How can you make AI performance evaluations more objective?

This is my #1 priority when staring an AI project. To make AI performance evaluation more objective, it’s essential to define measurable goals that include both technical and business metrics. This means not only focusing on technical aspects like accuracy or speed but also considering how these translate into business impact, such as customer satisfaction or operational efficiency. By establishing clear, quantifiable objectives that bridge technical performance and business outcomes, the evaluation process becomes more comprehensive and aligned with overarching organizational goals.

How to make AI performance evaluations more objective? 1. Most often AI performance evaluations are defined for higher accuracy but either unrealistic goal / a number that does not meet the business objectives. A way that has helped my team is to define the performance in terms of: 1. Accuracy above the values achievable currently for that process manually (without AI) 2. Computation needs - GPU / CPU cost and rest of the computations need to achieve the same. Relevance of this cost compared to its business value. 3. Real time / post processing - Does the process requires the performance to be real-time / near real time or is it post-events.

In the medical field where AI is used for diagnosis, treatment, or prognosis, it is extremely important to have objective performance evaluations. Good predictions could save lives. While project objectives can be clear (e.g., detection of cancer subtypes), there is a big unsolved problem in terms of selecting protocols, metrics, and datasets for universal objective performance evaluations. The key word here is 'universal' because these systems will be used anywhere in the world, and they must be developed and tested with 'universality' as a primary goal. It is a difficult problem but fortunately we are seeing progress, and hopefully, one day these medical AI systems will reach the population that actually needs the most.

It is most important to make any project success. Most of the times there is a mismatch in what business need is vs what technical understanding is. For an example, while building recommender or search systems, technical metrics might reflect better on fetched results (recall@k), however business might be all concerned about "conversion" metrics. So it becomes very important to have a clear objective about what your system is going to cater and then carefully drafting it

The objectivity of AI performance evaluations can be improved by following some guidelines: * Define clear and objective performance standards and ensure that they are consistent with the set objectives. * Define clear criteria for validating data and ensuring that the results correspond to reality. * Define clear criteria for validating models and algorithms used in AI. Finally, all these standards must be objective and applicable and should be updated regularly to ensure that they are in line with recent developments in the field of artificial intelligence.

Yes, that is extremely important. While dealing with regression problems, there are multiple performance metrics available such Mean Absolute Error (MAE) or Mean Square Error (MSE) or RMSE. However MSE is the most commonly used one since it measures the variance of residuals aka error. Similarly, for classification accuracy is the most common metric however it may not give a true picture for an imbalanced dataset. It is important to know what information each metric provides and what is the scope of project.

Finding the appropriate metric can be a bit tricky. First, You need to make sure about the parameters you want to evaluate (such as performance, bias, fairness, etc). The second step is to find the right metrics for the chosen parameter. These ones can be found within state-of-the-art and corresponding libraries. We also need to make sure how these metrics work mathematically. The final step is to make sure the chosen metrics are almost consistent with the other ones and if we can trust the results.

Be confident in the metrics you select actually optimizing the problem you’re solving. Find a way to measure the impact you’ve had to solve it. Picking the wrong metric can unintentionally optimize against the problem you want to solve for.

One point that needs to be remember in picking your metrics the nature of your data for example when in classification you know that you need to improve the area under the curve, it is good that you ask yourself if your data is skewed then is way better to use Area Under the Precision and Recall

Our evaluation toolkit should include both automated and semi-automated testing. Reference-based methods like ROUGE provide quantitative benchmarks, while preference-based methods leverage human choice for comparative evaluation – e.g. showing people 2 options and asking them to choose the better option, which we’ll use for evaluation with things like an ELO rating

To ensure fair and unbiased evaluations of AI, it's crucial to use diverse and representative datasets. These datasets should reflect real-world situations the AI system will encounter. Diversity in data includes various inputs, helping the AI generalize effectively across different scenarios. Representative datasets cover a wide range of inputs, preventing biased training that could result in unfair assessments. By combining suitable metrics with diverse and representative data, we can achieve a more objective and reliable evaluation of AI performance. This approach ensures that the AI behaves well in different situations, providing a more accurate measure of its capabilities.

Something everyone overlooks. Data provenance. The origin of data is now more important that the quality of data. Especially in the world of AI and machine learning. Data origination must be your first line of defence when building AI generative or specifically prescriptive models. These are very exposed to manipulation at origin. So consider this point when consuming large data sets to build and train models

“Garbage in garbage out” is fully relevant regarding the quality and the diversity of the data sets used to train machine learning models. This principle implies to think more globally the way to ingest, extract and process the data. Which approach is the most efficient? The data scientist preparing the data on adhoc basis by data wrangling or leveraging analytics to automate the ingestion, extraction and processing of the data for the ML training? On the debate ML versus Analytics, we have most probably a complementary approach when it comes to getting the right data with the right quality for training a model. Analytics is the way to operationalize the preparation of the data for ML. Analytics needs ML but ML needs also Analytics.

In AI evaluations, prioritize the use of varied and balanced data sets. Regularly audit data for biases and anomalies to maintain accuracy in performance assessment. Documenting data sources and methods enhances transparency, aiding in a more objective and reliable evaluation process.

For unbiased evaluation of AI models it is important to consider a diverse evaluation set, reflective of the real-world. Even strong LLMs have revealed surprising failure modes depening on the probability of the input data and even the target task's ground truth being found in the model's prior distribution. LLMs find it difficult to fare well on infrequent tasks than frequent ones even though the two tasks seem equally complex to a human. This highlights the need to use well-balanced evaluation sets.

A mix of quantitative and qualitative metrics is optimal. Additionally, companies should have a balanced dashboard of AI metrics that capture performance, QA/QI and auditing phases.

Our approach as to evaluation of AI systems has to comprehend the spectrum of tests on both functional and non functional dimensions. We perform rigorous testing on coverage security ethics and privacy dimensions of AI ML models through life cycle including at data level, during model building level, before deployment and post deployment.

These statistical methods are fully relevant in for a traditional machine learning era. There is a new paradigm with generative AI. Does it make it sense to use cross-validation when there is no training of the LLM model? What is testing in a context of prompt engineering? There is something new to define with new methods that can apply for generative AI. It opens-up new approaches for validating LLM models. That’s an exciting research journey ahead of us ??

To maintain objectivity, employ rigorous evaluation methods. Follow best practices and standards specific to your field and domain. Utilize appropriate tools and techniques, such as cross-validation, testing, or experimentation, to validate your results. Employ statistical methods like confidence intervals, p-values, or effect sizes to assess the significance and validity of your findings. Document your methods, procedures, and parameters, and make your code and data available for transparency and reproducibility.

Rigorous methods are a foundational, to his new world for trust and observability. BUT don’t rely on the fact alone that u have most governed, comprehensive all inclusive methods and processes, that are trustworthy at some point u will fail, and you wont even realise it. problem with comprehensive and rigours methods in a process is, eventually everybody starts to trust the method and process implicitly. take short cuts and become complacent. thats the first point The second point, is cyber warfare,this is fluid and agile, and changes in real time. More rigorous your methods and process becomes, the bigger your threat exposure to cyber criminals. Its easier to build exploits and countermeasures in a fixed system. Balance is key

Whenever possible, checking the literature for similar problems can be beneficial. For example, in Computer Vision, the ImageNet dataset is very popular, and we know how well humans and computers perform in that domain. If your results are orders of magnitude different, that could be a good indication to check your results. The same goes for forecasting (what is a "normal" RMSE in other problems), NLP, etc.

Ask other smart people to check and test your AI. If you get opinions from different kinds of experts, it helps make sure your evaluation is fair and unbiased.

By seeking feedback on the LLM outcomes, that’s probably the approach that deserves to be further developed with generative AI. The use of metrics can be challenging with generative AI so feedback and peer review is the most pragmatic mechanism to evaluate the results and the performance of a LLM model.

If you are technical person and working as a MLE or Data Scientist, once you come up with a set of evaluation metrics or metric it’s good to go back to the product owner or stake holder to see if these metrics have a simple and clear mapping function to a ROI or process improvement or any other business metric. One can even go ahead and make an example on how this metric translates into a business value. This way it will be much easier to iterate /fine tune the dataset or model once it goes to production and push future updates.

Elevating AI evaluation means embracing collaboration. Seeking feedback and peer review is the key—like refining a masterpiece with diverse insights. Openness to constructive criticism ensures robust, credible assessments. It's not just a step; it's a stride toward excellence.

Industry agnostic benchmarking and learning from real life implementations as many vendors out there with lab developed solutions with no proven results to talk about. AI solutions have ability to transcend industry boundaries and can make compelling case studies to push adoption and investments.

To enhance the objectivity of AI performance evaluations, it's essential to establish a baseline for comparison. This involves comparing the AI's performance against established benchmarks or previous versions of the system. Additionally, incorporating blind testing methods, where evaluators are unaware of the specific model or version they are assessing, can prevent biases and ensure a more impartial evaluation. This approach helps in achieving a fair and unbiased assessment of the AI's capabilities.

-Define specific, measurable metrics aligned with the intended purpose of the AI system. Clear objectives help in quantifying success. -Establish industry or task-specific benchmarks for comparison. This provides a standardized reference point for evaluating performance. -Test AI systems in real-world scenarios that mimic actual usage conditions. This ensures that evaluations are reflective of practical performance. -Use diverse and representative datasets for training and evaluation to avoid biases and ensure the AI model's generalizability. -Employ cross-validation techniques to assess how well the model generalizes to new data. Techniques like k-fold cross-validation help estimate performance across different data subsets.

It's critical to understand how and why the model is making certain decisions. A clear methodology to interpret the model is important. This will also help explain the model outcomes to non-technical audience.

Leistung ist nichts ohne Akzeptanz bei Nutzern und Mitarbeitern. Dies ist ein oft geh?rtes Problem beim Einsatz von KI. Es ist wichtig, entsprechende Ziele zu formulieren und Strategien zu entwickeln, um dies zu gew?hrleisten.