Back in October 2002, I appeared as a guest speaker for the Chicago (Illinois) URISA conference. The topic that I spoke about at that time was on the commercial and governmental applicability of neural networks. Although well-received (the audience actually clapped, some asked to have pictures taken with me, and nobody fell asleep) at the time it was regarded as, well, out there. After all, who the hell was talking about – much less knew anything about – neural networks.
Fast forward to 2014 and here we are: Google recently (and quietly) acquired a start-up - DNNResearch – whose primary purpose is the commercial application and development of practical neural networks.
Before you get all strange and creeped out, neural networks are not brains floating in vials, locked away in some weird, hidden laboratory - ala The X Files - cloaked in poor lighting (cue the evil laughter BWAHAHAHA!) but rather high level and complicated computer models attempting to simulate (in a fashion) how we think, approach and solve problems.
Turns out there’s a lot more to this picture than meets the mind’s eye – and the folks at Google know this all too well. As recently reported:
Incorporated last year, the startup’s website (DNNResearch) is conspicuously devoid of any identifying information — just a blank, black screen.
That’s about it; no big announcement, little or no mention in any major publications. Try the website for yourself: little information can be gleaned. And yet, looking into the personnel that’s involved we’re talking about some serious, substantial talent here:
Professor Hinton is the founding director of the Gatsby Computational Neuroscience Unit at University College in London, holds a Canada Research Chair in Machine Learning and is the director of the Canadian Institute for Advanced Research-funded program on “Neural Computation and Adaptive Perception.” Also a fellow of The Royal Society, Professor Hinton has become renowned for his work on neural nets and his research into “unsupervised learning procedures for neural networks with rich sensory input.”
So what’s the fuss? Read on,…
While the financial terms of the deal were not disclosed, Google was eager to acquire the startup’s research on neural networks — as well as the talent behind it — to help it go beyond traditional search algorithms in its ability to identify pieces of content, images, voice, text and so on. In its announcement today, the University of Toronto said that the team’s research “has profound implications for areas such as speech recognition, computer vision and language understanding.”
This is big; this is very similar to when Nicolai Tesla’s company and assets / models (along with Tesla agreeing to come along) got bought out by George Westinghouse – and we all know what happened then: using Tesla’s Alternating Current (AC) model, the practical development and application of large-scale electrical networks on a national and international scale took place.
One cannot help but sense that the other Google luminary – Ray Kurzweil - is somehow behind this and for good reason; assuming that we’re talking about those who seek to attain (AI) singularity, neural networks would be one viable path to undertake.
What exactly is a neural network and how does it work? From my October 2002 URISA presentation paper:
Neural networks differ radically from regular search engines, which employ ‘Boolean’ logic. Search engines are poor relatives to neural networks. For example, a user enters a keyword or term into a text field – such as the word “cat”. The typical search engine then searches for documents containing the word “cat”. The search engine simply searches for the occurrence of the search term in a document, regardless of how the term is used or the context in which the user is interested in the term “cat”, rendering the effectiveness of the information delivered minimal. Keyword engines do little but seek words – which ultimately becomes very manually intensive, requiring users to continually manage and update keyword associations or “topics” such as
cat = tiger = feline or cat is 90% feline, 10% furry.
Keyword search methodologies rely heavily on user sophistication to enter queries in fairly complex and specific language and to continue doing so until the desired file is obtained. Thus, standard keyword searching does not qualify as neural networks, for neural networks go beyond by matching the concepts and learning, through user interface, what it is a user will generally seek. Neural networks learn to understand users’ interest or expertise by extracting key ideas from the information a user accesses on a regular basis.
So let’s bottom line it (and again from my presentation paper):
Neural networks try to imitate human mental processes by creating connections between computer processors in a manner similar to brain neurons. How the neural networks are designed and the weight (by type or relevancy) of the connections determines the output. Neural networks are digital in nature and function upon pre-determined mathematical models (although there are ongoing efforts underway for biological computer networks using biological material as opposed to hard circuitry). Neural networks work best when drawing upon large and/or multiple databases within the context of fast telecommunications platforms. Neural networks are statistically modeled to establish relationships between inputs and the appropriate output, creating electronic mechanisms similar to human brain neurons. The resulting mathematical models are implemented in ready to install software packages to provide human-like learning, allowing analysis to take place.
Understand, neural networks are not to be confused with AI (Artificial Intelligence), but the approach employed therein do offer viable means and models – models with rather practical applications reaching across many markets: consumer, commercial, governmental and military.
And BTW: note the highlighted sections above – and reread the paragraph again with the realization that Google is moving into this arena; you’ll appreciate the implications.
But wait; there’s more.
From the news article:
For Google, this means getting access, in particular, to the team’s research into the improvement of object recognition, as the company looks to improve the quality of its image search and facial recognition capabilities. The company recently acquired Viewdle, which owns a number of patents on facial recognition, following its acquisition of two similar startups in PittPatt in 2011 and Neven Vision all the way back in 2006. In addition, Google has been looking to improve its voice recognition, natural language processing and machine learning, integrating that with its knowledge graph to help develop a brave new search engine. Google already has deep image search capabilities on the web, but, going forward, as smartphones proliferate, it will look to improve that experience on mobile.
So, let’s recap: we’re talking about:
* a very large information processing firm with seriously deep pockets and arguably what is probably one of the largest (if not fastest) networks ever created;
* a very large information processing firm working with folk noted for their views and research on AI singularity purchasing a firm on the cutting edge with regard to neural networks;
* a very large information processing firm also purchasing a firm utilizing advanced facial and voice recognition.
I’m buying Google stock.
What’s also remarkable (and somewhat overlooked; kudos to TechCrunch for noting this) is that Google had, some time ago, funded Dr. Hinton’s research work through a small initial grant of about $600,000 – and then goes on to buy out Dr. Hinton’s start-up company.
Big things are afoot – things with tremendous long-term ramifications for all of us.
Don’t be surprised if something out in Mountain View, California passes a Turing Test sooner than anybody expects.
For more about Google’s recent purchase of DNNResearch, check out this article:
To read my presentation paper on neural networks and truly understand what this means – along with some of the day to day applications neural networks offer, check out this link: