What can you learn about the latest trends in explainable AI?

one of the hot topics on explainable AI is counterfactual explanations. IT shows alternative scenarios that could have led to different predictions, helping users understand the factors influencing the outcome. They aid in understanding bias, fairness, and causality by exploring the impact of different variables on the AI model's decision-making process.

Increasing accessibility to Explainable AI (XAI) is essential for several reasons. It introduces diverse insights from various individuals, creating more innovative solutions. Increased accessibility means more people can learn and contribute to AI's evolution, regardless of background, driving its development. Greater accessibility improves AI's accountability, fairness, and correctness through more stakeholders understanding, developing, and monitoring its use. Lastly, increased accessibility can bridge the gap between AI users, developers, and policy-makers, leading to a more comprehensive system understanding.

We can also consider these points: 1. User Feedback Integration: AI learns from user feedback to improve concept-based explanations. For example, if users find certain terms confusing, the AI adapts its explanations to use more familiar language. 2. Natural Language Generation: AI expresses explanations in human-readable language, avoiding technical jargon. For instance, instead of using complex terms, the AI explains its decisions in plain and understandable words. 3. Interactive Explanations: AI offers interactive features like visualizations and step-by-step breakdowns. Users can explore how the AI arrived at a decision, gaining clarity by asking questions for further understanding.

Concept-based explanations in Explainable AI (XAI) address the challenge of making AI insights accessible to users. Instead of technical details, these explanations focus on high-level human-understandable concepts like emotions, goals, values, or categories. For instance, a facial recognition system's concept-based explanation might highlight a person's happiness based on observable facial features, fostering better comprehension and user engagement.

To address the "smiling" example. Yes - sure - this is an example of XAI. However a bias that the AI may experience is that correlation. Yes, the person may be smiling, but the eyes may be giving more subtle indicators that the smile is an angry smile or a forced smile through sadness. Training it on larger data sets should alleviate the problem somewhat.

Using counterfactual explanations to make AI systems more explainable is like understanding the different paths a situation might take. Imagine you're at a crossroads at night, deciding which way to go. You can learn from both the path you choose and the one you don't. By comparing what actually happened with what could have happened, you gain insight into why things turned out the way they did. It's about looking beyond the surface, understanding the underlying factors, and learning how to make better decisions in the future. It's a tool that doesn't just tell you what happened, but why it happened, much like analyzing the scene of a minor incident in the city to prevent bigger problems down the road.

Multimodal explainability is a multi-stage comprehensive process. It states the contribution of each modality to the underlying task. The overall impact of the modality presence or absence and the correlations and associations of the modalities at the individual and fusion/ integration level.

Consider an AI system used for loan approval. The AI might reject an application based on factors like the applicant's credit history, income level, and employment status. Just receiving a 'rejected' status from the AI with no explanation leads to frustration for the loan officer & applicant. With interactive explanations, the loan officer can probe AI's decision. If it was the applicant's income level, then the loan officer could explore scenarios and income ranges in which the loan might be approved. This interactive process not only makes AI's decision-making process more transparent but also empowers the loan officer to better communicate with the applicant and potentially find solutions.

❓⚡️Contradictions reveal guessing | 🧪💜 Test your hypothesis |🎭😻 In-context learning Why doesn't 🤖 spontaneously use emoji? ❓⚡️ Direct questions -> contradictory answers 🤖: Emoji were not excluded from pre-training data. 🤖: Emoji were not included in pre-training data. 🧪💜 Hypothesis: Emoji symbols = language Human: Show me 3 colored emoji 🤖: 💜 Purple Heart 💙 Blue Heart 💛 Yellow Heart Human: What is 💜? 🤖: Indigo heart Human: What color is 💜? 🤖: Violet Human conclusion: Language describes 💜 as different, related colors. 🎭😻Play with 🤖 to teach it 🤖: Ask me to respond 'in the style of' a sassy friend, so I occasionally respond to you with emoji. Human: Thanks for your suggestion! I’m off to play with CatGPT 😻

Interactive explanations in AI are akin to the communication between a watchful guardian of the city and the people he protects. It's about establishing trust, much like a seasoned detective working closely with a partner. They don't merely follow each other's lead; they ask questions, seek understanding, and provide feedback. In a medical diagnosis, for example, allowing the patient to question why a certain test or medication was chosen creates a two-way street. This level of interaction isn't just about following orders; it's about understanding the reasoning behind them. It makes the process more transparent and the relationship more trustworthy.

As AI continues to get deeper into our society, ethical explanations are more important than ever. While we push the boundaries of what AI can do, we must ask whether it's the right thing to do. Before explaining AI's decisions, we must ensure we're putting it to use in alignment with our societal values and ethical norms. As Spiderman's Uncle Ben would say, 'With great power comes great responsibility!'

Navigating the complex streets of ethics in AI is much like patrolling the dark alleys of a city at night, dressed up as a vigilante. It requires constant vigilance and a firm grasp of right and wrong. In any scenario, there might not always be a clear right or wrong decision. But the guiding principles must always remain firm, and transparency in those principles is key. For instance, the self driving car explanation.

COGNITIVE BIASES Data are the reflection of our reality and facts. However, they are ever evolving. Biased AI models have the potential to harm or derive dramatic consequences in legal decision-makings. e.g. a prediction mistake of a suspicious movement triggers an arrestation. And what about self-driving cars' Trolley Problems? Those are indeed solved case by case. ''Cognitive Biases'' are shortcuts of our flawed reasoning and also legacies of our social background. One way to mitigate them is by correcting Data and Model Drifts, be it in Computer Vision or Natural Language Processing, through hyperparameters tuning, metadata manipulation or right from the start, through bringing meticulous care to preprocessing a balanced dataset.

Incorporating diverse perspectives and voices during the development and training of AI systems minimizes the risk of perpetuating biases present in the data and algorithms. Ethical explanations play a pivotal role in this process, as they provide transparency into the moral principles and criteria guiding AI systems. By understanding how AI makes decisions and balances trade-offs, stakeholders can assess whether these systems align with societal values and norms. Implementing ethical explanations fosters trust and confidence in AI.

This is a key part of AI development and actually comes before AI training can begin. What are the ethical standards we want to build into these systems? We have so much diversity across the world and different ethical norms, so which ones are best here? Taking the self-driving car example, as a society what choice do we want if a car has to make a decision between hitting a child crossing the road or potentially killing the driver by crashing? From an ethical perspective, these discussions and debates need to happen now so that consistent standards can be adopted in AI development everywhere. This will be a major focus of policymakers, working with the industry, over the next few years

In line with the thoughts here, imagine you are using an AI-powered medical app to diagnose your medical symptoms. The app could provide you with a multimodal explanation of the diagnosis, including: A text summary of the diagnosis, in plain language that is easy to understand. (READ) A spoken explanation of the diagnosis, that can be helpful for people who have difficulty reading. (HEAR) Visual representations of the diagnosis, such as images of your medical scans or graphs of your test results to look at it closely. (SEE) By providing these different multimodal explanations, the app helps understand your diagnosis in a way that is both comprehensive and engaging and AI system more understandable and transparent to the user.

Comprehensive explainability models should consider the end user and : *. Explain the data - data source, dataset(s) name and author, reason for selecting datasets *. Explain the predications made - explain variables / features selection and weights, and justify reason for feature selection *. Explain the algorithms - the model type and reason (eg Image segmentation for pathology slides), explain the layers and the thresholds for the model predictors In my experience LIME works particularly well with whole slide images in digital histopathology for cancer detection to highlight tumour regions of interest and tumour burden estimation

Often overlooked is using simpler, less complex solutions. One should never use a complex solution when a simple one would do.

One thing I found helpful in explaining the AI use case is how it's already integrated in their everyday scenarios like shopping online (eg recommendations) or travel routes (eg using routing tools).

The complex data transformations and the hidden layer processing in deep networks make the model interpretability challenging and hence is commonly perceived as an opaque black-box model with a lack of understanding and explainability about the model's internal logic and decision-making process and the predicted outcomes. The widespread use and adoption of AI in real-life applications cannot rely merely on performance and accuracy metrics. The need for explainability integration in the model design to deployment phase is paramount for the practical adoption and usage of AI-based applications deployed in critical domains directly affecting human lives such as healthcare, autonomous systems, finance, legal, etc.

It is worth mentioning that the field of AI explainability has exponentially expanded since 2017 with an increased number of publications focusing on deep learning. Explainable AI techniques are split into three categories: sef explainable modelling, post hoc explainable modelling and global/ local explainability.