MIT’s Sarah Alnegheimish Unveils Orion: An Accessible Anomaly Detection Framework for All

In the realm of machine learning, accessibility and transparency are often seen as hurdles rather than stepping stones. However, Sarah Alnegheimish, a PhD student at MIT’s Laboratory for Information and Decision Systems (LIDS), is breaking down these barriers with Orion, an open-source anomaly detection framework designed for everyone.

Under the guidance of Principal Research Scientist Kalyan Veeramachaneni, Alnegheimish is dedicated to making machine learning systems more approachable and trustworthy. Orion, a testament to this mission, is a time series library capable of detecting anomalies in large-scale industrial and operational settings without supervision. This means that even those without extensive machine learning expertise can leverage its power.

Alnegheimish’s passion for accessible technology stems from her upbringing. Growing up in a home where education was highly valued, she witnessed firsthand the importance of shared knowledge. This early influence, coupled with her own positive experiences with open-source resources like MIT OpenCourseWare, fueled her desire to create tools that are readily available to all.

“I learned to view accessibility as the key to adoption,” says Alnegheimish. “To strive for impact, new technology needs to be accessed and assessed by those who need it. That’s the whole purpose of doing open-source development.”

Orion’s design reflects this philosophy. It utilizes statistical and machine learning-based models that are continuously logged and maintained. Users can analyze signals, compare anomaly detection methods, and investigate anomalies in an end-to-end program – all without needing to be machine learning experts. The framework, code, and datasets are all open-sourced, ensuring complete transparency and unrestricted access.

One of Orion’s key features is its transparency. “We label every step in the model and present it to the user,” explains Alnegheimish. This level of detail allows users to understand how the model works and build trust in its reliability.

Currently, Alnegheimish is exploring innovative ways to enhance Orion’s anomaly detection capabilities using pre-trained models. By repurposing these models, she aims to save time and computational costs while pushing the boundaries of what’s possible. Although these models were initially trained for forecasting, Alnegheimish believes they already possess the necessary patterns to detect anomalies through prompt engineering, eliminating the need for additional training.

Alnegheimish’s dedication to accessible design extends beyond the technical aspects of Orion. She has also focused on creating abstractions that provide universal representation for all models with simplified components. This approach allows users to easily adapt the system to their specific needs, as demonstrated by her mentorship of two master’s students who were able to develop their own models using Orion’s abstractions.

Furthermore, Alnegheimish has implemented a large language model (LLM) agent that acts as a mediator between users and Orion. Inspired by the user-friendliness of tools like ChatGPT, this agent allows users to interact with Orion using just two commands: Fit and Detect. This simplified interface makes AI more accessible to a wider audience.

Orion has already garnered significant attention, with over 120,000 downloads and thousands of users marking it as a favorite on Github. This widespread adoption is a testament to Alnegheimish’s vision of making AI more accessible to everyone.

“Traditionally, you used to measure the impact of research through citations and paper publications. Now you get real-time adoption through open source,” she concludes. The impact of Orion is clear: It’s empowering individuals and organizations to harness the power of anomaly detection, regardless of their technical expertise.