Is realtime AI a thing?


An AI service is built by transforming raw data into feature vectors to feed into a machine learning algorithm to produce a predictive model(*). The algorithm works its magic on the feature vectors and outputs a trained model from which predictions (or scores) on unknown data inputs are made.

Predictions remain fresh until new data becomes available when the model is re-trained to bring it up to date. The dual problem raised at this stage is the current model predictions are out-of-date and the length of time to re-train is proportional to the volume of data. The industry solution is to re-train the model as fast as possible.

On many levels, it is an odd situation because modern software systems don’t work like this. A good counter example is a database management system (dbms) which allows you to manage and manipulate data with simple crud and query statements. Data can be added, updated and deleted at any time and people can continue to query the database at the same time. A dbms is a dynamic system whereas an AI service is built on a regularly re-trained static model.

Discovering things in the here and now

AI is a category term for different deep learning and machine learning methods for specific tasks including regression, classification, clustering and recommendations. The focus here is on realtime AI for the discovery and detection of items (“things”) such as images, video, audio, db records, documents and so on and is characterized by:

  1. Automatic generalization: Given one or more things as a query the AI learns in realtime to discover other relevant things. This means there is no machine learning ‘model training’ phase.
  2. Interactive: Engage directly with the AI in different and unpredictable ways through multiple query options.
  3. Dynamic: Add new data to the AI as and when.

All three areas are covered over a series of posts and to kickoff, this slidedeck shows results using images(**):

(*) It is acknowledged in the image and audio space, trained deep learning models are employed without a machine learning component for object classification and recognition.

(**) In the example results, the feature vectors are generated automatically from a deep learning model.

Next post: Interactive AI … coming soon