly to the users. For example, an AI-driven credit card approval system furnished reasons to customers when declining their applications. Being transparent prevents miscommunications and fosters trust. 

Reproducibility in machine learning is essential for data scientists to investigate and scale the AI model in real-life. To do so, data scientists must obtain similar results each time they feed the same datasets to the same algorithm. Several factors influence AI reproducibility, including code, datasets, and the environment. 

Auditable IA ensures that an organization is accountable and ethical by documenting the development of the machine learning model and on-field deployment. Such paperwork not only helps you to comply with industry regulations but also resolves issues promptly. It provides data scientists with well-documented processes, explanations, and test results. 

An unbiased AI must not demonstrate specific preference toward specific groups of users. Data engineers must work towards mitigating preference-inclined behavior in the algorithm to ensure that the AI app treats all users fairly.  

For all the good it has brought to humanity, AI technologies have drawn flak in recent times. Take 

 on a report that suggests Instagram's unhealthy influence on users' self-esteem and mental health. Instead of making the research public, Facebook chose silence to prevent disrupting the engagement of one of its largest user bases.  

Given the spate of bad reps that AI systems suffered, businesses have every motivation to adopt a reliable AI framework.

The haste to adopt AI capabilities to enhance services or internal business processes might backfire. Although introduced with well-intention, flaws like discriminatory behavior, unpredictability, and inaccuracy can result in reputational and financial fallouts.

Facebook's reputation suffered a blow as its AI model's transparency was placed under tight scrutiny by the 

Dubious handling of data and manipulative practices will alienate privacy-conscious customers from the service. Meanwhile, a 

Hong Kong tycoon's lawsuit against an investment firm

 is a case of how an AI bot went wrong by losing more than $20 million due to an allegedly misrepresented algorithm.  

Companies have a collective responsibility to ensure that users are accorded fair treatment, privacy, and transparency when using AI technologies. Therefore, leading nations have adopted legislation to regulate the digital space. The European 

Digital Service Act (DSA) and Digital Market Acts (DMA

 set forth new frameworks for AI systems, including the call for transparency in conventional black box AI models. 

 that protects the privacy and promotes a bias-free environment. Meanwhile, the

 that it would investigate occurrences of discriminatory AI algorithms according to provisioned regulations. Transgressing such digital acts might result in hefty fines for organizations, as in the case of 

Criteo's $65 million fine for breaching the GDPR.

When deployed on a large scale, AI apps deliver significant impacts across the fabric of society. Any hint of discriminatory policies, intentionally or not, will be felt by the general population. For example, the 

policing software is accused of biased crime predicting that targets specific demographics, partly due to unreported crime data. 

Meanwhile, the successful implementation of reliable AI helps address social challenges effectively. For example, 

identifies youths with self-harm tendencies by identifying associated words in text messages with machine learning and ensures prompt counseling. Therefore, it's wise to revise your AI strategy to reflect your corporate's stance on social responsibility.  

What are the stages for building reliable AI?

Reliability in AI apps requires a mindful and meticulous approach throughout designing, training, and deploying the machine learning model. 

The design stage involves engaging a group of annotators to label training datasets. It is vital that the datasets remain free from bias, or it could jeopardize the entire algorithm training process. Therefore, you should ensure the annotators have adequate knowledge and competencies to label specific data types. 

Also, it's equally important to consider the broader perspective when annotating on a specific subject. For example, engineering professionals may use industrial terms to describe electronic components, while lay people use day-to-day words.

In such cases, enrolling annotators from a single group will lead to skewed datasets. And this will eventually affect the machine learning model when it is trained with biased data. As such, your team of annotators should comprise individuals from different fields, proficiency, and perspectives regarding the subject. 

Considering that the complexity of implementing a reliable AI model lies heavily in the data set, you'll want to establish a feedback loop mechanism for the annotators. As data volume grows, systematic guidance helps annotators produce quality training data and resolve ambiguity.

For example, we recommend setting a disagreement metric to handle exceptions during annotations. The disagreement metric alerts you of data that annotators struggle to reach a consensus on. Besides enabling mutually-agreed annotation, it also provides insights into the complexities of the data set. 

Remember that human errors are inevitable, regardless of how knowledgeable or well-trained the labelers are. Therefore, it's prudent to re-evaluate the labels for correctness, whether manually or via automated tools. It helps to engage an experienced supervisor to oversee the annotation and provide timely feedback to labelers. 

The AI model faces real tests once deployed in the production environment. At this point, it is imperative that you continuously monitor the AI app to ensure that it behaves predictably. Furthermore, it would be best if you continually fine-tuned the data sets to reflect the evolving environment in which the AI app operates. 

To preserve data quality and relevance, I recommend using a two-feedback loop to train your AI model post-deployment. For example, 

improve their ML training efficiency by iterating the feedback process with automated and manual audits. 

With the same strategy, you can hone into granular aspects that will substantially impact the retrained AI model. For example, your data engineers focus on increasing the data consensus, revising annotation guidelines, or reflecting new changes in the real world in the datasets. 

Levering on DCAI approach to build reliable AI 

So far, scientific discussions on AI have centered on algorithms and machine learning models. However, both aspects of AI have achieved a degree of maturity, and therefore, we should divert our focus to the data. After all, clean, unbiased data is the foundation of a reliable AI system.

Many AI systems demonstrate lackluster performance in production environments, despite being trained with cost-intensive machine learning models. This calls for a  Data-Centric Artificial Intelligence (DCAI) approach to strengthen the foundation of AI applications – data quality. Check out this

 to find out what this new ML training approach entails.  

The DCAI principle advocates clean and diverse data ensure optimal performance of your AI app. 

If you're concerned about the reliability of your AI projects, adopting the DCAI approach helps in several ways. 

Rather than solely banking on established AI models, DCAI refocused machine learning efforts to provide better datasets. The approach recognizes that quality data trumps data quantity. It won't matter if you have a massive data set if most are incorrectly labeled, biased, or redundant. 

The DCAI framework emphasized getting data annotation right from the start. It involves mitigative measures to prevent biased labels from compromising data quality and ensure diversity in the datasets. As such, your team has lesser corrective inspections on their plate as data quality was established from the beginning. 

DCAI aligns with Agile methodologies that most app developers are familiar with. Instead of waiting for issues to pile up, data scientists take an iterative approach to perform continuous checks. They apply auditing methodologies, such as the honeypot and consensus metric. As a result, the data engineering team can respond to changing environments, latent flaws, and new behaviors more efficiently. 

As AI makes headway in mainstream applications, data scientists seek to work collaboratively with app developers and other software experts. DCAI envisions a technological ecosystem that allows the respective parties to streamline machine learning processes for increased efficiency. 

For example, data engineers use Kili to reduce missing annotations by automatically keeping track of the labeling process. Kili uses machine learning iteration for shortening further the time it takes to train and deploy a complex machine learning model to the market. 

The call for accountability, responsibility, and ethics when designing AI-powered products is deafening. Authorities are clamping down on AI programs that infringe human rights, stray from ethical boundaries, or pose substantial risks to consumers' well-being. As a business leader, you need to adopt a new AI training strategy or risk the consequences of releasing a flawed AI app. We've presented the DCAI framework as the paradigm shift in AI development. It's about time that you pay equal, if not more, attention to the data that powers your AI model. 

How to build reliable AI, and why you should

The process of data exploration generally begins with setting up expectations. And each one of us can have different ones. In line with DCAI recommendations, you might be trying to identify issues with your original data set; you might also be looking for annotation errors, missing values, and unexpected outliers. Someone else could start exploring theirs to prove or disprove a hypothesis, test an existing model or to carve out a specific slice of existing data to be used for some other purpose.

Let’s take a look at an example. Let’s suppose that you have a dataset containing different characteristics of a group of students. We know their height, weight, gender, exam performance, etc. At face value, everything seems fine. When you start exploring the data, though, one thing just doesn’t feel right: the height range does not resemble anything you’d expect to see in reality. You then pinpoint the weirdest-looking bits of data and find out that some of the variables were provided in feet and inches, while others in centimeters. Thanks to your inquisitiveness and data exploration techniques you managed to avoid feeding wrong data to our model.

For another example, let’s look at image-based projects. In ML projects relying on large image sets, exploring can help you reduce the size of the data set, by enabling you to easily remove duplicates or images that look too similar to one another, images that are too dark, too bright etc. This should help you save time (for example spent labeling wrong images, answering a host of questions) and valuable resources (having to re-train the model).

The basic explore loop looks like the following:

 include seeing or hearing the data firsthand or using statistical methods like univariate and bivariate analysis to figure out data distribution as well as instances of outliers, missing values, collinearity, etc.

 include flagging a file or set of files for further human review in case of annotation errors or generating a new novel slice of the data that is easier to label. 

Depending on your needs, data exploration can be done at different stages of the overall process. With DCAI best practices in mind, you can run it during the annotation of a batch to do quality assurance (for example checking how well labelers do against the gold standard, trying to gauge annotator consensus or spotting edge cases) or to simply inspect examples. But there’s nothing stopping you from doing that 

 the annotation of a batch. Let’s say you’ve uploaded your data set and you’ve set the gold standard labels. Just to be on the safe side, you decide to run a pre-trained model on the same data set and then assess how the model does. While exploring and comparing both annotation types you find out that some gold standard labels are actually missing! The model found additional items and the tools that you used to explore the results made it super-easy to spot the difference.

The person who does the exploration may, or may not, be involved in the annotation process. More generally, the person doing the exploration may be different from the person who did any other process, including uploading, annotation etc. This may be a good thing, with a fresh set of eyes being able to spot new issues with the data set, but this may also prove to be a hindrance, as lack of knowledge of original workflow and organization of data may result in issues being raised in wrong places, leading to extra chaos and rework.

It’s worth noting that even if you’re the person who created the  data set or set up the workflow in the first place, this doesn’t necessarily mean that you’re immune to potential issues with exploring. With projects sometimes taking months or years to complete, after a while even your own data set can feel new or set up in a strange way.

Exploring your data set can be thought of as a “supercharged” version of looking through files on a regular file browser. There’s nothing in the process that you can’t do manually, but with the right set of tools and provided that you know the correct process and know how to avoid possible pitfalls you’ll find yourself doing your work (like discovering valuable insights, filtering data volume, comparing model performance, debugging the human etc.) way more efficiently.

Exploring the Data: A quick guide for MLEs

Exploring the Data: a quick guide for MLEs

What is the common point between Boeing and Toyota? 

They both face tremendous expenses related to the quality of their products: 

Beyond these amounts, the Automotive and Aircraft industries illustrate the challenges faced by the whole Manufacturing Industry. Although each specific field of the industry, at its scale, has its stakes, they all face the same challenge: ensuring the quality of their goods. 

Until recently, most of the Manufacturing Industry heavily relied on human vision inspection to detect these defects. Sadly, this method is not 100% effective. Each visual inspector judges the quality of the product differently, and human quality checks can't scale. 

Thankfully, visual AI technology now enables the Manufacturing Industry to overcome the limitations of Human Visual Inspection. Here's a guide on how computer vision for defect detection can help manufacturers and businesses and how to implement it.

Computer Vision for Visual Inspection: what’s in it for you?

Computer Vision comes with quite some benefits from manufacturers – but also for businesses. On the one hand, it fastens defect detection, and lowers costs, reduces delays and scrap. On the other hand, it brings clear improvements to the workflow and improves the accuracy of defect detection, model, and employee satisfaction...

If we were to sum this up, the benefit is twofold: 

Levering Computer Vision to optimize manufacturing processes and drive profitability

Defect detection using computer vision improves the Manufacturing process and optimizes costs. It minimizes the manufacturer's costs by identifying detects earlier and faster. It also reduces scrap and improves employee productivity. Increasing defect detection accuracy and reducing defect-related costs such as recall also play a part in preventing:

As the visual inspection process levering computer vision accelerate the global manufacturing processes, manufacturers using it are likely to speed up their entire operations. Freed from limitations from the traditional visual inspection performed by humans, they’ll no longer have to invest in costly employee training or additional recruitments. Additionally, they usually tend to see a reduction in their inspection issues, an increase in yield, and an overall increase in product quality with a reduced cycle time. On the midterm timeline, visual inspection powered by computer vision will likely improve the global manufacturing process and the ML model performances – hence being helpful to prepare for future purposes. One of these potential purposes is historical data tracking that Data Scientists of the manufacturing industry can leverage to pinpoint issues and, ultimately, how they can use them to learn and improve future production processes.

Implementing Visual Defect Detection to improve employees’ satisfaction

Human Visual Inspection is a tedious task that can cause visual fatigue and is vulnerable to errors. These cumbersome tasks can lead to:

Since computer vision solutions can detect defects at a faster rate than humans, they will not only improve staff productivity but also allow employees to allocate this extra time and energy to other areas.

Adopting computer vision will likely free employees from feeling bored and unmotivated and allow them to allocate their skills elsewhere, handle more productivity and reach better performances.

Depending on your business, your first point of call is stating the problem that needs to be resolved using defect detection. You could ask yourself questions such as: 

Once these questions are answered, you can move on to the next step: understanding and determining what hardware you need. For example, inspection stations with cameras, lights, etc. Consulting firms can help you implement this end-to-end process, which may require changes to current processes.

To implement an AI- integrated visual inspection system, you will need to collate and prepare the data that will be inputted as the training dataset for the model. This training data needs to be cleaned and happens before the model learns from the data.

In particular, it is crucial to consider the Internet of Things (IoT) data for manufacturing processes, as it provides consistent and accurate insights. The concept of IoT is that each object is tagged with an electronic tag and connected to the Internet, allowing for each thing to be easily identifiable on the Internet. Data can be in the format of video recording, images, etc. When collecting this data, the most critical element you need to consider is the quality of the image or video recordings. The higher the quality of the data, the more accurate results.

1. Using a deep learning model development service, such as Google Cloud ML Engine.

When building your model for visual inspection, there are several computer vision algorithms that you can use, such as image classification and object detection. 

Your choice of algorithm depends on different factors, such as:

Once your training data is ready, and you have your model prepared, the next step is to train your model. Data scientists will evaluate the model's performance and validate its accuracy level. Data scientists can fine-tune areas they believe need improving to increase the model's accuracy. During this phase, a test dataset can be very useful for conducting a further evaluation.

Once your model is producing accurate results, you are ready to deploy. However, the best practice is continuously improving your model – for example, inputting new training data.

Computer Vision for Defect Detection: a key factor of scaling

Scaling is one of the manual visual inspection's biggest and most common challenges. It's no surprise when considering that a single inspection can require 3 control inspectors; running 2 shifts; for 6 consecutive days a week. 

By implementing computer vision, businesses are more likely to optimize their processes and achieve their scaling goals. As the computer vision tool resolves the bottleneck of human quality control inspection, the manufacturers can build in other areas. 

Benefits of Computer Vision for Defect Detection and Scaling

Deep Learning-Based Car Damage Classification and Detection for Automotive Industry

Make Training Process Productive With Kili AutoML and Weights & Biases

Few-shot object detection: a fine-tuning approach

Automatic error detection for image classification in Kili

Technology providers, health authorities and early pharmaceutical investors have begun piloting different technologies like machine learning and cognitive computing to achieve AI to help address the challenges associated with large data volumes. These experiences provide cautionary insights considered as cognitive and AI computing in production systems. Similar to other medical fields, pharmacovigilance is currently facing increasing volumes of data that need support from smart technologies like machine learning and cognitive computing. These technologies follow various perspectives, from regulatory authorities and cognitive IT to the pharmaceutical industry. Therefore, machine learning in pharmacovigilance is progressing and can potentially boost risk-benefit estimation through prognostic aptitudes.

Benefits of artificial intelligence in pharmacovigilance

In health care, AI is used to detect and evaluate medicinal products, and is applicable in machine learning and deep learning. AI is used mostly in processing, extrapolating and reading large, unstructured datasets: medical industries are still early in implementing and integrating AI-driven solutions. AI tools in pharmacovigilance have reduced risks in medical processes, freed up pharmacovigilance experts from routine tasks and delivered high cost and time savings. These tools help manage increasing pharmacovigilance workloads and make the best out of human assets on the available pharmacovigilance (PV) teams. When used together with natural language processing, ML will help to unearth or surface data quickly. Therefore, pharmacovigilance leaders will spend less time identifying AE patterns, including the frequency and severity of AEs. Available ML and AI tools help pharmacovigilance processes translate large datasets from one language to another. This speeds up literature searches and social media scanning for AES to transform scanned documents and automate case follow-ups.

Ways AI automation can speed up your PV processes

With Mapper algorithm, a tool from Topological Data Analysis, create a powerful representation

We help customers to build AI that matters across industries

They trust us for their data-centric journey

Get Started