Monday, July 31, 2017

Is AI more Artificial than Intelligent?

Machine learning and Artificial Intelligence seem to dominate the tech media nowadays.  Lots of startups in the analytics field are being bought up by the tech giants.

There also is a lot of fear of AI taking over more and more types of jobs such as:



IBT recently had an article discussing a report put out by Stanford title Artificial Intelligence and Life in 2030 (it's a PDF).  The report breaks out various trends related to AI such as:

  • Large-scale machine learning
  • Deep learning -  (defined as convolutional neural network machine learning)
  • Reinforcement learning - a field of machine learning where the machine "learns" from it's mistakes.
  • Computer vision - a big user of deep learning
  • Natural Language Processing - Siri, Alexa, Hey Google, etc.
The report then goes into some of the great strides in AI that have taken place over the last few years.  For those interested in self-driving cars it had this summary:

During the first Defense Advanced Research Projects Agency (DARPA) “grand challenge” on autonomous driving in 2004, research teams failed to complete the challenge in a limited desert setting. But in eight short years, from 2004-2012, speedy and surprising progress occurred in both academia and industry. Advances in sensing technology and machine learning for perception tasks has sped progress and, as a result, Google’s autonomous vehicles and Tesla’s semi-autonomous cars are driving on city streets today. Google’s self-driving cars, which have logged more than 1,500,000 miles (300,000 miles without an accident), are completely autonomous—no human input needed. Tesla has widely released self-driving capability to existing cars with a software update.34 Their cars are semi-autonomous, with human drivers expected to stay engaged and take over if they detect a potential problem. It is not yet clear whether this semi-autonomous approach is sustainable, since as people become more confident in the cars’ capabilities, they are likely to pay less attention to the road, and become less reliable when they are most needed. The first traffic fatality involving an autonomous car, which occurred in June of 2016, brought this question into sharper focus.

It also talks about home-service robots such as vacuum cleaners and then gets into a current hotbed of machine learning - Health Care.  Could AI do some of things that doctors do now?

Looking ahead to the next fifteen years, AI advances, if coupled with sufficient data and well-targeted systems, promise to change the cognitive tasks assigned to human clinicians. Physicians now routinely solicit verbal descriptions of symptoms from presenting patients and, in their heads, correlate patterns against the clinical presentation of known diseases. With automated assistance, the physician could instead supervise this process, applying her or his experience and intuition to guide the input process and to evaluate the output of the machine intelligence. The literal “hands-on” experience of the physician will remain critical. A significant challenge is to optimally integrate the human dimensions of care with automated reasoning processes. 

So all of this is both fascinating and scary.   I don't think anyone can completely understand the long-term impact of AI technologies.  They types of tasks that we've been able to turn over to computers and other machines is, of course, staggering.  Who knows what they will think up next.  But, as anyone who has worked with computers knows, they don't think.  Machine learning is incredible, but machines don't think or learn like humans do.  They perform tasks, they mine data, they identify patterns, they monitor innumerable sensors and correlate data from them.  But they're not intelligent.

The Stanford report takes an optimistic view of the future of AI.  We're going to continue to try and find new uses of computers in our lives, it's happening whether we want it to or not.  But I think we'll find there are many things, that they just can't do.

Friday, July 28, 2017

Cameras and Computer Vision

Late last year, Intel bought the company Movidius and is betting big on Smart Cameras.  Here's something I found on seekingalpha about Intel's future plans:

Myriad 2 and Movidius are Intel's frontline assets to gain early lead in the fast-growing market for Computer Vision technology. Hikvision smart cameras are just among the many applications where Intel's Computer Vision processors/cameras could have long-term benefits.
CV-16 chart
I do not know how much Intel paid but buying Movidius Technology complemented Intel's RealSense 3D camera technology. A RealSense camera still needs the energy- efficient Myriad 2 to empower smart gadgets with Artificial Intelligence compute processing. Owning cutting-edge cameras and visual processors are important in Intel's overall Internet of Things strategy.


Our ViznTrac product is still in it's fledgling state, but we believe this industry is going to skyrocket.  ViznTrac's premise is basically to provide smart camera and IoT analytics to ordinary IP cameras.  I was asked recently how smart-cameras affect our vision.  I think it validates it and provides even more opportunity.  Currently, we're experimenting with Edge devices to enhance the security and usability of common IP cameras.  But, since we're coming at this from the analytics angle (correlate events from multiple devices both cameras and possibly other types), smart cameras are an asset to what we're trying to do.  Smart Cameras may eliminate the need for an edge device, but the intelligence and analysis that comes from managing and correlating input from multiple devices is still there.

Image Analytics will be Mainstream


I
mage analytics has exploded over the course of the past couple of years. The "Big Idea" that is grabbing all the headlines is self-driving cars. You've probably seen articles about Google's self-driving cars or Tesla's "auto-pilot" feature (if not, check out this video). As fascinating and futuristic as self-driving cars sound, the intelligent application of computer vision technology is taking off in kinds of fields and making its way into all kinds of practical business use-cases. Here are a few notable ones that have been mentioned in the tech media over the past year:
The use of IoT devices is continuing to grow, and camera-based devices are a huge part of that. Contextant's goal is to make this technology accessible and assist companies in providing pragmatic application of these technologies. This is what we mean by Applied Image Analytics. The idea behind Contextant's ViznTrac is make this exciting technology easily accessible to anyone.  With use-cases ranging from home-security using standard IP cameras to a broad range of industrial uses cases such as:
  • Warehouse Floor Layout Optimization
  • Counting people in a checkout line at a store
  • Count people entering and exiting a location
  • Use facial recognition to implement better security or kick off automation tasks
  • Use either bar-code or OCR to read labels on packages moving in a warehouse.
  • Home security and smart-home automation
It can be bewildering trying to keep up with all the advancements that have taken place recently in this area. What we are seeing is the result of the maturing of image processing technologies and robust machine learning platforms such as Google's TensorFlow coming together to make these technologies easier to use and more accessible to individuals and companies than ever before. The result is the explosion of use-cases across industrial verticals for companies of all sizes.
Our ViznTrac platform is currently in beta, with a planned Kickstarter coming in September.  With everything that has happened in this space just over the course of 2016 and thus far in 2017, it's exciting to think of everything that will happen over the next year. We believe that ViznTrac will allow both individuals and companies to more easily take advantage of this technology using hardware (such as IP cameras) that they already have.
If you're intrigued about applied image analytics and are curious about how ViznTrac could help you or your business, please drop us a line and somebody will contact you and answer any questions you have.

Thursday, July 27, 2017

How Machine Learning enables alternative User Interfaces



The way we interact with computers and computer-connected devices has come a long way.  In the beginning, computers were number crunching devices and people used punch-cards to load the data into them to crunch those numbers.  Then the command-line interface was introduced and it became a dominant (and still highly useful) interface.  Via the keyboard, you instructed the computer to do exactly what you wanted.

Then, in the late 80s we had the graphical user-interface revolution.  This is where the mouse,  a movable pointing device, became a key part of the user interface.  A command-line interface is focused on precision.  You must spell out the instructions that you want the computer to perform in fine detail.  Overtime, macros and other short-cuts were introduced to make it easier to type out frequently performed items, but the instructions were always very precise.  A mouse and GUI are very different.  You are now interacting in a three-dimensional space (even if it was only rendered in 2D with overlapping windows).  Distance became important as we would physically move items around on the screen in order to interact with them.  A mouse, although not as precise as the command-line, is still a very precise device for interacting with the computer.  We just had to rely on the OS to keep up with which XY position the pointer was at, which screen element was directly underneath it, which elements were overlapping each other, etc.

In the mid-2000s, touch interfaces became more relevant, with the 2007 iPhone being the break-away hit to popularize the touch interface.  The touch interface is a fairly mature interface now, although it is still not 100% complete.  It is something that is still being experimented with, such as utilizing pressure to enable new interactions with a device or computer.  A touch interface, although still precise, is not quite as precise as a mouse is.

Finally, over the last few years machine learning technologies have evolved to the point where they are becoming part of most software systems.  In business, we mostly think of ML as performing sophisticated analytics and predictions on data from our operational systems or IoT devices.  But it is also enabling the development and use of other types of input devices to directly interact with our software applications.  The key inputs that are becoming more dominant as alternative user experiences are based on sight and sound.

Voice has currently taken the lime-light as a key user-interface technology, with Apple (Siri), Google (Google Home) and Amazon (Alexa) fighting to become a dominant voice-based technology.  Wired recently had an article :  VOICE IS THE NEXT BIG PLATFORM, AND ALEXA WILL OWN IT .  In it they declared:

In the coming year, the tech that powers Amazon’s assistant will become even more robust. “We’ve made huge progress in teaching Alexa to better understand you,” Amazon’s head scientist and vice president of Alexa Rohit Prasad told Backchannel earlier this month. Amazon is making more tools available to developers. In November, the company announced an improved version of its developer tools, the Amazon Skills Kit. The company also announced improvements to the Alexa Voice Kit, the set of tools that allow manufacturers to incorporate Alexa into third-party devices like cars or refrigerators.

Machine Learning is the key technology backing voice-systems since they must learn and improve over time to recognize different  accents, different word orders, etc.  Although Apple, Google and Amazon get the majority of the press attention, they are not the ones provided voice interfaces and voice can be used in surprising ways.  My company, Contextant, specializes in using machine-learning technologies to help companies improve their business processes.  Currently, we're in the process of integrating voice-recognition technology with one our client's warehouse management system (WMS) to enable people in the warehouse to perform picking, inventory movement, receiving and packing operations via voice.  This frees their hands from having to worry about a keyboard or mouse.  Voice isn't perfect, but it can complement a standard GUI to help workers see what to do next, where to go inside a warehouse, etc.

Computer Vision is also an exploding machine-learning based technology that allows for alternative user experiences.  It has a come a long way in just the past couple of years, but is still not as mature or precise as the other technologies I've mentioned.  But the future is very bright for Computer Vision.  Our own ViznTrac service was developed out of an indoor-location tracking system to allow large facilities to keep track of where people are.  It is also a big-part of home-security systems (a market area that we hope to push ViznTrac into shortly) and with the development of facial-recognition, package detection and more sophisticated movement detection algorithms it will only improve.

These alternative UX features are still evolving, and they not as precise as traditional methods.  Voice systems can misunderstand what you say and computer-vision systems misinterpret what they are "seeing".  They won't replace the other forms of input (I'm using a keyboard now!) but they will continue to become more effective and complementary technologies.

Monday, June 26, 2017

Jobs of the Future



The Impact of AI and Machine Learning on our Jobs

Over the past year there have been an  increasing number of articles written about jobs that can be done by a machine versus a
person.  I tend to be pretty optimistic about the future, but I don't believe anyone can know how the nature of jobs will be transformed as automation is introduced into various aspects of life.

Here's an article from Fast Company that appeared 3 1/2 years ago about the changes coming from machine learning and artificial intelligence technologies.  I don't think it's aged well. Here are three things it listed:

1) Unstructured problem-solving: solving for problems in which the rules do not currently exist. Examples: a doctor diagnosing a disease, a lawyer writing a persuasive argument, a designer creating a new web application
When looking at the examples given, I think diagnosing a disease is probably the first one to be taken over by a machine.  But, depending on how you frame the problem, an AI could be written to determine which styles of argument may be most likely to succeed in certain scenarios.  Neither eliminates the need for a doctor or lawyer, but they are definitely tools that could reduce the time spent by the primary doctor or lawyer.  This could have significant downstream effects on the number of doctors, lawyers, research aids, paralegals, etc.  AI and machine learning are tools.  By themselves they don't do what a person can do, but they change the nature of the work, and the downstream effects of that are unknown.
2) Acquiring and processing new information, deciding what is relevant in a flood of undefined phenomena. Examples: a scientist discovering the properties of a medicine, an underwater explorer, or a journalist reporting on a story.
Here is an example of where some forms of journalism aren't safe from automation.    
3) Nonroutine physical work. Performing complex tasks in 3-D space, from cleaning to driving to cooking to giving manicures, which is thought of as relatively low-skilled work for humans, but actually requires a combination of skill #1 and skill #2 that is still very difficult for computers to master.
This is true, because we constantly break tasks down to their simplest essentials and teach machines to do that one task.  Then we start layering automated processes on top of each other.  Machine learning is all about determining what data is relevant.  A person sets the parameters of what kind of data is relevant, but machines are best at determining the specific data that will be treated relevant.  I think the rise of self-driving cars is a great illustration of this.  It's been true for a while that machines are better at routine driving than people would are (machines don't daydream or get distracted), but people are better at the unexpected (icy patch on a road).


Machines and the Future of Work


AI will change the face of the workplace.  The problems that you'll find with articles that try to predict what jobs will be safe, is that they think of the jobs as they currently exist for humans.  But, AI won't do the exact same jobs as people do, they'll do different jobs that will in turn change the nature of the jobs that people will be needed for.

Josh Wood (a tech recruiter), makes similar points that I was trying to make here and here.   Namely, that we can't really predict what jobs are safe nor that all the jobs will be automated.  It is the downstream effects of machine learning and AI on the economy that are utterly unknown.

He makes a statement that automation may be overblown:
After all, had the Blacksmiths Unions of the early 20th century been given the power to do so they would have outlawed the motorcar. My point here is that for every piece of ‘job-destroying technology’ there is a new job-creating industry being born which can never be properly foreseen. Winner of the Nobel Prize in Economics Milton Friedman accurately pointed out that most economic fallacies derive from the ‘zero-sum fallacy’ – the idea that if one party gains another party must always lose, when in fact provided all parties consent to a change all must be gaining (Otherwise, said party would not consent). The entire IT industry as we know it today would have been unthinkable to even the most prophetic sci-fi writers fifty years ago (We don’t all seem to wear identical silver clothing, either…), just as the number of Application Developers working today would have been unfathomable to leading technology economists writing in 2006.
The tech industry is huge part of the economy, and it is an industry where jobs are constantly being created that didn't exist before.

Things are going to change as AI and other ML techniques make their way through various forms of business and the truly beneficial uses of these technologies are found.  Change is constant, and the creation and destruction of the types of jobs that need to be done will continue.  At Contextant, a lot of our current work is focused on improving physical processes at warehouses.  The goal is to simplify the management of the work and try to make the work being done more efficient.  But what are the long-term down stream effects of these changes?  That is what is unknown. 

If something is completely unexpected or has never been considered before, then a machine won't make the same kind of value judgements that a person will.  Technology has always allowed us to build machines and devices that replace tasks or jobs that used to be done by hand.  AI and Machine Learning are no different.  They are tools that will transform what needs to be done.  But they are still just tools.  Machines will never be human.




Friday, June 23, 2017

When AI Can Transcribe Everything

We've done a little work with a transcriptionist client. Here is an oddly relevant article in the Atlantic concerning the promise of rapidly improving automated transcription technologies.






Across professional fields, a whole multitude of conversations—meetings, interviews, and conference calls—need to be transcribed and recorded for future reference. This can be a daily, onerous task, but for those willing to pay, the job can be outsourced to a professional transcription service. The service, in turn, will employ staff to transcribe audio files remotely or, as in my own couple of months in the profession, attend meetings to type out what is said in real time.
... the promise of ASR [Automatic speech recognition] programs capable of accurately transcribing interviews or meetings as they happen no longer seems so outlandish. At Microsoft’s Build conference last month, the company’s vice-president, Harry Shum, demonstrated a PowerPoint transcription service that would allow the spoken words of the presentation to be tied to individual slides. 

It’s difficult to predict precisely what this new order could look like, although casualties are expected. The stenographer would likely join the ranks of the costermonger and the iceman in the list of forgotten professions. Journalists could spend more time reporting and writing, aided by a plethora of assistive writing tools, while detectives could analyze the contradictions in suspect testimony earlier. Captioning on YouTube videos could be standard, while radio shows and podcasts could become accessible to the hard of hearing on a mass scale. Calls to acquaintances, friends, and old flames could be archived and searched in the same way that social-media messages and emails are, or intercepted and hoarded by law-enforcement agencies.

Monday, January 16, 2017

AI is a tool to be used.

Continuing with the theme "Jobs of the Future Don't Exist Yet" is this article about statements from Microsoft CEO Satya Nadella :

"The fundamental need of every person is to be able to use their time more effectively, not to say, ‘let us replace you’," Nadella said in an interview at the DLD conference in Munich. "This year and the next will be the key to democratizing AI. The most exciting thing to me is not just our own promise of AI as exhibited by these products, but to take that capability and put it in the hands of every developer and every organization."
Nadella is pushing Microsoft into consumer and industrial applications of software that can make inferences about its environment. He cautioned his own company and competitors about "parlor tricks" that show AI’s power without preserving workers’ dignity.

Thing are going to change as AI and other ML techniques make their way through businesses and truly beneficial uses of these technologies are found.  But it's going to change and create the types of jobs that need to be done.  A lot of our current work is focusing on improving processes at warehouses.  It simplifies the management of the work rather than the work itself.  Down stream effects, that is what is unknown.