Hacker News Reader

* Disclosure: I work for Google with the OP, and sit next to him in Japan.

"2. Google dispatched their solutions architect to work with them."

That's not what happened at all. He built the system entirely himself before getting in contact with us.

I think that while very complex machine learning tasks do require specific knowledge, creating a neural network that can do real work is within the abilities a single individual.

ktamura is not saying that happened, is saying that if that happened it would be more real than the current one. The current one (while true) looks like a TensorFlow advertising, making it sound easier than it really is.

Sure. It's just that we can't make it more genuine or plausible by writing it that way because it's not true.

>That's not what happened at all. He built the system entirely himself before getting in contact with us.

Good to know. While what I said was meant to be hypothetical, I am happy to stand corrected.

>creating a neural network that can do real work is within the abilities a single individual.

This really depends on said "individual." My biggest issue was the tacit conflation between two themes, one valid and another reeking of "feel good" marketing:

1. That a reasonably technical person can use neutral network to do useful things, no small part thanks to something like TensorFlow (valid)

2. That such individuals are prevalent in agriculture (???)

I am all for hero-making: it's at the heart of marketing and customer advocacy. However, as someone who has been doing technical marketing for awhile, I just find this story exceptional in both senses of the word (as others commented, perhaps I turn out to be wrong!)

That's fair. As the reader you can be the judge.

Advocacy is partly about inspiring others and I think this does that while presenting an accurate representation of what the farmer was able to accomplish.

What in the article made this sound like a complex application to you? It seemed to me like any number of vanilla NN's could have done the job as simple image classification tasks like these have been done successfully for decades. TensorFlow was chosen because he found out about it via Alpha Go, so Google's marketing is doing it's job. Given the man's embedded systems background, it seems entirely plausible that he could have designed and implemented the entire solution himself. (for most people, the 'hard' part of integrating the NN into the real world via the rpi would likely have been relatively easy for him given his experience.) While I generally tend to be cynical about most success stories, this one actually sounds pretty reasonable.

> I am not trying to be a hater. Deep learning has huge potential, and Google, among others, is doing a lot to make it accessible. That said, this type of over-promising ahead of market reality is what gives cutting edge technology a bad name.

Actually, this type of marketing is _exactly_ what is needed to get people interested in applications for this technology.

This particular farmer, perhaps, is an unusual case in that they went straight to machine learning with sexy new tools. However, machine vision has been used for applications in agriculture even as long as 20 years ago (I know this personally :-) )

It isn't at all surprising that a farmer decided to go "cutting edge" with a little help from Google. In fact the theory for the "diffusion of innovation" was based largely on observations of FARMERS!!(https://en.wikipedia.org/wiki/Diffusion_of_innovations). What I am saying is that although farmers seem to have an undeserved reputation for being low-tech, some are _very_ adept at technology especially when it has the potential to give them an advantage in the market.

Man, people just can't win on HN.

Just the other day on HN: people use excel because they don't have the elite mental reasoning ability that programmers have. Programmers are so awesome, it's cute that factory foreman use excel when they should be using my stellar programming skills for some hand wavy reason like "it's hard to tell what you did and reproduce it."

OK. Point taken.

Now here the sentiment shifts to the other extreme. People shouldn't go off and try to use cutting edge technology without the skilled supervision of an expert. Really, I would take this story better if Google were involved because they have a monopoly on the world's brain power.

---

I get that the common thread in both of them is that a programmer should be involved, and I get that I was overtly sarcastic.

But can't we just judge based on merit? Why not critique his approach, why not suggest what you would have done differently, or something to that effect.

Why do we immediately turn it into smarter than everyone programmer vs the idiot common folk.

Yes, thank you. I'm glad it's not just me noticing that the tone of HN commentary these days is increasingly reminiscent of freshman CS with That One Kid Who Knew Code Before College.

> these days https://news.ycombinator.com/item?id=9224

I read the parent comment completely differently. I don't think he was saying that non-technical-people are worthless and shouldn't try to do anything. I think he was saying that the tools are difficult to use in their current state. Perhaps in the future they will be easier to use and more available.

Similar to programming. Programming languages are difficult to get and learn to use. See e.g. my test here on how difficult it is for nontechnical peopel to even install python: https://news.ycombinator.com/item?id=11453086 Whereas Excel is widely available and taught in schools.

> in reality, much expert help is needed to make it work.

I totally agree. Having been taking classes and read ML/CNN articles online, this Google Cloud blog post made this project seem to be significantly easier than the actual effort that went into it. In particular, the blog simply covered the secret sauce to make it work in one line:

> Makoto used the sample TensorFlow code Deep MNIST for Experts with > minor modifications to the convolution, pooling and last layers, > changing the network design to adapt to the pixel format of cucumber > images and the number of cucumber classes.

I doubt this is a one man operation in just a few months of time, considering the computing power that's required to accomplish such feat (which it was said to take 2-3 days for each complete training set). Let alone the fact that they admit 7000 cucumber images are NOT enough to train the model to achieve the typical 95% accuracy that the original MNIST model it was based off.

Although promising, this blog post reads more like a PR campaign for GCP's ML offering (with the custom hardware devices to do the compute, perhaps TPUs?)

I just wish they'd move their CloudML thing out of alpha, I've given AWS too much of my money for their shitty GPUs.

What will the pricing for CloudML be? It will be difficult to bootstrap anything without a free tier of some sort.

I would be super surprised if it wasn't at least cost competitive with AWS/Azure GPU instances.

They didn't approve my request for alpha access though, so no real insight on my end.

What can you guess are their criteria for approval? I applied as well.

Probably how likely you are to spend a significant chunk of change on their cloud offering

I disagree. One of the first projects I tried in deep learning was to tweak the LeNet-5 style MNIST example to handle 64x64 color images. It was pretty straightforward. Pretty straightforward doesn't mean it took ten minutes; but it didn't take more than a couple of hours, for someone clueless about DNN models.

If you look at: https://github.com/tensorflow/tensorflow/blob/master/tensorf...

you could actually manage it just by changing some of the image size and classes constants. In earlier versions of that code, you'd have had to manually change a few constants that weren't extracted well. This is approximately easy, once you know your way around the python side of TensorFlow. (Which itself takes some work, but it's not rocket science.)

A better approach would probably add a third convolutional layer to continue to shrink the 80x80 input down to something easier to handle in the fully connected layer.

What's surprised me is actually how slow it is, knowing nothing about their model. My guess is they could speed it up by adding that extra convolutional layer to reduce the amount of work done in the final FC layers. The fact that they're that slow suggests the opposite of what you're noting -- I doubt they had serious assistance on the machine learning side. I'm more impressed by the amount of time it must have taken to get the physical parts working and to collect 7,000 training images!

Training MNIST itself is trivial. You can get > 99% accuracy in only a few minutes of training using a single GPU. This model should be at most 10-20x slower than MNIST (but, again, that FC layer would scale badly).

If you look at the stackoverflow tags for tensorflow, you'll see a lot of very similar problems being solved. ("I want to reuse model X on problem Y, where such and such is different").

TPUs are awesome, but you don't need them for this problem. You could run the inference at 1 cucumber per second on a laptop. And you could train it on a Titan X in a matter of hours at most.

(Source: I was and am still slightly on the Google Brain team, as a non-machine-learning person.)

Hardly.. the guy is only getting 70% accuracy. The hard part is the assembly line and he probably has a background in that.

Yeah I wish there was a library that people with a casual interest could use to just mess around. Or even just use deep-learning based tools built by other people.

I remember when the deep dream thing came out, there was a huge interest by people wanting to experiment with it on their own, or just use the existing code on their own computer. But installing these libraries is very difficult or impossible, particularly on Windows. People were trying to get it to run on virtual machines, which adds complexity and slows it down, and can't take advantage of GPUs which is necessary. They don't work on non-nvidia GPUs, etc.

The best example I found was Brain Simulator (http://www.goodai.com/#!brain-simulator/c81c). Which ran on Windows, had a GUI, and was intended to eventually be used in the game Space Engineers so nontechnical users could experiment with AI algorithms and build robots for the game.

I may write something like you have described. It would use OpenCL rather than CUDA. Though I would like it to be scalable, the starting goal will be simplicity and ease of deployment on a range of devices.

Do you have any suggestions for the API?

If you wrote some simple code that would work if there was a functioning learning module, what would it look like (if you feel like contributing to this)?

I am not an expert, but in my brief experimentation, I am a fan of the Torch API. It's focused on modules that you can easily build and combine together in very open ended ways. I have no idea what the best method would be though.

> That said, this type of over-promising ahead of market reality is what gives cutting edge technology a bad name.

Change "deep learning" to "expert systems" and "TensorFlow" to "decision trees" and that article could have been written in the 1980s: https://www.youtube.com/watch?v=6xpcES-ueKw

"Cucumber classification" seems almost identical to the letter classification task which is the beginner's tutorial on Tensorflow. I think anyone who knows Python and completed a Coursera or Udacity course could get something implemented.

> "Cucumber classification" seems almost identical to the letter classification task

Why do you say that?

You're looking at the shape of an object, and determining which category it belongs to. It is similar to looking at a handwritten "a" and determining if it belongs to the "a" category.

"Even with this low-res image, the system can only classify a cucumber based on its shape, length and level of distortion. It can't recognize color, texture, scratches and prickles,” Makoto explained. Increasing image resolution by zooming into the cucumber would result in much higher accuracy, but would also increase the training time significantly.

That seems incredibly reductionist, and you might as well say they're the same because "you add a bunch of numbers together and see if the result is bigger or less than zero."

(Also how do you know or why would you expect that it's categorizing by shape? Neural networks are infamous for finding "irrelevant" features in training data).

I don't understand your objection. Letter recognition takes a bitmap, and sorts it into one of 26 categories, cucumber grading takes a bitmap, and sorts it into one of 9 categories.

Of course it's reductionist. Generalizing an abstraction across many things that superficially look different is what programmers do. Reducing problem complexity to make something easier to automate is what programming is for.

Saying cucumber grading is like letter recognition is a true generalization that is also useful. Saying that they're both "adding a bunch of numbers" is a true generalization that is less useful. Programming, and (way more) broadly, science and engineering itself, is about finding useful generalizations.

"Cucumber classification"can be solved easily with a C3NN (Cucumber Classification Convolutional Neural Network).

Not to mention that if this solution took off, someone - the farmer, Google or an existing sorting-device company - would market it to most other farmers.

So most farmers won't be tweaking TensorFlow any time soon even once the product becomes mature - well, they might wind-up using a high quality UI for customizing said existing solution.

You'd be surprised... It's like an activation energy in chemistry. Just because two chemical reactants have enough energy to react doesn't mean it's likely to happen. You often need a catalyzer or have many many possible reactants to play the numbers game. Right now, this kinda of small scale farming is right at the "enough to react but not quite ready to chain react" phase. Computational costs need to fall some more to increase possible pool, but it'll happen soon likely.

This was a marketing piece for Cloud ML. It didn't seem like any overpromising occurred in the blog post. Saying that a commercialized out of the box product that has feature X which does not have feature X is overpromising. Promoting a developer platform with early adopter use cases seems like a valid thing to do, which is all they did here. Did I miss something?

Not at all. Admittedly I'm a strong developer, but very little knowledge of Deep Learning. I picked up tensor flow in a month.

Also, the guy is getting 70% accuracy. I doubt google helped him :p

Of course it's a marketing stunt but it also depicts one of the endless possibilities of deep learning. So I don't think its over hyped. For Google to invest more on making Tensorflow easy to use, there has to be early adopters showing enough interest. This will make Google invest enough resources to make it more mainstream.

Honestly, it doesn't seem that complicated. Just watch this educational video series https://www.youtube.com/watch?v=cKxRvEZd3Mw by Google, and you'll be able to do cool stuff with machine learning.

Google dispatched their solutions architect to work with them.

That's interesting. Is Google going into the consulting business? They might. That's much of what IBM and HP do now. If you're selling a new, complex B2B tool, you almost have to provide consulting services.

This seems to be the arc just about every large tech company goes in.

> While it's a cool concept and agriculture is ripe for technology disruption

Is it? From what I've heard farming is pretty high tech already.