Instead of telling their company\u2019s artificial intelligence design how to learn, Igal Raichelgauz and his team took a novel approach\u2014they let a brain tell them how to build an AI.<\/p>\n
A rat brain, to be specific.<\/p>\n
CEO Raichelgauz and his fellow Cortica cofounders decided that the advanced, heavy computer technologies being used to recognize and quantify images and sounds were not only inefficient, but downright bad at what they were trying to do.<\/p>\n
Those techniques \u201cstill fail to get to the level of a biological organism or even a baby when it comes to the ability to recognize objects and their different variations,\u201d according to the Israeli entrepreneur.<\/p>\n
While the human\u2014or animal\u2014brain has been an obvious inspiration for AI developers since the first notions of computer intelligence, the Cortica team took a different approach with that influence right from the company\u2019s inception.<\/p>\n
Working with a still-living piece of rat brain, the crew had the gray matter on a dish that gave them an electrical interface with nearly every neuron of the tissue.<\/p>\n
This neuron-level input\/output arrangement allowed the entrepreneurs to read responses from their electrical inputs to single, or multiple neurons, and various combinations over time.<\/p>\n
\u201cWe started to play with it more like hackers, like engineers, trying to understand what it does before even answering the question of how it does that,\u201d Raichelgauz says.<\/p>\n
He and his early Cortica lab were reverse-engineering the brain.<\/p>\n
They found that with some adaptation, a network of neurons could create what Cortica calls a conceptual signature\u2014and could do so without supervision or training.<\/p>\n
\u201cImagine if you have two images of the same person,\u201d the CEO says. \u201cThe network will produce a sequence of numbers for each of the images and there will be a very high overlap between those two numbers, while if the images contained completely different objects the overlap would be zero, or minimal.\u201d<\/p>\n
So instead of the labor-intensive method of feeding images to an AI model and telling it \u201cthese are cats,” a new computational framework emerged where various cats were mostly similar but unique linear representations.<\/p>\n
\u201cIt is not dependent on what you’re interested in doing with it,\u201d Raichelgauz says. \u201cLater you can associate it with what we want to label it and so on, but it has a very unique internal representation which does not depend on the task or the object of interest.\u201d<\/p>\n
In this case, the neural networks themselves are the building blocks for representation, not systems to be trained with our preconceived and inflexible definitions.<\/p>\n
Equipped with the ability to create its own representations, this form of AI can learn more like a baby or child through context. But instead of from a crib, this young learner\u2019s stimuli all come from the vast existing and updated knowledge on the web.<\/p>\n
And while leaving a toddler alone would be considered bad parenting, the creators at Cortica see great advantages to their brainchild learning unsupervised. Crawling and exploring around a practically unlimited set of images and associated metadata, the technology needs no human in the loop for its learning.<\/p>\n
\u201cIt\u2019s not about trying to classify [images],\u201d Raichelgauz says. \u201cIt\u2019s more about visual discovery of the world and clustering and understanding what is what and then creating this association with the language.\u201d<\/p>\n
The outcome, a computational system his team designed, creates surprisingly accurate results, according to the cofounder. It deals with a couple billion potential numbers in a sequence, with any given image selecting only a hundred or two on average.<\/p>\n
However, developing impressive artificial intelligence alone isn\u2019t enough for the company.<\/p>\n
Now that their AI has grown up a bit, it\u2019s time for the Cortica team\u2019s next step: applying that framework in a way that is helpful to potential customers.<\/p>\n
The company has largely focused on image interpretation because the visual sphere is key in most engagements today for consumers.<\/p>\n
But the group has also fed their AI other information like sounds, financial data and industry-specific medical imagery. The firm is in the process of looking for the right marketing strategies to launch its first app, and Raichelgauz says that\u2019s on the very near horizon.<\/p>\n
\u201cWhen it comes to images, we want to create an experience using technology that\u2019s really different from search or browsing or things that were designed originally for web or text,\u201d he says. \u201cI think we\u2019re going to bring something very unique there.\u201d<\/p>\n