Dated January 2, 2018
How will applications of machine learning change in the next 10 years
What are the shortcomings of machine learning
What language do you think is best to write a machine language application in
How can machine learning differentiate between a recorded voice and a real one
In what ways does Pindrop’s audio machine learning differ from conventional forms of ML
How is the Atlanta tech scene different from Silicone Valley
A month or so ago I found a discussion on Machine Learning on Quora. Are you familiar with Quora? It is a cool app. You pose a question to Quora requesting an answer, Quora puts the question out in cyberspace and whoever wants to answer can do so. Interesting to spend so time on.
The above questions were answered by Professor Vijay Balsubramaniyan PH.D. at Georgia Tech University in Atlanta Georgia. The first question he addresses is:
How will applications of machine learning change in the next ten years
Answer requested by Ashish Ranjan, Gyanendra Kumar, and 94 others
His answer struck me because I had not heard that this was the plan or I just wasn’t paying attention. But it makes great sense. He said that the future of ML or any technology makes it more human. I just heard an ad for Google. They claim that their computer voice for answering machines is indistinguishable from a human voice. That will be interesting to check out.
Machine learning he says is getting close in specific areas such as image recognition and classification, speech recognition for self-driving cars. “For example, Googles’ word error rate for speech recognition has gone from 30% in 2012 to 5% in 2017. That’s the difference between understanding 13 words out of 20 (2 sentences out of 3) to 19 words out of 20.” The difficulty is the last mile (95% to 99%).
He indicated that researchers expect to make significant strides in the next 10 years. Now that is the comment of a researcher. He didn’t say they were expecting to solve the problem in 90 days, 365 days but in terms of years. And quite a long time also. An interesting goal that they have also “computers will make our everyday tasks easier at home, at work, everywhere we go.”
What are the shortcomings of machine learning
Answer requested by Gyanendra Kumar, Hasha Ramanagoudra, and 44 others
Machine learning has several shortcomings he admits. First ML is only as good as the quality of data used. The old “garbage in, garbage out.” As everything gets more voluminous and complex the challenges increasing can lead to false conclusions such as ‘seeing what you want to see.’
He states that “ML helps technology become more human but it is necessary to guard against it absorbing our deeply ingrained human biases.”
“The ML model that you are working with needs to be critically evaluated using traditional experimental methods. Its strengths and weaknesses have to be identified. All of this is more challenging than developing the model in the first place.”
A Short Aside
Vijay and two other PhDs founded a company called Pindrop Security. They serve the global marketplace by phone anti-fraud and authentication technology. They primarily work with Call Centers. Their applications provide audio analysis technology analyzes 147 different features of a phone call. In 2015 Pindrop screened more than 360 million calls and raised $122 million in funding. In 2017 they protected 410,826,795 phone calls, detected 401,926 fraud calls, generated $200,500,000 in savings for their clients. The company’s mission statement is: “Our Mission is to provide security, identity, and trust in every voice interactions.”
What language do you think is best to write a Machine Learning application in
Answer requested by Jamie Corkhill, Harsha Ramanagoudra and 79 others
According to IBM research, Python is the most popular language currently used for Machine Learning. It has flexibility, an abundance of strong ML packages like stikit-learn, and they can easily replace critical routines with C/C+ when necessary.”
He did suggest keeping an eye on Golang which is gaining in ML support and popularity.”
How can machine learning differentiate between a recorded voice and a real one, given that the quality of the recorded voice is fairly good
Answer requested by Pushkar Sharma, Hammad Arshad and 9 more.
“Even the most realistic recorded voice that sounds legitimate to the human ear retain some trace features that are different from the live voice of the genuine speaker. These features are often emphasized (example played back in higher volume) while analyzing the spectral characteristics of the signal.”
Thoughts on Adversarial Machine Learning
Answer requested by Amit Jadhav, Praveen Krishna and 14 more
“Adversarial machine learning is a particular class of online machine learning that is intended for computer security.”
While typical (offline) machine learning tools use fixed training data and assume that the unseen test data follows the same distribution as the training data. Adversarial machine learning tools are continuously adapting to the ever-changing distribution of data.
This is often the case in authentication systems where malicious impostors keep looking for new vulnerabilities to defeat them.
“Adversarial machine learning is very important at Pindrop. Fraudsters keep changing their attack techniques and are becoming smarter over time.”
“Our fraud prevention system is in a continuous evolution to keep catching new kinds of fraud activities over the voice channel.”
We have heard that refrain several times lately. Elon Musk has proclaimed it, Steven Hawking has warned us of the same.
So Trump portrays Mexico and China and global trade as the enemies. What is he talking about?
Back on January 26 of 2017, he declared: “It has been a one-sided deal from the beginning of NAFTA with massive numbers of jobs and companies lost”. Hmm
President Obama stated “The next wave of economic dislocations won’t come from overseas. It will come from the relentless pace of automation that makes a lot of good middle-class jobs obsolete.” He offered these words of caution during his farewell address.
Research supports Obama’s claim. Far more jobs are lost to robots and automation (better technology) than trade with China, Mexico or any other country.
America has lost jobs to trade, but robots are the big threat. Any job that is repetitive in nature and even more so if it is hazardous to worker’s health are in danger of being replaced by robots.
In Manufacturing, nearly 5 million jobs have been lost.
Real Numbers for Real Issues:
U.S. Trade with China killed 985,000 American manufacturing jobs between 1999 and 2011
per MIT professor David Autor
U.S. Trade with Mexico cost roughly 800,000 jobs between 1997 and 2013
per Robert Scott, Economic Policy Institute
That might sound high. But last year alone the US added more jobs than those losses combined. And remember Mexico imports 40% of its goods from the United States.
Additionally, two Ball State professors found that between 2000 and 2010 about 87% of the manufacturing job losses stemmed from factories becoming more efficient.
The main driver of more efficiency? Automation and Better Technology.
J. Bradford Jensen, an economics professor at Georgetown University stated “There has been a lot of technical change that has reduced the need for labor – some of it is automation, some of is design, more software, less hardware.
So why not crack down on robots?
“It’s harder to demonize what everyone sees as technical progress. It’s easier to demonize the foreigner,” Jenson added.
It’s hard to really get solid numbers when talking about automation because it also creates jobs. ATMs are a case in point. They perform jobs that bank tellers one did. But there isn’t much evidence that bank employment tanked as a result of ATMs. MIT professor Daron Acemoglu
One last point to consider. Mexico imports goods. 40% of those goods are American.
Editors note: I recommend reading a post I wrote New Year’s weekend. It is on my blog “dabblerducksbutt.com” The title is Capitalism Overreach.
It is a story of coal mining in West Virginia. As you know the miners of West Virginia did not get a chance to enjoy the fruits of the twentieth century. Hopefully, we will use this example to ensure that no family, no community is left behind in this Fourth Industrial Revolution.
If the job loss rate is even close to 47% the pain felt by the middle class and working families is going to be unimaginable. There won’t be other jobs to go. Nowhere to live.
I have found so much more material that I am going to digest and file another post shortly after this one.
Thanks for Stopping By!