1.
Accuracy 0.90 achieved so easily, makes me wonder if I've done something wrong. Lol.

2.
My neural net models are the only things in my life doing well. I think I chose the right career. Lol.

3.
Rerunning experiments is not fun. But getting better results is really... Ego stroking.

90% accuracy is fun app level. Aim for at least three nines for anything serious (and way more if you think about using it for something really dangerous - like classifying collected mushrooms).

As long as you don't overfit, this is a nice start :)

@Oktokolo it's real life application. 90 is pretty decent considering noise and delay.

@NEMESISprj no overfit. It's eval acc. 😜

@NoMad nice! What real-life application is it used for if I may ask? Always interesting to see what people do

@NEMESISprj considering that I'm publishing a paper on it, (fingers crossed, if shit doesn't go wrong) I can't actually discuss it without giving away my name. 😜 (And this is devrant. I've ranted behind ppl's back pretty often)

All I can say, is that it's robotic. 🤐

@NoMad
Ah its for a paper. Yes, then its fine. Code produced for a paper normally gets never touched again anyways.

@NoMad ah yes, well good luck in that case! At least robotics sounds good already haha

@Oktokolo it really depends on application and risks attached... 90% could be sufficient. And we also know nothing about whether or not this is the final model. I agree that you'd want it as high as possible, but sometimes things like data gathering, preprocessing, noise and anomalies can be a real pain in the ass.

Oh and the "so easily" part in the rant probably indicates that it's in early stage of development but that might be me

@Oktokolo well, that's complex. The ten percent is on softmax output. I'm waiting for kfold to finish so I can see how the it actually looks in the end. My guess is, better than 90%.

@NEMESISprj
Of course it isn't easy.
But tell anyone outside IT that you have a ten percent error rate and he will think you are joking.
Outside of IT, ten percent error rate is something to be deeply ashamed of.

Also as a starting point it is indeed fine.
And for a proof of concept it is fine too.

@Oktokolo in medical 90% is godlike

@electrineer
Imagine every tenths pacemaker failing after a month. Or hart rate monitors omitting an average of ten in hundred spikes. Or every tenths patient experiences a failing cardiopulmonary bypass...
The manufacturers of these devices would get bankrupt pretty fast...
...And some of the higher level employees would get prison time.

@Oktokolo it's probably the exact opposite: outside of it applications for neural networks (like medicine and life sciences for exanple), 90% is incredible because you work with limited data most of the time as far as I'm aware.
But with medicine 90% could be a death blow to your model if this means that you have high false negative/positive error rates causing lives to be lost or something..

@Oktokolo those don't work with ML tho.

All in all neural network accuracies are really only to be evaluated in their application, let's keep it on that lol

@NoMad
Exactly. Regulators did not catch up to the new ML hype yet. So we can still see the absolute absence of any quality standards whatsoever in that new field.

@Oktokolo even in medicine, I'd argue if you're down to the last 10% if your model is working or not, then your design is completely wrong. Also, wouldn't those applications have a secondary/tertiary checks? It's not down to one model's accuracy, and if it is, then who's gonna trust those?

I agree regulation and testing in medical apps are absolutely necessary, and even then, long term monitor by a human expert are necessary.

@Oktokolo i wish netflix had 90% eval acc. in their recommendation engine.

@Oktokolo @NoMad

It really depends on the application though. If that 90% is significantly better than what any human can do, it's really really good, but if failure is fatal to the robot (or slightly less important: humans) then it's really really bad.

It depends on what success/failure is exactly: A self driving car which classifies traffic lights incorrectly 10% of the time is extremely dangerous, but a self driving car which mistakes a cat for a dog 10% of the time... meh, as long as it detects "some moving object" at a near-100% rate, and the safety margins for evading the pet are good enough, it doesn't have to be a problem.

Also, if you have 90% accurate weather bot -- or a trading bot which can predict all stock price movements a few days into the future with 90% accuracy, you'd quickly become the richest person in the world.

@Oktokolo ahahaha this is so true. thanks for the laugh

@bioDan
Content recommendation in general would do just fine with a plain simple and dumb non-ML approach:
Just select unconsumed items matching tags where the overall rating of consumed items is positive. More positive -> more items having that tag.
Then add N% randomly selected items for debubbling (also called capitol raid prevention).

Obviously, debubbling isn't an issue at Netflix. But users getting bored because they never get anything recommended which isn't the dsame as they already know, might be an issue there.
so content recommendation is indeed the area where you actually want to not have 100% accuracy...

@Oktokolo i get your approach but in many cases it will miss the objective. How do you handle what happens when you have many tags for an item, just one tag, or none? Will you always recommend the same results?

Recommendation should take into account many other features like age range, region, sex, and more (given they are available in your data)

Also, it doesnt work in most cases where complementary products are recommended, like in amazon when buying a product X you get "users who bough X also bought Y, Z, .."

@bioDan
The amount of tags doesn't matter, you can weight them individually and then aggregate the tag values for each item. Can even take into account whether tags are specific or generic by weighting values differently...
However complex you go there: It will not get computationally more expensive than using an ANN for it.

ML might or might not work better than pure statistics for recommending prducts.
Even a purely random selection would beat Amazon's current recommendation system though...

Also both, taking into account the buying patterns of groups of people having same attributes assigned to them, and recommending products often bought together, is solveable by pure statistics.

General rule of thumb: If you can solve it without ML, do so.

There are plenty of areas, that are currently not solveable without ML because we don't know any suitable algorithms yet. Apply ML there.

@Oktokolo ANN is in the area of DL, classic ML is based on pure statistics.

I agree with the rule of thumb.

Some solutions to challenges can be solved with a simple algorithm like you suggested or a decision tree type algorithm if your data and the constraints on your data are finite.

But in todays world data is only growing in veracity, volume, and velocity.

So when optimizing for recommendations for any system with big data, the more data, variety of data, and relationships within the data you have, the more complex the task becomes. That's why challenges in the field of big data require ML or DL solutions.

Making the most basic linear regression without embedding the knowledge in a model will not work.

@Oktokolo @NoMad imagine a treatment that has 50% success rate of saving a life, and there's no better one available.