MIT’s Sarah Alnegheimish Unveils Orion: An Open-Source Anomaly Detection Framework for All
In the ever-evolving landscape of machine learning, accessibility, transparency, and trustworthiness are paramount. Sarah Alnegheimish, a PhD student in Principal Research Scientist Kalyan Veeramachaneni’s Data-to-AI group at MIT’s Laboratory for Information and Decision Systems (LIDS), is tackling these challenges head-on.
Alnegheimish is the driving force behind Orion, an open-source machine learning framework and time series library designed to democratize anomaly detection. Orion empowers users, even those without extensive machine learning expertise, to identify unexpected patterns in large-scale industrial and operational settings.
Growing up in a home where education was highly valued instilled in Sarah a deep-seated desire to make machine learning tools accessible to everyone. Her positive experiences with open-source resources further fueled this mission, making her realize that accessibility is key to adoption. To truly make an impact, new technologies must be easily accessed and assessed by those who need them most, hence the open-source nature of Orion.
During her bachelor’s degree at King Saud University (KSU), Sarah frequently utilized MIT OpenCourseWare to teach herself. She later joined the King Abdulaziz City for Science and Technology (KACST), where she began conducting research with Veeramachaneni. This collaboration eventually led her to pursue graduate studies at MIT under his guidance.
Orion is the culmination of Alnegheimish’s master’s thesis, focusing on time series anomaly detection. By identifying deviations from expected behaviors, Orion offers crucial insights for various applications. From detecting cybersecurity threats in network traffic to predicting equipment failures through sensor readings and monitoring patient vital signs to reduce health complications, the possibilities are boundless.
The framework employs statistical and machine learning models that are continuously logged and maintained. Users can analyze signals, compare anomaly detection methods, and investigate anomalies within an end-to-end program. The framework, code, and datasets are all open-sourced, fostering collaboration and transparency.
The design of Orion prioritizes transparency by labeling every step in the model and presenting it to the user, which helps to build trust. The goal is to consolidate various machine learning algorithms into a single, user-friendly platform, enabling anyone to leverage these models out-of-the-box. Its success is evident as public users install the library and run it on their data, leveraging the latest methods for anomaly detection.
For her PhD, Alnegheimish is exploring innovative approaches to anomaly detection with Orion, repurposing pre-trained models to save time and computational costs. While these models were originally trained to forecast, not find anomalies, prompt engineering allows them to be adapted for anomaly detection without additional training.
Orion’s accessible design enables users to develop systems that make them accessible and adaptable for others. She uses abstractions that provide universal representation for all models with simplified components. The models follow a sequence of steps from raw input to desired output, standardizing the input and output to allow for flexibility in the middle steps.
Alnegheimish has also investigated the use of large language models (LLMs) as intermediaries between users and the system. With her software, users can train their model using the “Fit” command, and detect anomalies using the “Detect” command. This simplifies the interaction with Orion, making it even more accessible to a wider audience.
With over 120,000 downloads and thousands of users marking it as a favorite on GitHub, Orion is making AI accessible to everyone. This real-time adoption through open source demonstrates the impact and value of Alnegheimish’s work.
