20 Jun Everyone is Wasting Money on Simplistic Ground Truth
As machine learning (ML) algorithms grow up, the first most common ones utilize an array of strings (String) as input, so it is logical by extension that the state of the art tagging companies and software support tagging images using a flat array of tags (e.g. “Car”, “Bike”, “Person”, “Tree”). Startups such as Scale.ai ($18M Series B), LabelBox.com ($10M round A) and even large cloud providers such as Amazon’s Sagemaker are all setting the current low bar of image annotation. These companies have yet to encounter more complex operational requirements and as such are building on top of research and not reality. Today’s labels might be acceptable to the general consumer needing help with their family photos, or basic object detection to avoid collisions, but they do not cover the majority of real-world data in health care, government, Fortune 5000 and sports.
The reality is that objects, actions, and activities within images and even more so, within video, are more semantic in their representation. (https://en.wikipedia.org/wiki/Semantics) The next generation ML algorithms will utilize the upper or lower neural network layers for semantic learning, and the GPUs are just now starting to enable cost-effective memory in deep learning methodologies. Much like ML was developed by modeling the human synapse, semantics are used in our brains as a critical layer that help humans identify objects. We first identify a set of likely objects, such as “apple”, “orange”, “peach”, and then use the semantic context, such as “texture is smooth”, “color is orange” to finalize. (Here is fMRI research combined with ML showing this in action: https://www.nature.com/articles/s41598-018-28865-1) Customers working hard at labeling their data today might as well consider it a one-time effort. These tags are inflexible, overly simplistic and unable to adapt to future requirements.
To solve this problem and ‘future proof’ ground truth data, Orions Systems has invented and patented the “Tagonomy”. This is the only known way of generating semantic data from a directed graph. Computer algorithms and humans think taxonomically, while the mechanics of the world work more semantically. A taxonomy is a directed graph (DAG, https://en.wikipedia.org/wiki/Directed_acyclic_graph), but it is not flexible enough to describe complex relationships without repeating tags exponentially. For example, one tag might be “Clothing”, and its child nodes “Top Clothing” and “Bottom Clothing”. Each of these might also have color as well as many other properties. In a taxonomy, the only way to handle this is by repeating all properties for each “Top” and “Bottom”, duplicating colors for every clothing type. Lastly, tagging anything using string values only locks the data into a rigid structure. This introduces a wide array of new challenges, such as management of future tagging changes, sharing of data with others and enabling multi-lingual crowdsourcing. An ontology is a graph with any structure, while a taxonomy is a directed graph and a tagonomy is a directed graph where any node can have one or more taxonomic properties. Additionally, by traversing a tagonomy, each node can execute one or more commands as well as produce semantic data. In short, a tagonomy can generate ontological data, freeing the data from the limited structure and bounds of either being a simple string array or a rigid taxonomy.
Please contact us to learn more and understand how to create valuable, rather than temporary, ground truth. Don’t lock your company into overly simplistic data models which will surely not adapt over time. Tagging content using tagonomies allows “auto flattening” into a string array, allowing for compatibility with today’s machine learning algorithms, however it is not possible to auto convert the other way around. Ground truth generated today with string arrays will quickly become outdated, and any data labeling will have to be repeated using advanced tagonomic techniques. By using Tagonomies, you protect the future usage of your data while also dramatically increasing the amount and type of ground truth created from it.