⚪️🔵 Edge#82: Fiddler is Bringing ML Monitoring to Enterprises

What’s New in AI, a deep dive into one of the freshest research papers or technology frameworks that are worth your attention. Our goal is to keep you up to date with new developments in AI in a way that complements the concepts we are debating in other editions of our newsletter.

Share


💥 What’s New in AI: Fiddler is Bringing Machine Learning Monitoring to Enterprises 

Monitoring and explainability are some of the toughest challenges faced in real-world machine learning implementations. The constituency and lifecycle of machine learning programs are fundamentally different from other software technologies. As a result, most of the monitoring, debugging, and interpretability tools that we use in traditional software architecture are relatively useless when it comes to machine learning. To that, we need to add that modern deep neural networks remain a black box for most data scientists, which, sometimes, hinders its adoption in mission-critical systems. Solving monitoring and interpretability remains a pivotal challenge for the mainstream adoption of machine learning solutions, especially in enterprise environments. Among the platforms tackling this important challenge, Fiddler AI stands out as one of the most complete and innovative technology stacks in the market.  

Each relevant architecture in the history of the software industry has sparked the creation of monitoring platforms that accompanied its evolution. The networking and distributed computing era powered companies such as Computer Associates and BMC, which dominated the application performance monitoring (APM) landscape for decades until the emergence of the cloud space, in which the baton was passed to companies like New Relic and App Dynamics. Machine learning takes this challenge to a new level, given that we are not only talking about a new runtime architecture but a completely different structure for the programs living in that architecture.  

The Challenges of Machine Learning Performance Monitoring  

Effective monitoring of ML models remains a tough challenge that becomes especially distinct when working with machine learning systems in the real world. While there are plenty of challenges associated with machine learning model monitoring and explainability, most of them can be summarized in the following categories:  

  1. Biases: Machine learning models are a natural mechanism for amplifying data bias. Quantifying the impact of bias in the outputs of machine learning models is far from trivial.  

  1. Real-Time Model Performance Without Labels: Explaining performance metrics during the training of machine learning models is relatively easy as everything can be reconciled back to labeled datasets. That picture looks very different once the models are deployed into production, and they need to operate against unlabeled datasets.  

  1. Real-Time Performance Metric Calculations: Quantifying performance metrics of machine learning models is relatively expensive from a computational standpoint. As a result, it is difficult to maintain a real-time performance view of machine learning models.  

  1. Auto-Retraining: A prevalent way to address the performance decay in machine learning models is to train a new version with an updated dataset and deploy it side by side with the current model. That process is certainly effective but presents a challenge from the model monitoring standpoint, as the performance metric of the new model can be drastically different from that of its previous version.  

  1. Accuracy vs. Interpretability: Finally, one of the most famous dilemmas in the current machine learning ecosystem is the friction between model accuracy and interpretability. In general, models that are relatively easy to interpret might not be super accurate, while more accurate, complex deep learning models prove to be incredibly difficult to monitor and explain.  

Addressing these challenges requires monitoring and explainability platforms that are highly tailored to the machine learning space. Adapting traditional APM stacks to monitor machine learning models has proven to be a largely fruitless effort. Instead, a new generation of companies has emerged to enable monitoring and explainability as a first-class component of machine learning programs. The evolution of APM capabilities into machine learning works has come to be known as model performance monitoring (MPM) and has become one of the pillars of the MLOps movement. Conceptually, MPM enables the key building blocks to track and monitor the end-to-end lifecycle of machine learning models.  

Some of the principles of MPM are beautifully outlined in “Introducing Model Performance Management,” authored by the Fiddler team. Fiddler has been one of the pioneers in the MPM space, adapting many of its concepts to modern machine learning stacks.  

The Fiddler Platform 

Fiddler is one of the undisputed early leaders in the machine learning monitoring space. The Fiddler platform enables a foundational set of capabilities to streamline the interpretability and monitoring of machine learning models. One of the key differentiators of Fiddler is that it does not take a static view. Instead, it enables visibility and interpretability across the various stages of the lifecycle of machine learning models, from training to deployment.

Image credit: Fiddler 

The Fiddler platform was built around the simple principle of enabling “explainable monitoring” for machine learning models. To achieve that, Fiddler can seamlessly enable monitoring capabilities across many machine learning frameworks such as TensorFlow, PyTorch, and Scikit-learn.  

Similarly, Fiddler supports first-class integration with a large number of mainstream data sources commonly used in the machine learning pipeline. This latter capability is important as Fiddler can tie model performance metrics all the way back to the source training dataset. For instance, the Fiddler platform supports native data drift detection and resolution capabilities that can impact the performance of machine learning models.

Image credit: Fiddler 

Another key building block of the Fiddler platform is outlier detection. Anomalies and outliers are among the top factors influencing performance degradations in machine learning models. Fiddler’s monitoring engine allows one to easily detect outliers and generate one-click explanations of the causes of anomalies.  

Image credit: Fiddler 

Another key area of Fiddler’s innovation is related to explainability. Differently from other stacks in the market, Fiddler has interpretability interfaces for both technical and non-technical users. For data scientists and machine learning engineers, the Fiddler platform provides a visual interface to conduct feature attribution and sensibility analysis. A very nice complement is Fiddler’s simpler model interpretability interface for non-technical users.  

Image credit: Fiddler 

One of the biggest nightmares of machine learning explainability is that many of the techniques vary depending on the target datasets, such as text, tabular, or image. This problem increases if we are dealing with models that can have inputs of different types. A model that can have both text and image inputs could be drastically harder to interpret than one with inputs of a single type. Fiddler addresses this challenge with a capability known as hybrid explainability, which enables a very clean interface to generate explanations for models with different feature types such as numeric, text, and, very soon, images. This is an area in which many interpretability tools fall short when applied to real-world multi-class machine learning models. Fiddler’s hybrid explainability interface can generate simple explanations for different data feature types, using a consistent user experience.  

Image credit: Fiddler 

Bias and fairness monitoring is one of the emerging areas of research in machine learning interpretability and one that is particularly hard to crack. Fiddler is one of the first platforms to incorporate bias and fairness indicators for both models and datasets in a very intuitive user interface that helps data scientists actively mitigate these factors throughout the lifecycle of machine learning solutions.  

Image credit: Fiddler 

In addition to its robust set of machine learning monitoring and explainability capabilities, Fiddler created a series of enterprise-grade features such as alerting, collaboration, and access control that make the platform very well equipped for enterprise environments. In its initial phase, Fiddler has seen strong adoption in mission-critical machine learning applications within companies in sectors such as finance and technology.  

Conclusion  

From CA to New Relic, APM systems have been an integral part of the evolution of the software industry, and machine learning will not be different. The unique characteristics of machine learning applications require a new form of monitoring and interpretability platform. The Fiddler platform has been one of the most active innovators in this nascent phase of the machine learning monitoring space. Combining explainability and monitoring capabilities with cutting-edge ideas, such as bias-fairness analysis and a strong set of collaboration features, makes Fiddler one of the most attractive offers for enterprises and startups looking to incorporate monitoring and interpretability capabilities as part of their machine learning systems.  


🧠 The Quiz

Every ten quizzes we reward two random people. Participate! The question is the following:

  1. What is the role of Fiddler’s hybrid explainability capability?

Check your knowledge