The article is a little dated. It was written in the wake of Intel acquiring Nervana and Movidius, which it used as a hook to talk about Wave Computing. Wave was founded a few months before Nervana in 2014, and raised a similar amount of money (~$24M), which is just enough to get to your first chip if you don't waste resources. There are other companies tackling this (Cerberas Sys.) and other technologies that can get you acceleration (FPGAs).
Handwritten digits / OCR was one of the first practical uses in the late 80's / early 90's for CNNs (check reading, addresses, http://yann.lecun.com/exdb/lenet/). Integrating an LSTM would be useful for providing higher level sequence recognition.
I know just when my bank "upgraded" their check reading software last year. I was taught to write checks in the form:
$1234 66/100
in the amount column. I've done this for decades. Suddenly, my account was only debited $12.34, and I received dunning letters, interest, and penalties. Showed the bank the check image, they fixed it. (They also admitted that the OCR software made no attempt to read the handwritten amount, nor the signature.) I thought it was a fluke, but it happened again for the the next two checks.
So I started writing the amount as:
$1234.66
and things started working again. So no, I am not impressed with the OCR used to read checks, and it is clearly not remotely ready to read general handwriting.
I get tired of the fees charged for electronic transfers. It should be the other way around. Many places will tack on 3% if you use a credit card. Paypal is what, 1.5%? Wiring money is expensive, Western Union is even worse. Added up over a year, all those "convenience fees" and crap can be a hefty bill.
Checks still have zero transaction costs for me and the depositor. When people claim I didn't pay them, it's nice to show them the cancelled check with their signature on it.
When bargaining with someone, showing them a signed check made out to them can clinch the deal :-) Of course, cash is even more persuasive, but I don't care to carry around cash and again, I like having a cancelled check as a receipt.
I am not criticising you for using checks. As you rightly point out electronic transfers should be the no fee option - the idea that pushing around bits of handwritten paper is cheaper than electrons is crazy.
My bank wants me to use electronic billpay, but they want to charge for it. I say no thanks, you can keep dealing with my free paper checks which cost you more.
I've been using ATMs since 1979, and I picked a bank that did not charge for using it. It's just nuts to charge for an ATM when using their tellers for free costs them far more.
I recently tried to do this with Acrobat Pro with varying degrees of success, you might want to give that a shot if you can get access to the program (slightly cost prohibitive).
I wish I could try various ones, but as you say, cost is an issue.
I do have some OCRs, and while they do a credible job with printed text (95%), they fail miserably with typewritten letters (75%) (the blotchiness of a mechanical typewriter throws it off) and fail 100% with handwritten letters, even carefully done block letters.
I know there is success with domain specific handwriting recognition, like envelope addresses, but nothing that can handle a cursive letter.
You'd think with all the facial recognition progress this should be a much simpler problem.
(To do a good OCR, it'd have to examine the context of a muddled character to determine what it should be, just like a human does when reading cursive. This should be an ideal candidate for "deep learning".)
While I think your situation is unique, it would be nice if there was an out-of-the-box ML system that could be trained by individuals for any specific use case (such as transcribing letters).
In the future, I think this might be the case where you "teach" a computer or program to recognize and do tasks the same way you would a child or pet.
PS- Transcribing letters brings to mind an app on the iPhone that I used to use for instant language translation. Can't quite remember what it was called but Google ended up acquiring them for their OCR technology.
It is possible that handwriting recognition is not an attractive target for AI research because few handwrite things anymore. I heard that cursive is not even taught anymore!
That leads one to suspect that one can private communication by simply writing in cursive!
Smarter memory design means that data does not need to be constantly reloaded, saving time that would otherwise be wasted on data transfer.
Using lower-precision, fixed-point arithmetic is more efficient than standard-precision floating-point arithmetic. As a result, less time and components are wasted on unnecessary precision.
Someone more knowledgeable about processors can better describe the benefits of the Wave architecture.
In terms of benchmarks, Wave claims[1] that a single Wave machine is 25x more efficient than 100 GPUs for Google's Inception network.
Wave seems to be referencing this[1] Google Research blog post. I couldn't find any reference to what GPU cards Google uses. It still provides a basis of comparison to Google Cloud at least. But like you say, a specialized data center deployment probably wouldn't show such a dramatic difference.
Out of the ones mentioned in the article Movidius was/is specifically focused on low power vision tasks, the rest are fundamentally same stuff as we have today with CPUs/GPUs, just faster, lower power.
The GPU is hard to beat for training purposes unless you're Google scale, but not acceptable for some of these lower power inference scenarios. You can't reasonably make a product with a GPU today under 10W.
