AI Summary
This video from Simplilearn provides a comprehensive introduction to Explainable AI (XAI), explaining why it is essential for understanding AI-driven decisions in critical fields like banking, healthcare, and cybersecurity. It uses a relatable bank loan approval example to illustrate how XAI techniques such as SHAP and LIME work, and it also covers the concepts of global and local explainability, along with the limitations of these methods.
Chapters
The video opens with a scenario of a bank loan rejection where an AI system made the decision without any clear explanation, highlighting the core challenge of AI transparency and the need for explainable AI (XAI).
Explainable AI (XAI) is defined as a set of technologies that help humans understand why an AI or machine learning model makes a prediction or decision. It identifies the factors influencing a result, unlike traditional AI which only provides the outcome.
Many advanced AI models are so complex that even their developers cannot easily explain how a specific decision was made. This is known as the 'black box problem', which XAI aims to solve.
XAI is not a separate model but a technique applied to existing models. The process involves: 1) Input data (features), 2) AI model prediction, 3) Applying an XAI technique (like SHAP or LIME), and 4) Interpreting the explanation to understand feature influence.
Global explainability provides an overview of which features are most influential across many predictions (e.g., across 10,000 loan applications). Local explainability focuses on a single prediction to explain why that specific result was produced.
SHAP (SHapley Additive exPlanations) assigns contribution values to each feature, showing how they move the prediction away from a baseline. In the example, a 60% baseline minus 15 (credit score) minus 12 (existing debt) plus 4 (income) plus 3 (employment) equals the 40% approval probability.
LIME (Local Interpretable Model-agnostic Explanations) creates variations of the original input, observes how the model responds, and builds a simpler, interpretable model that approximates the complex model's behavior around that specific prediction.
XAI is applied in healthcare (e.g., predicting heart disease risk), banking (e.g., loan approvals), and cybersecurity (e.g., flagging fraudulent transactions) to help professionals investigate and trust AI predictions.
An explanation does not guarantee the prediction is correct or fair. XAI helps investigate model behavior, but accuracy, fairness, and reliability must be evaluated separately. A model could be biased even if its predictions are explainable.
Explainable AI is crucial for building trust and understanding in AI systems, especially for high-stakes decisions. While techniques like SHAP and LIME provide valuable insights into model behavior, they are tools for investigation, not proof of correctness or fairness.
Mentioned in this Video
Tutorial Checklist
💡 Key Takeaways
Definition of XAI
Provides a clear, concise definition that forms the foundation of the entire video.
03:13The Black Box Problem
Explains a core challenge in AI that motivates the need for XAI.
06:21SHAP's Baseline Calculation
Demonstrates a concrete example of how SHAP calculates feature contributions, making the abstract concept tangible.
15:12LIME's Approach
Highlights a different, model-agnostic technique for local explanations, showing the diversity of XAI methods.
19:46Limitations of XAI
Crucial caveat that explanations do not equal correctness or fairness, preventing over-reliance on XAI.
29:08Full Transcript
[00:07] receive a message saying loan rejected. But here's the question, why was your loan rejected? Was it your credit score, your existing debt, or something else entirely different? Now, imagine the bank tells you that an AI system made
[00:20] this decision, but nobody can clearly explain how it arrived at that result. the one of the biggest challenges with artificial intelligence today. And that artificial intelligence today. And that exactly is where explainable AI or XAI
[00:34] comes in. In simple terms, explainable AI helps us to understand why an AI models makes particular predictions or decisions. Instead of just telling us what the result is, it helps us identify the factors that influence that result.
[00:47] For example, if an AI system rejects your loan application, explainable AI can help show whether your credit score, existing loans, or income contributed to that prediction. And as AI becomes increasingly involved in banking,
[01:01] healthcare, recruitment, and also other important decisions, understanding these predictions becomes essential. So in this video, we will explore what explainable AI is, why we need to know it, and how it actually works, and two
[01:14] popular techniques called SHAP and LIME, and all through a simple real-world example. Before we start, if you're looking to build a career in AI and develop the skills to create real-world AI solution, here's something worth
[01:28] checking out. The Microsoft AI Engineer program is designing to help you understand the complete AI stack from Python and machine learning to deep learning, generative AI, NLP, and also Azure OpenAI. Through live expert-led
[01:41] sessions, hands-on projects, and also multiple capstone projects, you will build practical experience by working on real-world AI applications rather than just learning concepts in isolation. The program also includes Microsoft aligned
[01:55] learning live masterclasses with Microsoft certified trainers, Microsoft Learn certificates, and also preparation for the Microsoft Azure AI-901 certification exam. You'll be working on practical use cases involving AI
[02:08] assistant, predictive models, autonomous driving, customer analysis, sales forecasting, risk prediction, and also generative AI applications, helping you understand how AI solutions are built and applied across industries. If you
[02:21] want to build industry-ready skills in machine learning, generative AI, and AI and gain practical experience developing AI solutions, check out the link given in the description box below and in the pinned comment. Now, let's get started.
[02:34] Before we begin, here's a quick question for you to answer. Which is the primary purpose of explainable AI? Is it to make AI models run faster, to help humans understand why AI makes a prediction, or is it to eliminate all errors in AI
[02:49] prediction, or is it to replace human decision-making? Let us know your Now, without any further ado, let's start. Now, let's start with the initial question of what is exactly explainable
[03:02] AI. Explainable AI is a set of technologies that help humans understand why artificial intelligence or machine learning model makes a prediction or a decision. In simple terms, traditional AI often tells you what the result is,
[03:18] while explainable AI helps you understand why the AI produced that result. Let's understand this with a simple example. Imagine you're applying for a bank loan. You apply for, say, personal loan of 5 lakh rupees. The bank
[03:32] will then use the AI-powered system to evaluate your applications. You provide the following information, such as monthly income, credit score, existing loans, employment type, and also loan requested. Now, the AI will take these
[03:48] factors and process the information. The system will consider these feature and system will consider these feature and will predict if your loan is approved or not. And these are the features that are applicable for those. The system gives
[04:03] you a prediction like if the loan is accepted or rejected, but doesn't clearly explain which factor influenced it. It might be the credit score or anything else. So, the flow happens in this way. You first give it a text, then
[04:19] the machine Repeat, repeat, sorry. First, you are going to input the features in it, and then the machine is going to process that information and give you a clear output whether you are selected or not. In this case, it is a
[04:33] will say the system gives you a prediction, but does not clearly explain what are the factors in what information that we have given. So, as you can see, we know the result, but we don't know
[04:46] what are the factors. Is it the salary? Is it the credit score that is actually influencing this decision? Now, imagine the same thing. Bank adds an explainable AI technique to its prediction system,
[04:59] which is AI with an explanation. Now, instead of just seeing loan rejected, you can see which factors influenced the model's prediction. So, as you can see here, we have loan assessment model, and why did the model predict or reject it?
[05:14] Is it the existing debt, credit score, or stable employment? So, here as you can see, existing debt is the major feature which is affecting the loan rejection. And then, credit score also is not that great. But, stable
[05:29] employment is on the positive side. So, you got few points because of this. This is the basic idea behind explainable AI. The bars above are fictional. In real application, explanation methods calculate feature contribution from the
[05:43] trained model. So, by this, you will have an idea of you have to clear the existing debt, improve your credit score before getting a loan. But, why do we actually need explainable AI? Think about how AI is being used today. AI
[05:58] models can help predict diseases, evaluate loan applications, detect fraudulent transaction, and also recommend job candidates. But, here's the problem. Many advanced AI models are extremely complex. Even their developers
[06:13] may find it difficult to explain exactly how a particular decision was produced. This is often called as the black box problem. Now, what exactly is this black box? As you can see, we already know we are going to input the data, and then a
[06:29] AI model predicts it, and then the decision is given. Now, what happens decision is given. Now, what happens exactly in this part, inside the AI model's brain, we cannot predict. Even the developers who actually made this
[06:43] model wouldn't be knowing it. So, this is called as the black box problem. To solve this problem, explainable AI is recommended. For low-stakes applications such as movie recommendation, this may not be a major concern. But, imagine AI
[06:57] predicting a patient has a serious medical conditions or not. A doctor would want to know which clinical measurements or test results influence that prediction before deciding what to do next. Explainability helps humans
[07:11] investigate such predictions rather than accepting them blindly. Now, how does explainable AI actually work? Here's something important to understand. Explainable AI is not necessarily a separate AI model. Often, it's a
[07:25] technique applied to the existing machine learning model to help interpret its behavior. Let's understand how this actually works using our bank loan example itself. Let's go step-by-step. The first step here is input data.
[07:40] First, the bank collects information about the applicants such as your income, credit score, the type of employment, etc. So, they'll have the document given. This is basically the history of the employment. These details
[07:55] are called input features. So, the input features will be fed inside the machine. features will be fed inside the machine. This features will be fed inside the AI model. In our example, the application has monthly income, a credit score, and
[08:10] existing loans, and also the type of employment has in the input features. Next, in step two, which is the AI model, it makes prediction. So, this is then passed to a AI model, which analyzes the current information. So,
[08:25] this is the AI model that processes the information. The model analyzes the input features based on the patterns it has learned during the training. Now, let's say the model predicts a 40% probability of the loan approval. So,
[08:39] this will say that it is 40% probability. Let's assume the bank has sent an approval threshold of 50%. As you can see, the loan will be rejected because it is 10% less than the approval. But, here's the important
[08:54] question. Why did that model predict only 40% chance of approval? Was it because of the applicant's credit score, existing debt, or something else? That's what we need to investigate. And the here's where explainability AI technique
[09:08] comes in. Step three will be explainable AI. So, consider any models which is sharp or lime will be introduced in the same step as well, after the model
[09:21] accepting and rejecting loan applications, we can apply the explainability technique such as shape or lime to analyze the model prediction, and that is step three. These techniques help us identify which input feature
[09:35] influenced the predicted 40% approval probability. For example, the explanation might reveal that the applicant's credit score and existing debt pushed the model prediction towards a lower probability of approval.
[09:48] Meanwhile, their monthly income and stable employment may have pushed the prediction towards a higher probability. So, here the explainable AI gives you what pushed to positive and what pushed to negative, and then they come up with
[10:02] the total sum. Here, what happens we begin to understand which factors influence that result. And finally, for the fourth step, we have interpreting the explanation. Finally, it will give us an explanation of why the 40% was
[10:17] given. This is to understand which feature had the strongest influence on the model's prediction. So, we'll see further how the numbers are predicted. In our example, credit score and existing debt may have contributed, and
[10:30] also monthly income and also employment type on a positive side, and also credit type on a positive side, and also credit score and also the debt negative side. By examining these contributions, we better understand why the model produces
[10:43] a 40% approval probability and why the prediction fell below the bank's assumed 50% approval threshold. However, there's an important distinction. These explanations help us understand how the model behaved. They don't automatically
[10:57] prove that the loan rejection was correct or fair. And that's the basic workflow of explainable AI. But here's something interesting. Explainable AI can help us understand not just one individual prediction, but also how AI
[11:10] model behaves across thousands of predictions. And that brings us to the next two important concepts, which is global and local explainability. Explainable AI can help us understand predictions in two different ways. So,
[11:24] predictions in two different ways. So, let's say explainable AI can predict. Now, in two ways it can predict. One is global. Repeat. One is global and the other one is local type. It helps us understand the overall behavior of an AI
[11:39] model and identify which features are generally most influential across many prediction. Now, let's consider the same example of loan applications we discussed earlier. Instead of analyzing thousands of prediction, what if you
[11:51] want to understand why a model is particularly this application was rejected? That That's where local explainability comes in place. It focuses on specific predictions and also help identify which features influence
[12:05] help identify which features influence the result. Sorry, can we repeat from the result. Sorry, can we repeat from after this image? Please, can we repeat here? We'll understand the difference by using same bank loan problem itself.
[12:17] Imagine bank uses an AI model to evaluate 10,000 loan application. The bank wants to understand how the model generally makes predictions across all of these applications. For example, does the model rely heavily
[12:31] on credit scores, monthly income, or existing debt? This is called the global explainability. It helps us understand the overall behavior of an AI model and identify which features are generally most influential across many prediction.
[12:45] So, as you can see here, we have global feature importance. The most important feature shown here is the credit score and next the existing debt. These two features will be more implemented and also will be major two features
[12:59] affecting the loan prediction. Then we have a monthly income, employment status, and also age into consideration. So, this is out of 10,000 application we have considered. Now, let's focus on loan applications we discussed earlier
[13:14] for a particular loan. So, as you can see here, we have local explainability. It focuses on one specific prediction and helps identify which feature influence the result. So, here as you can see for that particular
[13:28] applications, we have given inputs like your monthly salary, type of employment, your credit score, etc. So, all the values are here, and based on this feature, what is the baseline approval? We'll talk about this later, but can you
[13:43] see the values exactly are predicted using this. And this is for the exact one application that we have given. And this is called local explainability. So, here the difference is very simple. So, as you can see here, the entire
[13:57] differentiation is given, and global explainability explains over model behavior. So, here all the applications are considered and said. So, here we have shows which features are most important across many predictions. So,
[14:12] we'll have an idea which is the most major um feature which is considered. And then we have helps understand general patterns, and also model logic. And then useful for model monitoring and building trust.
[14:26] Here the question that it answers to is how does the AI model generally behave? Whereas, when you come to local explainability, it explains a single predictions. If you want clarity about your own applications, it will give you
[14:39] that. And here if it is shows how each features contribute to a specific features contribute to a specific decision. And then we have helps answer why particular results was predicted. And then we have useful for users who
[14:52] want to understand their individual outcomes. And the question that this answers is why did AI model make this particular prediction? So, this is for all applications, and this is for a particular explanation. And to generate
[15:07] these kind of explanation, we need to use techniques such as SHAP or LIME. use techniques such as SHAP or LIME. Let's start with SHAP. S H A P. SHAP stands for Shapley Additive Explanation. In simple terms, SHAP helps us
[15:20] understand how much each input feature contributes to a model prediction compared to a reference point. But, how does this actually work? Let's understand it using our bank loan example itself. Now, remember the AI
[15:35] model predicted a 40% probability of the approval resulting in loan rejection under our assumed 50% threshold. But, knowing the prediction isn't enough. We need to understand which feature pushed that probability higher or lower. That's
[15:50] that probability higher or lower. That's exactly what SHAP helps us investigate. So, as we all know, these are the inputs given to the AI model. Now, among all these features, which feature actually pushed for prediction? Now, SHAP assigns
[16:06] contributed values to different features, showing how they influence the model's output. In this hypothetical example, the application's credit score and existing debt pushed the approval probability downwards. So, as you can
[16:20] see, the credit score is negative, the existing debt is also negative. So, the bar, as you can see, it is pushed towards the negative side and also negative side. There is particular values as well given to it. So, it is
[16:35] minus 15 and minus 12. We'll get back to this, what are the importance of this, etc. We'll come back. Some features contribute positively as well, which is income, which is given a point 0.04% positive, and the employment type as
[16:49] well. But, how do these individual contribution produce the final to learn about something called as baseline. As AI model's reference prediction before considering the
[17:02] contribution of this particular application features. Let's say our baseline is set to 60%. So, as you can see, the baseline is set to 60%. And [clears throat] we already know the
[17:16] values as -15, -12, +4, and +3. Now, how did this 40% come comes into place is very simple. So, they'll take the baseline
[17:28] So, they'll take the baseline prediction, which is 60. It'll be -15, +4, +3, and that is a 40% chances of approval of the loan. And that's how the prediction was done. Now, keep in mind
[17:44] examples to help us understand how features contribute work. In a real SHAP analysis, these contributions are calculated using a trained machine learning model. And their values depend on the model output
[17:58] being explained. For example, some SHAP explanations use probability values, while others use a different scale. But the underlying idea remains the same. SHAP helps us understand how different features move the model's prediction
[18:12] away from a baseline. And this is where SHAP visualizations become useful. On the left, we have global feature importance. This helps us understand which feature have the greatest average influence across many predictions. So,
[18:25] this is done across many prediction. The credit score is really high, the prediction, the debt ratio is second position, the income is affecting the third, the age is then coming in, and then the employee mental years, that is
[18:40] how much duration they're already doing the employment basis comes in. So, let me just zoom in for your better understanding. So, this is global importance. This is across a thousands of applications across the AI model.
[18:53] Now, coming to our own observation, which is local observation, as you can see, this is the base value, then we have the credit score, which is On the right, we have a local SHAP. This focuses on one particular applicant and
[19:08] shows how individual feature influence that specific prediction. In SHAP waterfall chart, we can see how these contribution move the prediction from >> [snorts] >> The important thing here to remember is
[19:21] that SHAP can help us understand both how the model generally behaves and why it is produced a particular prediction. Now, SHAP isn't just only one technique used for explainable AI. There's another popular method that takes a different
[19:35] approach to understanding individual prediction, and that is LIME. Now, LIME stands for local interpretable model agnostic explanation. The name might be so complicated, but the idea behind this is actually quite simple. While SHAP
[19:51] explains us how to measure how individual features contribute to the prediction, LIME takes a slightly different approach. It tries to understand how complex AI model behaves around one specific predictions by
[20:03] creating a simpler interpretable model. But, how does that actually work? Let's return back to our bad example problem. Now, let's see how LIME explains the same loan application. Remember our model predicted a 40% probability, as
[20:19] you can see here, it's 40% probability. Unlike SHAP's feature contribution approach, LIME explains predictions by creating variations of the original application and observing how the model responds. That means it will have
[20:32] multiple variation of the same application whether they'll change change slightly the credit score, and see the output, and then compile them together. Let's see how that actually works. So, as you can see here, this is
[20:46] the original applications that was there, and we have similar application where slightly the credit score has been incremented. Then we have earning loans implemented, that is reduced, and then we have monthly income increased. So, we
[21:01] have multiple variation of the same application. So, this is a classic application's credit score is slightly higher? What if their existing debt is lower? Or what if their monthly income changes? Lime sends these modified
[21:16] applications through an original AI model and observe how its predictions changes. Next, Lime uses these observations to build a simpler interpretable model that approximates the original AI model's behavior around
[21:28] this particular applicant. Think of it like creating a simplified explanation of how the complex model behaves near one specific prediction. And because Lime focuses on nearby examples, it's called local explanation technique. So
[21:43] here as you can see, if we have example, this is the original one, what is the results that is there? Example two, example three, example four, right? Then we give it to a simple interpretable model where it has some calculations of
[22:00] the approval score for each individual applications. Then it is given to a local approximation where all these individual simple local model scores are
[22:12] then using a complex AI models are approximated by a local machine again. Which is called a simple local model. Now, this is basically the approximation of all the values that is received for every single variation of the
[22:29] application that was submitted. And finally, the Lime presents an explanation showing which features supported or supports the predicted outcomes. For example, the explanation might show that applicant's credit score
[22:42] and existing debt work against the loan approval. Meanwhile, monthly income and stable employment may support approval. So as you can see, it is given the So as you can see, it is given the result of monthly income is positive,
[22:54] stable employment is also positive, and also the credit score, existing debts, and loan requested, which is maybe the amount also might have affected. Now, it which is 5 lakhs. It would have a different variation where it is of 3
[23:09] lakh and it would have a better score. So, the loan requested is negative because originally it is 5 lakhs, so it is negatively implementing here. So, as you can see we have extra in-depth knowledge of how this is going to work.
[23:23] This helps us understand which feature influence the model's behavior around this particular application. And here's what makes LIME very interesting. It's model agnostic, which means it can be applied to many different types of
[23:35] machine learning models without needing to know their internal structure. However, LIME provides an approximation of the model's local behavior, not a complete explanation of everything happening inside the AI system. So,
[23:48] while SHAP uses feature contribution relative to a baseline, LIME builds a simpler model around one prediction to help us understand it. Both techniques helps us complex AI prediction easier to interpret, but they approach the problem
[24:03] differently. Now that we understand how SHAP and LIME works, let's see how to actually interpret the explanation they provided. We'll continue using our bank problem itself. Remember, our AI model predicted 40% probability of the loan
[24:17] approval. But what does the prediction tell us about the features influencing the decision? Let's find out. Now, let's see how interpret this SHAP waterfall chart. Start with the baseline approval probability of 60%. You've already done
[24:32] this, so we'll take that probability of 60%. The red bars show that the negative contribution. So, the credit score reduces it by 15 and the existing debt
[24:44] by 12. The green bars shows a positive contribution, which is monthly income, which is plus four and we have stable employment, which is plus three. And if
[24:56] you add all these up, we will be getting 40%. 40%. Now, the approval rate is 50%. Since the approval probability is just 40%, there's a 10% difference, the loan will
[25:10] be rejected by the AI model. Remember, these values are illustrative. The chart influenced the prediction, not whether the decision was correct or not. So, according to me here, the major role player is credit score, which is low for
[25:26] the loan approval. Now, let's look at the LIME explanation of the same application. Here we have a horizontal bar chart showing how different features influenced the model's predictions around this particular applicant. The
[25:39] green bars are represent factors that support loan approval, while the red bars represent factors working against approval. For example, monthly income and stable employment support approval. On the other hand, the applicant's
[25:51] credit score, existing debt, and requested loan amount works against it. Notice that the longer the bar, the stronger the feature's influence is in a local approximation, and that's how I identified that the existing debt and
[26:06] also the credit score are the major influence here as per this map. However, these LIME values are estimated from a simplified local model. They don't necessarily match the exact contributions calculated by SHAP. So, we
[26:20] saw a major difference in SHAP calculation and LIME calculation as well. So, now let's compare SHAP versus LIME. So, what we can learn from these two explanations? SHAP shows how individual
[26:33] features contributes move our predictions away from the baseline. LIME shows how a simpler model approximates the original AI model's behavior around one particular prediction. Both help us identify influential features, although
[26:47] they are different methods, the although they are use different methods, and also may produce different explanation values, and that's why these techniques are useful. Instead of blindly accepting an AI-generated prediction, we can
[27:00] investigate the factors that influenced it, and also make more informed decision. Now that we have seen how explainable AI works in practice, let's explore these techniques are being used in real world. Let's look at three
[27:13] important areas where understanding AI decision can make real difference. The first one will be explainable AI in health care. Imagine a hospital using an AI model to predict whether the patient is at risk of developing a heart
[27:28] disease. The model analyzes information such as blood pressure, cholesterol levels, age, and also medical history. But simply predicting that patient is at high risk isn't enough. Doctors also need to understand which factor
[27:42] influenced that particular prediction. Explainable AI can help highlight those factors, allowing medical professionals to investigate the predictions alongside their clinical knowledge. Next, we have banking and finance. As we have seen our
[27:56] loan approval example, banks can use AI models to access credit risk, and also evaluate loan application. Explainable AI can help identify whether factors such as credit scores, existing debt, or income influenced a particular
[28:11] prediction. This helps finance institutions review model behavior, support more transparent decision-making. Another important application is cyber security. Imagine an AI-powered security system flagging a
[28:26] financial transaction as potentially fraudulent. Instead of simply showing a fraudulent alert, explainable AI can help reveal which features contributed to that prediction. For example, the transaction amount might be unusually
[28:40] high. The location might be unfamiliar, or the activity might differ from customer's usual behavior. Security analysts can also use these explanations to investigate the suspicion transaction and decide whether the further action is
[28:54] required. So, whether it's health care, banking, or cybersecurity, explainable AI helps people understand and evaluate AI predictions instead of simply accepting them. But, there's something important we need to remember. Just
[29:08] because an AI prediction can be explained doesn't necessarily mean that the prediction is correct or fair. And that brings us to the limitation of explainable AI. Now, here's something important about explainable AI. Suppose
[29:22] a SHAP shows that the applicant's credit score and existing debt pushed the model towards rejecting the loan. Does that automatically mean the rejection was correct? Not necessarily. SHAP explains which feature contributed to the model's
[29:36] prediction. It doesn't tell us whether the model itself made a reliable decision. And just because some model relied on credit scores or existing debt doesn't prove those factors caused the applicant to become a risky borrower.
[29:50] There's also the question of fairness. What if the model was trained on biased historical leading data? Even if SHAP explained the prediction, that explanation doesn't guarantee that the model treats different groups fairly.
[30:03] Explainability helps us investigate an AI model's behavior, but accuracy, fairness, and reliability still needs to be evaluated separately. The one thing you should remember is artificial intelligence answers what is
[30:17] the prediction. Explainable AI helps answer why did that model make that prediction? And that brings us to the end of this video on explainable AI. Today, we explore what explainable AI is, why understanding explainable AI
[30:31] predictions matters, and how techniques like SHAP and LIME helps us interpret the decisions made by AI models. We also looked at how explainable AI is applied in health care, banking, and cybersecurity. As AI continues to play a
[30:44] bigger role in important decisions, understanding why a model produces a particular result becomes increasingly important. If you found this video subscribe to Simply Learn for more such AI and technology tutorials. And let us
[30:59] know in the comment section below which AI concept you would like me to explain next. Thank you for watching and keep learning with Simply Learn.
