Hacker News Reader

Comment by sweezyjeezy | Hacker News Reader

The DeepMind datacenter project was very interesting, but a lot of the ML people I spoke to were quite dubious about how much of it was genuinely down to new AI/neural networks, and how much of it was Google PR to justify how much they spent on DeepMind.

> DeepMind trained a neural network to more accurately predict future cooling requirements, in turn reducing the power usage of the cooling system by 40 percent.

But when you look at the DeepMind blog post (https://deepmind.com/blog/deepmind-ai-reduces-google-data-ce...), it looks like the 40% model is comparing to a baseline of doing nothing. So the question is : is this really something you require an AI research powerhouse like DeepMind for, or is it something a regular data science team could do?

View on HN

I think this question applies to all applications of machine learning. PR pieces love to label every application of ML to a novel field as "artificial intelligence," even when the predictions are simple enough to be predicted by statistical regression (and in many cases that's effectively what is happening).

So yes, many applications of ML are something "a regular data science team could do." However, isn't the main benefit of ML removing the need for a team of data scientists? Of course you will need some data scientists / ML engineers to support the ML product itself, but ultimately the gains will be realized from the ML doing the work that traditionally required a team of data scientists.

That's true, but I've noticed quite a few pieces of deep learning NLP papers take pains to use competitive baselines (eg state-of-the-art non-DL techniques, more complicated than regression) and have heard that good venues will reject your paper/talk if use poor baselines

I don't think your second point follows from your first. ML is not easy to set up or maintain, even if we're just talking about linear regression. Companies trying to make 'data scientists obsolete', usually frame ML as a black box that you can plug data in one end and spit out profits the other. But in reality it doesn't work like that, significant effort has to build and maintain data pipelines, to make sure data quality is good, and to monitor that the ML model is doing anything useful in the end.

It can be easier to set up and maintain than hand-rolled solutions. Best example I can think of is speech recognition - previous systems used hidden markov models, gaussian mixtures, triphones, and all sorts of complex and obscure things that had to be tuned by experts. Now they just use an end-to-end neural network that goes directly from sound to letters (almost anyway; I believe the input is still MFCCs).

I agree it's not magic but it neither are 'manual' methods and at least the maths is a lot simpler.

> Now they just use an end-to-end neural network

You've missed my point. The thing is that building the ML model is NOT the hard part of machine learning in industry. The hard part is building an infrastructure that can make that machine learning model do something useful. It is much harder than people imagine. See for example this great paper by google for more details : https://static.googleusercontent.com/media/research.google.c...

I don't think either of you are strictly wrong, though--the comment's point seems to be that doing things by hand was so much harder that it couldn't even be done to the same level of success that machine learning achieves.

Unless you know something I don't, the SOTA in ASR is a hybrid system that still has all the complexity of the traditional generative model plus NNs on top.

I was going to say, I feel like this could be done with excel and some historical data combined with weather forecasting.

that being said, even if ML gets you a 0.5% advantage over spreadsheet math, that is a non-trivial amount of savings on a national scale.

But is it non-trivial net of the expense of gaining it?

That is a much easier decision to make, but i have no idea what the costs and savings involved are so I do not know.

On the other hand, if ML gives you a 0.5% disadvantage, that too is a non-trival amount of losses on a national scale. Because they compared against baseline rather than against another data science team (or both), it's difficult to know if ML actually brings an advantage here.

Can't we assume they already have a data science team? Optimizing the grid is their job already right?

From what I've seen in several places though is companies tend not to listen to their own people, but are willing to make changes based on the advice of hired "experts". So the experts will come in and ask everyone what the challenges are and what they might do to improve. Then the "experts" will pick several low hanging fruit and find a few things that cross internal boundaries (the harder things to change) and make recommendations. Upper management will go with it and improvements will be made. I suppose the way to sum it up is that they bring objectivity to the table.

Yes, and there are two other points to follow up this one:

If one ML solution is better than the other one by only 0.01%, in this context it is worth a lot. It is a winner-take-all type of situation.

Edit: the second point was this: If one ML algorithm gives us a 1% saving and the other one gives us a 0.9% saving, you could say that is a 0.1% improvement or you could say that is a 10% (of 1% saving) solution. A percentage sign without context is not that meaningful.

What I'm wondering is why I can't find anything about ML applied to finding and curating data, which is the most tedious part of data science. That would be an interesting way of using ML without fuzzy stuff.

There are a few current projects, particularly in database research. I don't know how many of them use ML in the traditional understanding. Current projects I know of are Wrangler, Mimir, Katara, MayBMS (in no particular order).

Did you try any of them. Wrangler seems like a tool similar to OpenRefine to me.

Would you pay for it?

>> Google PR to justify how much they spent on DeepMind - why would google try justifying for Deep mind and not to other tens-to-hundreds of companies it acquires in a year ?

DeepMind's expenditure is probably upwards of $1bn/year.

What costs do they have that would be one billion a year? I can understand servers and salary for the researchers but not to the scale of 9 figures a year.

I think you could make a similar argument for a lot of important choices: when you pick a doctor, an employee, a job, a computer, you're going to pick the best you can afford even if the middle of the road doctor or computer could get the job done. Generally people want the satisfaction that they chose the best in the business and so aren't leaving anything in the table.

From the blog post, it seems that the 40% reduction is compared to the current utilization:

"by applying DeepMind’s machine learning to our own Google data centres, we’ve managed to reduce the amount of energy we use for cooling by up to 40 percent. [...] Given how sophisticated Google’s data centres are already, it’s a phenomenal step forward."

That "up to" renders the rest of the phrase meaningless. Was it 40% less for one second/minute/hour/day?

I assume, because they haven't actually given an impressive number without massive wiggle room, that it's because they are clever enough to know the real number is embarrassingly low, but would welcome corrections.

I worked at one of the datacenters where they rolled this out. It definitely wasn't all that impressive in terms of real energy savings, it would temporarily make our PUE drop, but we'd have to make up for it after the recommendations were secured in order to account for the temporarily relaxed set points. It was mostly a PR thing from my point of view.

Not an expert but I expect that what AI/machine learning/NN shines is the ability to produce very high quality solutions (or approximations) for large-scale NP-hard problems. With regard to the "NP-hard" part, I don't think regular data science can be a sufficient substitute.

What do you mean by "regular data science"? Deep learning is just another tool in the belt of a data scientist. A deep learning solution doesn't just spring up on its own. At this point in time anyway, all solutions are engineered by engineers and data scientists.

DeepMind will go down as one of the best acquisitions ever

Reply on news.ycombinator.com