How do you secure AI performance?

When I oversea the training of models, my particular emphasis is on the quality of the fine-tuning data, and there are compelling reasons for this approach. The base model is typically trained on a massive dataset in an unsupervised manner. Thus, fine-tuning with excellent data provides significant value—it's a matter of getting more bang for your buck. A dataset consisting of just a few thousand examples can yield outstanding results. When evaluating a dataset, two factors are paramount: the quality and diversity of samples. The importance of quality samples cannot be overstated, and diverse samples enhance the overall effectiveness of the model.

The saying "garbage in, garbage out" is key for Machine Learning applications. Most databases and data tables are only as good as the data used to populate them. Most of the times, this data is generated upstream and if that presents quality issues it will have a detrimental effect downstream. Improving the data acquisition process to ensure quality is the highest at the acquisition phase ensures that downstream systems, including Machine Learning models, have better chances of higher quality.

A majority of training data that exists today is from a handful of developed nations who had access to advanced technology for the past few decades. Emerging nations in the AI space in places like Africa are poorly represented in data models. Data must also be inclusive - we must take steps to ensure data diversity so we don’t further marginalize people who are underrepresented in datasets.

Recent advancements in AI have unlocked a plethora of possibilities, but they have also introduced new vulnerabilities, especially. One such vulnerability is prompt-injection, a form of malicious input designed to manipulate an AI system's behavior. It leverages the model's capacity to potentially corrupt a system. A beneficial practice is to adopt the perspective of an attacker and attempt to inject malicious prompts. Moreover, introducing a classifier layer could aid in detecting anomalies may bolster the system's resilience against such threats.

Testing goes beyond accuracy and performance. It also includes aspects such as robustness , interpretability, reproducibility and system integration. Key challenges to highlight: - Accuracy: required rates of accuracy are dependent on the nature of AI applications & associated risks. Autonomous vehicles need higher accuracy rates compared to movie recommendation engines. - Bias: Thorough understanding of data is required to test for bias. If bias is not addressed at the early stage of the design, it becomes costly to remove it afterwards. - Reproducibility: this is a key issue for AI models. This is more relevant when the environments associated to training, testing & deployment are very different such as in AI healthcare applications.

Models are like living beings. They must be nurtured, fed, and their performance can vary. So it's essential to know when to update the models. Key is to: 1??Select monitoring metrics to detect various potential issues such as drift detection algorithms to track data changes or rule-based methods to detect one-off events, and track data inputs/outputs to help MLOps teams diagnose problems 2??Automate monitoring and alerting. Companies are transitioning from homegrown tools for this to specialized ones. Options include tools for end-to-end ML lifecycle mgmt such as IBM Watson Studio or Vertex AI, or startups focused on this such as Arize, Fiddler, etc. 3??When AI performance drops, consider if you should update or replace your model instead

Protecting the data used to train and test the AI model is crucial. This involves implementing secure data storage, access controls, encryption, and proper data anonymization techniques to safeguard sensitive information.

Continuous assessment allows for identifying biases, errors, or unintended consequences that may arise over time. By regularly reviewing model performance and addressing any issues promptly, you can demonstrate a commitment to maintaining transparency, fairness, and ethical use of AI technology. This proactive approach further enhances trust, confidence, and the overall acceptance of AI models in practice.

I'm just getting started working with Vertex AI models, so pardon any rookie feedback. To evaluate AI models and their performance I'm looking at quantitative metrics like accuracy, precision, recall, and mean squared error for objective assessments. And I believe qualitative evaluation means grading the output on bias, fairness, explainability, and user feedback to gain insights. Finally I know that it will soon be necessary to check if the model complies with laws, regulations, and ethical guidelines. Which can vary, particularly with the EU AI Act, and the upcoming US legislation.

Hi Ramon nice write up.. I have a quick question I’m assuming in your criteria when you were for referring to “models” to me the most important aspect would be the “features” and the process used to evaluate including document, available lineage just curious. i’ve been in BI forever only in AI the last couple years. Thanks