Omid Madani 

omidmadani att yahoo dot  com [click on picture, for another version of me, with my son, from just a few years back!]

I am interested in all aspects of intelligence, and all matters mental! (specially from a computational perspective) From 2014 to 2024, at Cisco, I developed machine learning and data analytics pipelines, almost entirely unsupervised, to help define and enforce network policies and make data centers more secure. Previously, in reverse chronological order, I was at Google (the perception group and ml for YouTube, NLP), at SRI (AI Center), Yahoo! Research, U Alberta, U Washington, U Houston, Saddleback Community College, and spent my formative years in Iran (Bandar Abbas and Tehran), then Dubai, UAE. My research has included the following threads:

• I am inspired by how our minds might work! In particular, the type of problems that we solve. An important feature of our intelligence that perhaps separates us humans from many other animals is the huge number of inter-related concepts that we (apparently) acquire and develop and effectively use. Here, by "concept" I mean a recurring (and useful) perceptual pattern (supported by a locus of representation in our brain), such as words and phrases uttered in continuous speech, or visual objects and entities such as books, faces, common action sequences, etc. What tasks and problems, processes and algorithms, and representations and data structures, support developing such complexity? These questions are very broad and provide a starting point. I hope to contribute to answering some such questions, in particular from the perspective of computational learning and development.
• I have worked on mostly AI, and machine learning in particular, now for more than 2 decades! Key problem properties have revolved around unsupervised and self-supervised learning, online learning, large-scale learning, models of active learning, multiview learning, data mining, situated/embedded systems (in rich environments), imperfect/noisy features and labels/feedback, multiple (learning) systems interacting, and so on.
• I have been working on efficient learning algorithms for supervised learning problems with high input and output dimensionalities (potentially huge numbers of features and classes/concepts). The focus has been on online algorithms that can handle dynamic growing sets of classes, possible nonstationarities, etc.
• Other related threads in my work include empirical and theoretial analyses of algorithms (in particular, motivated by AI decision-making/planning problems, such as Markov Decision Processes (MDPs)), exploration of applications, and discovery of new problems in the applications. Past and future application areas include information retrieval, text and natural language processing, game playing, personalization, vision, and so on.

Please see the following links for further information on my work.  

• Publications
• Selections from my work:
  • NEW: Tracking Changing Probabilities via Dynamic Learners, in arXiv, 2024. Continues the work on prediction games: online (lifelong) multiclass probabilistic predictors for non-stationarities.
  • New: An Information Theoretic Score for Learning Hierarchical Concepts, Frontiers in Computational Neuroscience (in celebration of the 75th anniversary of Claude Shannon's paper on a mathematical theory of communication), 2023.
    • Earlier: Expedition: A System for the Unsupervised Learning of a Hierarchy of Concepts, a revival of Prediction Games! arXiv, 2021.
  • New Dataset: A dataset of 22 graphs (our work at Cisco), is now available at the Stanford SNAP (thanks to Rok Sosic). It contains edges (TCP/UDP) from distributed applications, and two graphs have reference groupings of nodes (ground truth). Our IWSPA 2022 paper describes the data (here's the README file).
  • A dataset of multimodal feature vectors, YouTube Multiview Video Games, available at UCI repository, specially useful for multiview (multimodal) machine learning research (also here in smaller partitions).
  • Presentation video on "index learning" (efficient linear classifier learning for many classes, 1000s and beyond) (a precursor to prediction games), posted by sfbayacm.
  • A Python implementation of sparse EMA ("Emma"), a version of (sparse) index learning, suitable for non-stationary many-class problems (thanks to Jose Antonio).
 
A poem by Omar Khayyam.