Order in Data, Laws of Physics, and Seeing through Walls

In today’s issue of our #InfocusAI digest, we will discuss the importance of getting your data in order for successful AI performance and tell you about the world’s first library of CV algorithms based on the laws of physics, a smart walking stick for the visually impaired, and a WiFi-based method for tracking people through the walls. You will also learn about large language models exceeding in some ways GPT-3. 

Issue 10, 12–26 January 2023

A recent study by NewVantage Partners shows that only 24% of 116 global Fortune 1000 businesses consider themselves data-driven. Even fewer of them (21%) claim that they have what could be considered “data cultures.” And only about a quarter of the companies in this survey think they are doing enough to ensure responsible and ethical use of data. These findings are far from being optimistic. Experts confirm that outstanding issues of data management are a major hindrance to the development of artificial intelligence. A Forbes article citing Mona Chadha from Amazon Web Services says that principal bottlenecks include poor data quality, unfair bias, and lax security. Poor data quality can result in inaccurate computations and inconsistent behavior of AI models, unfair biases may lead to discriminatory behavior, and AI vulnerabilities can play into the hands of fraudsters. Lack of order in data and absence of a data handling culture can bring businesses nothing but financial and reputational losses from the use of AI. So if businesses want to use artificial intelligence to their own benefit, they will have to take this long and thorny path to data-driven management. Read this study to learn more on how global leaders are tackling this problem. 

Researchers from the University of California in Los-Angeles have released the world’s first Python library with computer vision algorithms that simulate light propagation based on the laws of physics. The library was named PhyCV. Unlike traditional algorithms, which are a series of hand-crafted empirical rules, this new class of CV algorithms relies on the laws of nature as the backbone for computations. PhyCV currently includes algorithms for Phase-Stretch Transform (PST), Phase-Stretch Adaptive Gradient-field Extractor (PAGE), and Vision Enhancement via Virtual diffraction and coherent Detection (VEViD). CPU and GPU versions of the algorithm are available. See the full library on GitHub. 

Engineers at Colorado University in Boulder have developed a walking stick with a built-in camera and ML algorithms to help people who are visually impaired better understand their surroundings and manage their daily routine. For example, the device can take pictures of the streets during a walk and use computer vision to decide what its owner should do and when, in order to arrive at their desired destination. The stick conveys this information to its user with voice or vibration. The cane uses the same technologies to read and speak out the ingredients of foods in a supermarket, or vibrates when approaching a certain grocery item. It can also guide users to their best choice of table in a cafe. This device will make the lives of visually challenged people much easier, but it will take time before they can actually buy it – engineers still need to make it more compact. See the Colorado University website for more information on the technologies used in the device and on some alternative applications for the smart walking stick. 

When it comes to large language models (LLMs), GPT-3 is the first thing you would name today. However, LLMs are much more than just that. The Indian Express calls out several other large language models that, even though not as popular as OpenAI’s brainchild, are certainly worth the notice for their highly valuable traits. For example, ERNIE Titan language model designed by China’s tech giant Baidu was trained to tell the difference between text created by a human and text generated by itself. This helps the model rate the credibility of generated content, which makes it more valuable in the eyes of consumers. Or take YaLM 100B by Yandex. It may not be as powerful as GPT-3 (100 billion parameters versus 175 billion), but it is available for free and can be used both for research and commercial applications. The model has been published on GitHub under the Apache 2.0 license. There is also Megatron-NLG created by NVIDIA and Microsoft with 530 billion parameters – three times more than in GPT-3. The list also includes Gopher by DeepMind and BLOOM by BigScience. Read this article for more on their strengths versus GPT-3. 

Researchers from Carnegie Mellon University have come up with yet another way to “see” through walls – using WiFi. They developed a deep neural network that matches the phase and amplitude of WiFi signals against the UV coordinates of body parts to determine a person’s posture just as well as other equivalent applications would do. This research paper offers an in-depth insight into the technology. It also describes how this technology paves the way to cheaper methods of tracking people without intruding on their privacy. The latter, though, raises some  doubts…