Classroom recording of a lecture given at Tilburg University in 2023.
So yes today we’re going to take something a slightly different topic uh we’re going to talk about electricity generation and we’re also going to talk about cost benefit analysis and critiques of uh certain interventions those two topics are kind of separate but we we’re also going to put them
Together um and then uh to the extent there’s time left we’re going to do a little bit about the carbon footprint of AI itself um a lot of what uh I want to talk about today is kind of inspired by this uh paper which is called tackling climate
Change with machine learning um and what you see there is a table from it so it’s uh it was just put onto arive in uh 2019 and it’s a big paper it’s like 111 Pages uh it’s very big and it goes through it’s really a textbook B it’s a
Textbook it’s a PDF right and um it is really kind of an encyclopedia at least as of 2019 of of different lots of different ways you could uh use machine learning to help with climate change so in this in this particular table this is just the overview and you
Can click on each of these sections to find out about it right we’ve got different application areas so electricity systems Transportation industry uh all the way down to societal impacts geoengine excuse me geoengineering and these are all completely sort of separate application domains they don’t necessarily have
Anything to do with each other except that they’re uh working towards uh mitigating climate change and then on the top the The Columns here you can see we’ve got causal inference computer vision interpretable models NLP time series transfer learning so the different types of machine learning that
You can apply here and the dots tell you which bid is relevant to which other bit right so a lot of the methods that we we’re looking at here a lot of different methods within machine learning of course and uh we can apply them to different things so I think this is a
Great starting point like up until came out uh climate change and especially machine learning with climate change was such a new topic that there was not much of a guide um this is a good starting point and uh it’s connected to the climate change. a uh Community let’s say it’s it’s kind
Of a website and they organize uh conference workshops so if you’re looking for what to do uh in this domain then climate change AI is is a good place to start okay so some of the inspiration for what we’re going to talk about today comes from there and um as I
Said in lecture one AI in most cases AI isn’t actually the solution that we’re looking for because we’re looking for reductions in emissions we’re looking for a lot of physical things but AI helps to implement to control to manage the solutions um today we’re going to look at electricity systems in particular
Again this diagram comes from the uh the 2019 paper um each of the uh labels you see here is basically a different subdomain within uh energy and electricity systems there’s a lot you can do and they’re all again very different so we could uh it could be forecasting demand so forecasting demand obviously
Has a lot of social context like who who’s going to be um driving this weekend who’s going to be watching TV this weekend you know it’s related to consumer behavior um and it could be as far as accelerating Fusion science so that’s a completely different domain each one of these again very different
Domains in all cases we can use AI to to accelerate some of the uh better ways of doing things right so for the for the energy system which is a complicated uh system there are lots of ways that we can uh make things better um we will talk quite a bit about
Renewable sources I think I showed you this before okay so there’s been a massive growth in renewable energy and it’s the growth keeps growing because it’s just uh yeah 2022 was another record installation especially for the Netherlands um and I saw this uh plot
Last week and this is kind of um yeah I think it’s kind of astonishing this plot shows you how many hours in the year the Netherlands produced more uh renewable electricity than it used so when you produce more energy than you use you sell it to Belgium you
Sell it to Germany um but you’re basically self-sufficient in electricity for that hour and if we look at 2021 that’s a very flat line because the Netherlands wasn’t producing more than it needed in 2022 it started to boost and then and then something you know the massive
Installation in 2022 has led to this massive generation in 2023 so I mean things are really changing even just in the past two years the level of renewable energy production has really it’s it’s amazing uh I find it amazing and this is um so this is good news um but it also
Shows the importance of managing uh renewable energy uh resources and as you can tell a little bit from that graph and you know perfectly well renewable sources are variable and they distributed now I’ll show you this video is that right yeah this is not very let’s try the
Lighting now if I change the lighting a bit okay so here’s um the uh the cloud and the solar energy generation over the UK and what you can see here is it’s a composite we see every time this the screen goes dark that’s nighttime and uh and then when it
Come it comes through the day you can see the clouds um the actual uh time is up there and so each of these dots is a solar domestic uh solar energy system that’s been logging the amount of energy it produces and you can see how much it
Varies from day to day this is a very cloudy day no one’s producing anything right uh so everything stays Blue and then the next day things are a bit different different the way that these light up so each individual point is someone’s home and and whether or not they’re going to
Be generating energy today it varies massively and especially according to the cloud cover uh which in places like the UK and the Netherlands is very varied yeah that’s a sunny day okay so um the other thing is that the grid was not built for this so the grid was
Built back in the day when you just had a handful of power plants and the power just went outwards to everyone um variable and distributed was not part of the plan so actually that means that a lot of the system doesn’t the the the the the things that have previously been installed they
Don’t automatically sense how much electricity is coming back into the grid from these things so it may seem strange but there’s a kind of a lack of monitoring uh in even the UK and the Netherlands uh rich countries but the the uh we don’t have you know
These these flows of data telling is exactly what uh Power is coming from which um individual home so one of the things we want to do is to forecast what’s going to happen now why is that firstly the amount of electricity that you produce has to match the amount of
Electricity that is consumed because otherwise the the excess energy in the grid it leads to bad things so if you are using more then you need to produce more and that really has to be kept in sync so the people who work at the electricity grid have to make sure this
Is happening sort of live they have to make sure they’re producing enough if the production is variable because of renewable energy being variable then uh one of the things that we have to do is to sort of keep some backup power uh so that’s what I’ve drawn on the right here
Spinning Reserve so there is some some sort of fossil fuel being burnt just in the background just to make sure that there is a generator ready to go if the renewable energy dips and so of course this uh this backup Reserve that’s still generating uh carbon emissions because it’s fossil
Powered and if we had better predictions of what’s going to happen in the next hour in the next six hours we would know whether or not we’re going to need the fossil Reserve so what we want to do is predict basically what I showed you in the video
We want to predict whether the where and when the uh renewable energy is coming from and uh in this plot we can see what that kind of looks like so it’s a Time series prediction problem and uh this we show three days here um you can see a
Prediction in red and you can see the actual uh outcome in blue so of course we know that on a daily cycle there is a general pattern but but what the details are of that it’s actually quite difficult to forecast because you can see a cloud goes over something a cloud
Goes away from something there’s a lot of very rapid variation the better we can get these predictions the better we can manage our overall electricity system so how can we do it okay if we want to make short-term predictions of solar energy well if we think about the video
We’re going to need to know the cloud cover and that’s forecast right so we saw a video of what the cloud actually was but we’re going to need to forecast the cloud cover and then we put that together with simply the geolocations of our solar energy generators and that helps us to
Give our predictions right so far so good now that’s the plan and that would be great if we had these two things but neither of those two things are trivial um I spoke before about uh crowd sourcing and geolocations right so we have uh here’s a large solar farm and
You can see in this overhead aerial imagery it looks nice and obvious you may want to do machine learning to detect this thing using remote sensing on the other hand these are quite big and obvious and I I think uh especially with crowd sourcing or possibly with government data you can actually find
Out directly where these things are the challenge comes when you have these tiny things here I use this photo to illustrate this is a street in London and it illustrates the if you were going to do M computer vision for this it’s actually a really difficult challenge because even just these three houses
Next to each other they are all solar panels on top um but they look completely different this is uh solar panels on either side side of a rided roof this is solar panels standing up on a flat roof and this is solar panels lying down on a flat
Roof and then this is Complicated by the fact that okay here are some cars here’s windows it’s really difficult actually it’s it might not sound like it’s uh the flashiest challenge for a computer vision expert but it’s a difficult one so actually detecting these things is
Kind of difficult is it important uh in the UK at least and I’m sure it’s the same in the Netherlands it provides one3 of the installed capacity it’s probably more in the Netherlands actually because we have a lot of uh domestic rooftop solar so it even though these things are small
There’s a lot of them it’s a it’s actually a big part of our energy generation so that’s um why uh we want to collect the uh the the geolocations of these things so that we can help to predict I spoke about this before and I said that we um through a project in
2020 we created a data set of the UK data other people did it for India for France um and these are not exactly the same data sets but you know maybe we can gather them all together and do useful things with them um this field is moving
Quickly because uh ever since I made that slide at the start of the course uh this other one came out for for Germany so now there’s a a data set of German uh solar panels together with um high resolution uh satellite Imager so maybe that German data set could be
Very good for computer vision and a big part of the software engineering challenge here is how we’re going to pre-process those data sets so those four data sets from the previous page they all come in slightly different formats uh you need to do your feature engineering and you need to be able to
Handle your latitudes and longitudes correctly as we’ve seen plenty of times um what I’m showing here is from our uh UK data we went through this process of we had um I know these initials don’t mean much to you but we had government data we had uh citizen science crowdsource
Data and one of the problems here is if you’ve got multiple data sets is that you could have duplication so we could have a solar farm according to the government a solar farm according to Citizen science and it’s the same thing so among other things we have to do D
Duplication lots of pre-processing so um if the total capacity is expressed as megawatts or gigawatts you need to map these things right so there’s a lot of feature engineering a lot of software engineering that goes to create the final uh usable data set excuse me
Um and what’s shown here is the sort of Comm uh the citizen science Loop in in how we did it so here’s our pre-processing and we create a data search and then we actually feed back to the community that’s collecting data and say um we’d like people to label this
Attribute of the solar panels or we’d like people to collect more data over here so there’s a whole citizen science cycle here and that refers back to what we spoke about last week but let’s assume we’ve got the data all right so now can we train a machine learning
Algorithm because we’d like to that’s what the the name title of the course is is implying all right so we’ve got some geolocations and so on um another problem is in fact the imagery data itself if we wanted to do computer vision to recognize those solar
Panels um this is a quote from some people I collaborated with they need to load two and a half gigabytes of uh imagery off disk per second and the reason for that is that the machine learning algorithm you know machine learning we have a deep learning algorithm running on a powerful computer
With GPU or something like this so there’s a powerful algorithm it’s training it needs to suck in this many uh images per second in order to actually be you know properly utilized so again software engineering comes here comes in here again because loading an entire highdefinition movie per second into the
Algorithm it’s not trivial so I’ll leave that kind of thing for your uh software engineering uh course but it’s difficult and so here we’ve got uh uh this image shows uh standard imagery from the Netherlands so that’s public data published by this uh this uh organization so 25 cm squared is about
This uh and then the number of pixels to cover the Netherlands right so if we have it like this then um a standard solar pack panel is going to be like a 16 pixel Square something like that and to cover the entire country is going to be a lot of
Pixels it’s going to be 600 billion pixels and we’re going to want to process them all right so if we were going to analyze the Netherlands Okay so we’ve already talked about how we might do it I’m not going to go over that again so some of the
This is a classic example of what we might do in in remote sensing we’re we’re scanning imagery and we’re trying to detect these objects right fine um I’ve put on the screen here just a couple of examples of people of research papers where people have been doing that
And feel free to look at those so this is we don’t know where the solar panels are we train some computer vision algorithm to find them but that’s that’s only a step on the way to where we to go what we actually want to do if you remember is
Is to make predictions about the solar generation so uh here’s here’s the first step where machine learning might help to detect solar panels and the second part we actually want to do time series prediction so completely different so instead of computer vision here we’ve got time series forecasting um these are
Well it’s it’s it’s a different situation but I’m sure you’ve come across it before we we are going to train on you could imagine for example that I train on the first two days and then try to predict the next day this is the kind of setup that we’re going to do
In Time series forecasting yeah and I will show you an example so if you remember the UK uh video that I showed you before oops wrong button video I mean there we go all right so now what do we see on the left is kind of similar to what we had
Before uh now nwp here means numerical weather predictions so this is basically the data that goes into a weather forecast it’s trying to predict how much Sunshine there will be there is some sort of model making these predictions so now weather forecasting of course has been worked on for a long
Time and it needs to be very Det detailed to put into these kind of models but basically we’re going to say uh whatever runs the weather forecast system we’re going to take those and those are going to be input features to our model so if I want to predict the
Solar uh energy in a particular location I’m going to take the uh the irradiance so the sunshine prediction for that location and that’s just going to be a feature that goes into my model right so that’s going to be a predictor now this middle column shows you actually what did happen the country
Is broken into pieces here according to the sort of energy generation regions so all National Grids are kind of broken down into sort of cellular regions um and so that’s why it looks like it does you can see the actual generation and it’s very patchy and very
Variable and then next to it you can see the prediction so we’re taking uh we’re taking this data and we’re taking the G locations and we’re making the prediction here of course it’s very easy to predict the general it goes up it goes down but then the variation you can
See it’s okay it’s pretty good but it still needs improving now the thing that I’m showing you here is the uh prediction that is run at it is the prediction that is used in the UK National Grid so what they’re going to do is they
Are going to be in the control room for the whole country’s National Grid uh looking at data like this to decide whether they need to uh activate more gas generation or something like this um this is genuinely what is happening in the control rooms and so what we want to do is
Improve this so that it gets closer to that yeah any questions about this is there anything that’s unclear so that’s about the uh the the the generation and prediction sorry the yeah it’s more like the prediction of uh solar power in order to manage it I’ll just quickly show you a couple
Of other examples that where things come into sort of the electricity system like completely different applications of machine learning um this is one uh about wind turbines now wind turbines you may have heard there can be collisions uh uh with bats and birds and uh some nature organizations have uh raised questions about
This um we want to reduce it you can see uh this is pretty rare footage uh we’ve got a sort of uh infrared uh footage here where there is a bat flying past and yes it gets past the wind turbine but yeah you can see the um the turbines
Are moving fast so it’s not always obvious for a bird or a bat if and how to get past it they’re not familiar with it we want to avoid that there’s going to be any um uh fallout from that and one thing that is done is to
Stop the wind turbines at the points when say migrating birds or foraging bats are going to be flying through we heard of this this is U it’s it happens in in pretty much all the uh the wind generation when there are important moments for like Wildlife are going to
Be passing through they simply stop the wind turbines now that’s a big deal right because we are now not generating anything from the wind turbine so it’s it’s it’s an expensive thing to be doing um and let’s see this uh we have some labels in here but if we uh if we don’t
Pause the turbines at all then okay we’re going to generate as much energy as we like or at least the maximum that we can and then okay whatever the level of um mortality for uh bats passing through it’s okay it’s full blast what is very common is to have a fixed
Schedule so a particular time of night we stop the turbines we get a little bit less uh electricity generation but the mortality decreases if we can do it automatically then we can get even more mortality decreasing we can really avoid the the the bats and birds
Collisions and that if if we do it well then it doesn’t lead to any less energy so of course this is better than the fixed schedule kind of an interesting question I think is how are you going to sense these animals right so let’s imagine it’s bats or Birds
You’re going to put a device on a wind turbine that’s going to detect these animals uh will you use image radar sonar or audio just think about how you might do it I I’m going to ask for a show of hands just to see which methods are
Preferred so this is for daytime and nighttime on top of a wind turbine birds and bats pretty far away but we want to detect them who would do it with images radar sonar audio okay and that’s that’s quite an even spread I think radar was quite an interesting option uh yes
Um there are different methods that are tried and I think uh the the ideal would be a mixture of them image I think is actually quite difficult because by the time that you want to spot the animal I mean it’s a long way away and it’s just
Going to be a tiny pixel on your image um one of some of the advanced systems that that do exist they use a kind of radar to detect objects nearby um and then they actually once they’ve detected something then they zoom in and get a kind of telephoto image to to decide
Whether it is in fact a bat or a bird or whatever it is um so it’s it’s not a trivial problem at all but like up on the top of wind turbines they’re actually applying all these different automatic detection methods and uh trying to do it all right so that’s
One another completely different example um smart grids and micro grids right we only talked about the generation so far but the the actual consumption and use um is a place where we can put some some more intelligence uh now this image which comes from this paper shows you a little
Bit of a diagram okay now we’re on the sort of consumption we’re at sort of domestic user side of things the national grd is out there providing some amount of electricity at some kind of cost and from the point of view of a very modern house we’ve got our own
Maybe we’ve got our own renewable energy generation we’ve got the use that we made make so let’s say it it could be uh well things like TV that are not particularly related to any of the other factors um you can also choose when to activate your heating or when to
Activate your high power units um and then of course you’ve got storage so if you have solar and storage then there’s quite a lot of decisions to make there and because this is set up pretty much as a market system so there is a price that you have to pay
To uh to take some kilowatts of energy and there’s a different price that you can sell back to the grid if you got access so now there’s a kind of um Market optimization problem here how are you going to satisfy your needs at the lowest cost or the highest
Profit you can decide when to charge and discharge the battery you can decide when to buy and sell electricity you can decide whether to consume El electricity or whether to wait with the washing machine so all of these things I mean this sort of domestic electricity situation even just at the domestic
Level has become more and more complicated and making the right decisions in these situations is not trivial uh so yeah we can apply machine learning um in fact the method here is quite nice but it uses reinforcement learning I won’t talk about that right
Now if we have extra time at the end of the course we might come back to the actual algorithm use used but you can see here even for domestic consumer in this country there is quite a good argument for for for having some smart uh uh control of these things the same
Happens if we’re off- grid so if we’re in some uh well not in the Netherlands but somewhere else where we’re off grid you can take away that uh external grid and still we’ve got we’ve got an optimization problem here how can we make sure that we’re not going to run
Out of energy um uh you know in the optimal way um I don’t know I don’t know if we’ve come across reinforcement learning who’s who’s aware of reinforcement learning it’s a bit oh wow that’s a load okay great that’s that’s interesting okay have you have you used reinforcement
Learning no okay fine so but but it’s come up so anyway this is a nice application of it and I think it’s it’s kind of nice to look into it I guess I guess you might have reinforcement learning now that it’s used for um Alpha
Go and things like that I guess so yeah kind of everywhere now right so those are a few different ways in which we can use uh machine learning to make a difference in the electricity grid microphone on yes um so maybe they’re not going to tell you but uh you could
Measure how many computers you sell to Google and that might tell you implicitly how much uh carbonis embodied there so two very different ways from the inside and from the outside and these things can be very hard to get exact numbers for we need to estimate
Them as you saw on the previous Slide the magnitude the orders of magnitude of these things can vary massively so the most important thing is to is to get a good estimate of the order of magnitude of these things even if you don’t get an exact number those kind of estimates can
Help so um let’s try and itemize the carbon emissions used in developing a model an AI model so I mean uh in training it and making sure it works versus deploying it so that’s running it and monitoring it so uh I’m going to put this let’s say put this on here Um yeah so this is for develop in and this is for deploying okay so in terms of carbon footprint uh let’s let’s make a a quick list what kinds of uh what kinds of uh what should we put on the list for training an algorithm carbon footprint y okay so um y anything
Else yeah so I’m going to write that down as data center yes yep and the data center cost you have that for development and you have that for deployment as well yeah anything else for development or for deployment well I’ve already mentioned the embodied carbon so embodied essentially means or embedded yeah embedded
Um so the embedded carbon of the um uh of the computer equipment so all of the carbon that that is used in manufacturing it and that goes for both as well so there’s a few different things going on here some of some of the things are are incurred at development stage some
Of them are incurred at deployment stage some of them happen once some of them happen lots of times so um the the the training a model uh maybe it will happen once optimistically what else have we got here okay so yeah two other things which
I didn’t have on this list uh one is cooling the servers that actually takes uh a significant amount of energy and the other one is the cost of transmitting information so trans you know on the internet uh you know information seems to travel freely um but it’s not actually free because of
Course we have to pay for our broadband and so on and everyone does and there is a carbon footprint and so if you take the carbon footprint of the whole uh let’s say um internet backbone and you can spread that carbon cost over all the information that we’re traveling there
Is for every email for every piece of data you know a small carbon footprint and that needs to be added on so all of these things go together now important question is how often does each of these run so i’ I’ve separated into development and deployment now I’m a researcher so I do
A lot more development than deployment right so so my carbon footprint is is is doing this a lot and the deployment doesn’t happen very often if I was working in Industry I might do the development not very often but the deployment if my algorithm has a million customers using it every
Day that’s a heck of a lot of uses so the the the balance of these things really depends on the use case and to make it more precise right so here we have the power usage of training a machine learning algorithm right so of course it depends on how long the
Training takes now you’ve seen it yourself sometimes it takes 10 seconds sometimes it takes two days we measure it in hours times the number of processes times the power consumed per processor now when your processor is maxed out we can kind of assume that it’s it’s using it’s efficient
Maximum capacity so you can simply uh you can find these numbers in the sort of statistics for your uh your your processor so we can actually directly work this out so the the power usage which is in megawatt hours is simply the time times the power and there’s another Factor the Pue
Which is a uh a multiplier of about 1.58 this is just a nice uh 1.58 it varies but this is a good example um and this is about all of the other stuff that comes in now it becomes a bit of an estimate here but when we’re running our
Algorithm in a data center so this is if you’re running on Amazon cloud or Microsoft cloud or Google Cloud right so they are running a data center for you or if you have your own data center whatever still there is some kind of factor which is basically comes like a
Multiplayer so if I’m running my uh algorithm this much in terms of carbon footprint then there’s something like a 50% 58% extra which is the carbon footprint of all the cooling and other things needed to keep that thing going so this uh value is not trivial right an
Extra 58% it’s not trivial U it’s something that we should always remember in the process and so to turn that power usage into carbon impact well okay Power isn’t carbon impact but we can we we can say for a particular dat sorry a particular energy source it has a particular carbon
Intensity so these are things again these are statistics that you can find out if I’m going to burn coal to make electricity what is its carbon intensity how much carbon uh equivalent do I emit per megawatt hour if it’s solar how much carbon equivalent do I emit per megawatt
Hour right so these can be standard numbers but they can vary and so directly there right you can go from how many hours did my algorithm take to train to a kind of estimate or measure of what’s this carbon footprint now uh you might not have all
Of these uh data so the the power usage that’s basically what’s going on at the data center and if you can control the data center improve the data center then you can improve this part of it and uh excuse me the carbon intensity of the energy that’s of course the
Responsibility of the energy provider and you can change energy providers but it’s it’s it’s there that you would make any improvements to that part of it so here’s some examples this one um shows the carbon footprint of a kilowatt of electricity oh sorry no no no not a
Kilowatt of electricity um this is this is the example of the full system so they’re tra training Bert we’ve probably heard about ber it crops up in NLP um so this is about training The Bert model and that calculation from the previous slide how does it play out in real life
In different countries it varies if you can think about Germany Australia Central United States this slide is uh I don’t know let’s see two or three years old now but um still Germany central USA Australia what do they have in common they burn a lot of coal and that’s why
They have the higher carbon footprint for um electricity and thus for the AI and of course it varies through the year uh this is the starting from um well January through to December and if we look at Western Europe so the car the the carbon footprint of training our
Algorithm as the days get sunnier the carbon footprint goes down because we have more renewable energy so it varies according to place it varies according to time and that of course makes things harder to calculate we’d like it to be a fixed number if we want to calculate it but it’s a bit
Tricky this slide shows a similar thing coming from Google’s data so Google’s own data centers basically where it’s green is where they are um uh carbon free energy they call it excuse me and uh and so yes it varies by time of day and it varies according
To where you get the energy from so one of the things they want to do is kind of optimize when and where they use the energy the headline about uh AI taking five uh cars worth of of uh carbon came uh from one of the original attempts to
Quantify this so it came from this paper by strubel ATL and uh you can see here we got the uh travel okay we we don’t see all the details but anyway we’ve got the carbon footprint of an average car and so on and then NLP pipeline um and that’s
Where they got this uh this estimate this first estimate that it cost about as much as five cars in their lifetime um so how does this happen well firstly okay we’ve got the car here the including the fuel and the embedded emissions 126,000 pound of CO2 uh the NLP running
Training one model takes 39 so actually what they’ve done in this paper they’re taking something like Bert where you’re not just training one model but you’re running a whole process to train the model lots of different times different hyper parameters each time so we’re actually training the model
Hundreds and hundreds and hundreds of times right so they take this number 39 here multiply it by quite a lot be based on um knowing what the optimization algorithm was and that’s where it comes from um the number has been criticized for various reasons it’s fine to say that okay this optimization pipeline
Including the hyper parameterization optimization can take a lot of uh energy that’s absolutely fine and then you see how the headline ends up to say training a single AI model it’s not quite what the headline says but it’s yeah so we now we understand what’s going on um but
This was a first estimate and it’s not a great estimate I must say but but things have moved on since then there’s there’s better data now um there’s this paper from Google themselves uh if you take that uh as a good sign or a bad sign but the
They in 2022 wrote a kind of update based on their own data um it include improvements so the model um they changed their actual algorithm so that it was an algorithm that ran in 4.2 uh times less time they improve the machine they have their own
Tpus instead of gpus and that makes uh they they’ve got more power efficient um uh tensor processors here 13 times lower the data center themselves so Google um manages and runs its own data centers they were able to make more efficient data centers and so that means that this
Usage uh Factor goes down they reduced it to that much and they’ve also invested in some uh renewable electricity sources and also this kind of scheduling to say let’s run the algorithms when it’s better so their uh their carbon footprint per kilowatt has gone down the really important thing about
All of these calculations if you think back to the first calculation that I showed you it’s multiplying right it’s multiplying all of these things together so if you can make even a little Improvement on all of these things and here they’re they’re making pretty good improvements on all these things you end
Up with a massive difference relatively to their earlier work they’ve got a 700 times lower carbon footprint which is is fantastic um so that’s just one example and that’s Google talking about themselves so you know you can uh you can be skeptical about that but it does
Show how all of these considerations go together one of the things that we can do is log um the the carbon footprint of an experiment so this uh yeah okay so this experiment impact tracker I don’t think it’s actually used very commonly but it’s a nice idea which is
That you can Implement a little bit of code and it says you know every time you run your training algorithm it’s going to make a little note of when it was and how long it ran and and what the sort of cost of energy was at the
Time and then when you finished your experiment it sort of automatically outputs this statement it says we use this much electricity this much CO2 and so on and it’s so that’s specific to where you ran the code specific to when you ran the code and of course how much
Uh training you run so it’s a nice idea and this is about the whole experiment right so it’s not just about training it’s not just about inference it’s not just about evaluation it should be the whole thing and that’s what we want to do we want to evaluate the whole life
Cycle let’s see I’m not going to go through this in a lot of detail given the time available but if you wanted to train a CNN for a um a small business what could you do to reduce the footprint just take a moment to think so we’ve been through some examples
Already and the examples from that Google paper could be one thing but you know a small uh business for example they probably don’t control their own data center so they can’t do things to the data center you know there are things that are available things that are not available um if we’re thinking
At this uh small business scale there are still lots of things we can do so uh we can have uh more efficient uh cnns convolutional neural networks we can change the hardware to more efficient Hardware it’s it’s worth mentioning that of course throwing away your old hardware and buying new hardware has
This embedded carbon footprint aspect so that has to be considered but it’s pretty common that uh upgrading the hardware or you know changing your provider so that they have more uh Modern Hardware um can make a big difference to the uh yeah here we are performance per watt so so the the
Magnitude of the difference can be massive uh yeah you can use green energy of course but there’s actually quite a lot so we can use pre-trained models that’s something that we’ve spoken about and that means that you don’t need to train so much because we’ve got a pre-trained model it already exists
We’re just going to train it a bit more so actually the machine learning strategy relates to the carbon footprint um we can use smaller algorithms we can use a single algorithm rather than an assemble of lots of algorithms lots of things going on um don’t produce outputs
You won’t use so we have you know if we have these monitoring situations where we’ve got lots of images lots of audio we can just run the whole thing through the uh the algorithm but maybe it’s more efficient to say I’m not going to run the analysis until you know I’ve
Selected my data points to be analyzed lots of ways that we can do it there’s a list coming to the uh summarizing uh stage basically it’s all about the cost versus the benefit and so what’s the in in CO2 terms what’s the carbon footprint of our algorithm as we just spoke about versus
What are the benefits right so if we’re thinking about AI for biodiversity or for climate change essentially we’re aiming to make this uh this as big as possible this as small as possible and as long as the gain is positive we’re not doing a bad thing right so it’s all about accounting those
Two things and I’ve said it before AI is not the solution but it can help to monitor these Solutions and uh hopefully part of the picture and that is why we spoke today about cost benefit at the same time as speaking about the electricity sector so
Think about like right back at the start of the lecture we talked about different things we can do in the energy sector now of course it’s uh the energy use of our algorithms is going to be part of the equation so if I’m going to implement an algorithm and it’s going to
Be running on everyone’s um let’s say anyone everyone’s Smart Homes or everyone’s solar panels are not making a big difference to the the carbon footprint okay there’s a bit of an issue there so that’s why these things come together we talked about the electricity sector and the carbon footprint of AI