Worse, even if they recognize that privacy concerns are rampant among consumers, many in tech dismiss concerns like Zizi’s as irrelevant.
“In the tech industry they say ‘people like their privacy but they use the product anyway so it’s a fake problem.’ But people use the product because they have no choice,” Zizi explained in an interview. There are plenty of examples of people in the past grudgingly doing things in large numbers, and then stopping as soon they are given a choice, he notes.
Giving them a choice is the idea behind Aerendir, which seeks to deliver cloudless AI as a way around forcing people to share sensitive data through the cloud in order to establish the trust to carry out a transaction.
That sharing could become even more prevalent as the IoT comes online, and processing data at the edge can allay privacy concerns, but Zizi notes that deploying AI as it is normally conceived of to IoT devices is cost prohibitive, because of the GPUs required by neural network implementation.
“Most items in the IoT will cost less than $12,” he states. “It becomes an economic equation.”
It will still be possible to perform biometric identification and a wide range of other AI applications on such devices, but not in the usual way. Zizi points out that most AI systems rely on big data and image analysis, a task to which $30 GPU chips are ideally suited. Graphical databases have become standard, he says, due to the proliferation of GPUs in cloud data centers, and subsequent industry inertia. There is, however, another way.
“What if we dump graphics and go to vector?” Zizi asks. “That’s the secret of my company. We can get the same level of precision in terms of AI or whatever you want to call it, using vectors that are shorter, so we need less number crunching. We don’t need the power of neural net implemented for vision. We implement it just for digits.”
The principle is inspired by a technological throwback; televisions used to produce video by beaming a single line, or vector, across a cathode-ray tube. Given that the AI needs of most edge devices will be specific to a certain function or small set of functions, vector analysis on chips that cost a fraction of the amount that GPUs do will provide the computing power necessary for IoT devices to perform their tasks. Chips costing $2 to $12 can be used for dynamically trainable devices, according to company analysis.
That shift is just the beginning of a potential democratization of AI, however, as many advanced applications can be performed by systems that are not dynamically trainable. Systems can be trained in the cloud and then implemented at the edge, for instance. Some applications that are novel, and in some ways highly advanced, like determining a person’s age based on neuro-muscular data collected through micro-mechanical influence on device sensors, does not require a self-adaptable neural network, and can be implemented in ultra-cheap chips, costing as little as 2.7 cents, Zizi claims.
Aerendir is currently working with a vaping company on exactly that application, building an age-blocking capability into vaping devices that can make a determination of the age of the person holding it within a second. Zizi is enthusiastic about the project, though he is unable to name the company’s vaping-industry partner.
“This is a real society solution based on just the fact that we can have intelligent systems at costs that are low enough to be implemented,” he says. The solution may be ready in six months.
Other potential near-term applications include remote controls that can collect age and gender data, but no identifying data, for market research purposes, and USB keys and door handles without GPUs that can recognize fingerprints.
Not only can AI theoretically be extended to more places in the nascent IoT with vector-based processing, the less expensive chips can help device makers improve their profit margins as they do so.
This could enable Zizi to build the company he envisions one in which “droplets” of computing resources around us form an empowering “mist.” The mist is a bubble of control across a highly connected environment.
“The mist is what is around us. We all have a perimeter,” Zizi explains. He laughs at the metaphor, but he is quite serious about the potential for a new model of networking, which utilizes biometrics and other AI applications, but without the same risk to privacy and data security. The usual tech sector approach of dealing with those risks by adding a user interface layer to existing systems is simply not adequate, according to Zizi.
“We have an ecosystem that has breached data and personal information for ages, and has tried to put lipstick on a pig with UI,” he says.
Ultimately, what Zizi is talking about may seem disruptive in some ways, but it is a combination of leveraging existing computing technology and a new approach that is intended to help the sector realize an IoT future it has long planned for.
“By reshaping the dataset, that’s the key part, through vectorization, you can use vector processing units, which are the basic legacy computer chip,” he explains. “You can use one of them, ten of them, whatever power you need, but they’re cheap compared to GPUs.”
GPUs will still have their place at the edge and in the cloud, and existing biometric technologies will continue to be part of the identity ecosystem, at the edge and elsewhere. It is how the technology is delivered, not what the technology is, that is important.
“I’m not attacking other biometrics, because every single ID can be good, but it’s how you implement it that makes a difference for the user,” he insists.
That means databases do not need to be aggregated to make an engagement tool, he says. It means not commoditizing privacy.
This interview originally appeared on Biometric Update.
Aerendir | AI | artificial intelligence | cloud | edge computing | GPU | IoT