Quelles sont les étapes à suivre pour garantir la précision de la conception d’un algorithme d’IA ?

Algorithm design in ML should be based on the problem/challenge. A use case is a project I worked on in MedTech. The objective was to identify correlation + causation to patients that are likely to be infertile within the next 5 years. The type of ML problem in this use case was a binary classification problem (YES/NO) A list of the inputs are too long to include. We used relevant data that was required by our Chief Medical Officer who guided us. The output is the likelihood of Infertility in the next 5 years. Regarding unfair bias, it appears that the fault of infertility typically falls on women vs men. We factored in a balance of different types of causes of infertility in both men and women.

Defining the problem statement is the cornerstone of AI algorithm design. SMEs are essential in shaping the problem statement and project objectives. Project objectives should be precise and measurable, tied to business and soft impact metrics. Business impact metrics quantify benefits such as revenue and cost savings. Soft impact metrics evaluate qualitative changes like enhanced collaboration and communication. A well-defined problem and clear objectives provide a roadmap for AI solutions. Collaboration among AI experts, SMEs, and stakeholders aligns the project's purpose and goals.

There are various types of ML problems: Clustering, regression & classification. And within each there are multiple algos 1. Choose type of ML problem: For e.g., we were working on a problem of defining target segments for a consumer goods company. This is an example of clustering. 2. Choose the right algo for your problem: There are multiple algos for clustering: DBSCAN, K-means clustering, heirarchical clustering etc. How to decide which one to choose? K-means: Use when you have a fair idea of the number of clusters/Target segments. Heirarchical clustering: USe when you dont know the number of clusters/target segments DBSCAN: When you want number of clusters/TG to be automatically determined. Hope this helps :)

It’s beneficial to engage with domain experts to delve deeper into the intricacies of the problem you aim to solve. This collaboration can illuminate the path towards a more exact problem statement while foreseeing potential hurdles. Equally important is the alignment among all stakeholders involved. A unified understanding of the problem, the objectives, and the success metrics is crucial. Any misalignment could lead to an inaccurate problem definition, subsequently manifesting in flawed algorithms. It's imperative to ponder the ethical ramifications of the problem you are tackling and the potential biases that may emerge. Ethical oversight could aid in defining the problem in a manner that mitigates harm and bias.

I think the question should be to what extent can we assure data accuracy. Unless your data is infinite and your model is a magic wand, we will always have blind spots. We don't know what we don't know. So it is important to understand the dataset we are using and its limitations. The same applies to the models. A model is always an approximation. Not to forget our human, cultural and so many other biases. I think we should always keep in mind these issues and properly frame our results, especially in highly sensitive applications.

The chosen approach heavily relies on the project's objectives and data availability and quality. Typically, models with greater interpretability tend to be less complex. The specific algorithm selection is influenced by the desired outcomes. For instance, in a root cause analysis project, complex PCMCI models were employed to identify correlations among various data facets. While the model itself may be complex, the outcomes were designed to be easily understandable and actionable.

The approach to design an algorithm would depend on many factors related to the path toward the solution. A good example of this is perhaps a Cancer Diagnosis or Prognosis. If we want to detect breast cancer we could deploy ML/DL algorithms with the objective to automate detection. The approach would be to use historical data of let's say mammogram images labeled as benign or malignant. Regarding feature extraction we may include tumor shape, size and texture. Or any relevant data approved by a medical professional that will provide the framework. In the model training phase we would deploy CNNs to classify the images. We can ofcourse implement Reinforcement learning to continue to improve our models as new data is introduced.

If your task is to recognize objects in images, computer vision with deep learning might be a good pick because it's good at that. However, deep learning needs a lot of data and strong computers, so if you don't have those, a simpler machine learning model might work better. Also, think about what's more important for your project: is it more crucial to get very accurate results, or to have a model that's easy to explain to others? Each AI method has its own set of pros and cons, so choosing wisely is key to building a reliable algorithm.

One of the key challenges i have seen are with how the tail starts wagging the dog...i.e. we become so focussed on data driven analysis - that we ignore what the lack of data in certain segments could be implying.. eg. implicit bias against certain groups even without intending to do so. Here in it is not so much the algorithm's flaw - as in the algorithms' design principles which includes the data considerations.

To test an AI algorithm effectively, use evaluation metrics like accuracy, precision, and recall. For instance, in a medical AI diagnosing diseases, accuracy measures the overall correctness, while precision checks how many true positive cases were correctly identified. Recall measures the proportion of valid positive cases that the algorithm correctly identifies. It's similar to ensuring that all critical components in a machine are in their proper places for it to function effectively. High precision means fewer false alarms, while high recall captures most real cases. For example, in spam email detection, you aim to find the most spam (high recall) without wrongly labeling real emails as spam (high precision).

When implementing algorithms, architects should prioritize considerations like code modularity and reproducibility. After constructing the model, it's crucial to visualize and continually optimize the results. This approach isn't merely about model selection; it's a method to ensure the selected model's suitability for evaluation. Once models are identified, I frequently employ ensemble models to gain insight into their performance across diverse aspects, encompassing linear, non-linear, bagging, and boosting models.

Reproducibility is often overlooked and is a very important aspect of coding in AI operations. Algorithms that depend on stochastic frameworks (e.g., deep learning, clustering, and XGBoost) require a random seed as part of their initialization, and for their results to be reliable and reproducible, a large number of seeds can be used to create an ensemble of trained models before results are trusted. From experience, controlling the random seed from Python's random library or NumPy is straightforward, but doing it with TensorFlow comes with its challenges.

A very widespread approach is the use of methodologies that allow the optimisation of hyperparameters of the models, such as Gridsearch. These algorithms allow us to make combinations, measuring the accuracy or the chosen metric. It allows us to select the optimal parameters for each model.

In my early days, I faced challenges with incorrect parameter tuning which led to poor model performance. It's vital to have a structured approach. Firstly, ensuring your code is clean, well-documented, and modular will save you from future headaches. Secondly, choosing the right frameworks and libraries is crucial as they provide the necessary tools to build, evaluate, and fine-tune your algorithm. Lastly, being prepared to face unexpected data patterns and learning to adapt your algorithm accordingly is a part of the journey. Each challenge faced during implementation sharpens your problem-solving skills, leading to a more accurate and reliable algorithm.

As Faris Mismar already said reproducibility is very important key for a success. Not only seeds, but versions of every library you use also should be tracked. One line of changes in any library your code depend on can affect your progress. It is also important to build result evaluation tools available not only for developers but also for a product owners. This tools should be functional anaugh to make decisions.

I've found that testing is not just a validation step but also an opportunity for refinement. Beyond traditional metrics like accuracy and F1-score, it's crucial to consider domain-specific metrics that align with business objectives. For example, in sentiment analysis projects, metrics like 'Area Under the ROC Curve' may be more indicative of performance than mere accuracy. Additionally, comparing your algorithm against baseline or benchmark algorithms provides valuable insights into its relative strengths and weaknesses, guiding further optimization.

Testing is a crucial piece in the pipeline and should be treated as such. Understanding what each metric means in relation to your data is important (i.e don't just use one metric to validate your model). The pitfalls here can sometimes be due to lack of data. If this is the case, a good overview of synthetic data approaches should be investigated to aid in testing different "real-world" scenarios.

Mechanism for 'Test the Algorithm': Adoption: Data Split: Partition data into training and testing sets. Cross-Validation: Use k-fold cross-validation for unbiased error estimation. Environment: Ensure a controlled testing environment, mirroring real-world conditions. Measurement: Metrics: Depending on the problem, use accuracy, F1-score, RMSE, or AUC-ROC. Confusion Matrix: Evaluate true positives, false negatives, etc. Performance: Monitor algorithm speed and resource consumption. Feedback Loop: Continuously refine based on test results.

Your algorithm might have perfect scores, but still not solve the initial problem. This is also the moment to get back with stakeholders, show what the algorithm can do so far, and make sure you are on the right direction. Align with your team, preferably before testing the model(so you don't move goalposts), what is the threshold score for your model to be good enough for decision making. This will depend a lot on how critical the application is.

In my experience, testing the algorithm means trying it out in various scenarios and datasets to confirm its accuracy and functionality. I use different testing methods and metrics, like precision and recall, to assess its performance and compare it to benchmarks.

Mechanism for 'Debug and Optimise the Algorithm': Adoption: Error Analysis: Identify patterns in misclassifications or residuals. Profiling: Pinpoint bottlenecks in code or computational inefficiencies. Hyperparameter Tuning: Use grid or random search for optimal settings. Measurement: Performance Metrics: Monitor improvements in accuracy, F1-score, or RMSE. Resource Consumption: Track CPU, memory, and GPU usage. Iteration Speed: Measure time taken for each optimization cycle. Convergence: Ensure algorithm stability post-optimization.

Balancing Accuracy and Efficiency in AI Design: In AI algorithm design, achieving a balance between accuracy and efficiency is pivotal. While we aim for highly accurate results, efficiency is equally crucial, especially when dealing with vast datasets in real-time. Designers often employ techniques like approximation, which yields solutions close enough to the exact one, saving time and resources. Heuristics, using educated guesses, prove handy in complex scenarios where precision is challenging. Optimization, on the other hand, helps find the most efficient solution among alternatives. These approaches are valuable tools in the AI designer's toolkit, ensuring the right trade-offs to meet the demands of real-world AI applications.

Often you will hear a lot of advice about picking the problem well and sticking to the right approach, yet the most wisdom of dealing with statistical algorithms lies in being iterative and understanding what the data is telling you. To quote Sherlock, probably an applied mathematician at heart: "It is a capital mistake to theorize before one has data. Insensibly one begins to twist facts to suit theories, instead of theories to suit facts". This is why the best algorithm designers first create a fast/small end-to-end pipeline of their system and put in the right rigorous checks/measurements/quality-controls/metrics. Then the art of creating the magic lies in iterating with this small system relentlessly to a point, of enough progress.

In my experience, after testing, it's essential to debug by identifying and fixing errors in the code or logic. Optimization enhances performance, efficiency, and quality using tools like debugging, logging, profiling, and hyperparameter tuning.

- Debugging is critical to make the solution work when taken into production. Else we have to regularly have down times as there will be issues, which is not acceptable. - Before taking solution into production, we have to test if all edge cases are covered. In software engineering, some professionals are more accustomed to saying testing all scenarios, both mean the same. - deep learning and training Large language models, are computationally very expensive. That’s when optimisation turns critical. Just as an example, graph technique has turned so popular due to its efficiency.

As stated in a recent position statement in Lancet Digital Health, relying solely on quantitative metrics for evaluating the performance of AI algorithms in clinical settings can be of concern, as it may lead to underestimation or overestimation of their clinical utility. The writers suggest the involvement of healthcare professionals in the development and validation of AI-based tools to ensure alignment with clinical needs. Combining qualitative insights from domain experts with quantitative metrics enhances utility and builds user trust.

Ensuring an accurate AI algorithm design involves a crucial step: Robustness testing. It means evaluating algorithms across various scenarios, including offensive breaches and tricky input variations. This thorough examination helps uncover vulnerabilities and strengthens the model's performance. Moreover, ethical considerations are paramount. Every AI algorithm should be developed with a strong ethical foundation, addressing concerns like fairness, bias, privacy, and security. By integrating ethical principles into design and adopting sound data handling practices, AI algorithms can be built to produce impactful and non-discriminatory outcomes. These aspects ensure the reliability and ethical integrity of AI algorithm design.

From my experience, after testing, it's crucial to validate the algorithm, ensuring it meets requirements through methods like cross-validation or user feedback. Then, deploying it involves making it accessible, which can be achieved through cloud computing or web services.

Be comfortable to be transparent in this process - this is usually helpful. If you can afford (i.e. it is not a trade secret) to be be open source for your design principles - please do so.. the more you share with the community the more the community can keep you safe, especially validating blind spots is very helpful - and i think the product is more credible.

The best algorithms can't thrive without meaningful data. In data-centric AI, begin with a clear problem definition. Collect diverse, high-quality data. Preprocess and clean rigorously. Utilize training, validation, and test sets. Continuously improve and ensure ethical and compliant practices. Diverse, relevant data is the key to accuracy.

I enjoy designing algorithms… It’s like putting many pieces to a puzzle together. There’s something rewarding about it. Even if there are challenges, it is empowering to grow your problem solving skills. Algorithm design is a critical part of the objective toward the success as well as the reliability in the long run. Defining the problem and brainstorming on the data needed for data collection and ingestion is imperative. As the problem tells us what data we are going to need, the data and trustworthiness of that data is going to lead us to greater outcomes. In the end we want accuracy, reliability, trustworthiness, a decline in unfair biases, as well as robustness and scalability that we can achieve overtime.

Here’s what else to consider: Often overlooked but crucial in AI design is the "human touch". A recent story highlights this. A startup, eager to launch its AI solution, meticulously followed all design principles. Yet, upon deployment, users felt the AI was impersonal. The solution? They introduced a feedback loop, allowing users to share how they felt about AI decisions. Over time, this feedback refined the AI, making it not just accurate but also more 'human-friendly'. The takeaway? Beyond mathematical correctness, understanding user sentiment and the human context can make an AI solution truly resonate. It's the bridge between a technically perfect AI and one that truly fits into human lives.

In AI algorithm design, the primary focus is understanding how input connects to the desired output. Unlike traditional programming, where fixed rules dictate this connection, AI relies on evolving models guided by programmers. For instance, in face recognition, AI models learn to map images to people's names, gradually improving their accuracy. This dynamic approach sets AI apart from conventional programming, making algorithms more adaptable and intelligent, and it's at the heart of creating systems that can learn and evolve to perform better over time. It's like teaching a computer to recognize people's faces by showing it many pictures and letting it adjust its "thinking" as it learns from its mistakes.

Ensuring accurate AI algorithm design is a multifaceted process that starts with a solid foundation in data. Often overlooked but crucial are the initial steps: exploratory data analysis and data preprocessing. Once you've defined your problem, the data takes center stage in selecting, tuning, and validating algorithms. Understanding your data is the key to building algorithms that deliver accurate results.