Neural network used as "doping"

Question

A genius programmer (someone with initials PJ) created a neural network aimed to provide the best results in running marathon for human.

Thanks to Google deep dream project, this programmer also realized what each of the cells in his neural network does and how the neural network actually works.

In other words, given a usual human we can determine what that human should eat, drink and how the subject should generally behave in order to run marathon perfectly.

My story centers around idea that this programmer devotes two to five years of his life to get himself trained by this neural network. He basically lets this network decide what and when he should eat, he feeds it with all his biological data and so on.

Can we expect he will win the marathon race?

In other words:

• We start with an average human
• This human does exactly as told by the Neural Network
• The Neural Network in question is trained perfectly to understand the human body and its goal is to run marathon as fast as possible (minimize marathon running time)
• This human obeys everything the neural network tells him to do for a maximum of 5 years in row
• At the end of this training session, this human is so good in running marathon that he can win

P.S.: This human does not use any illegal doping, but is willing and able to eat food supplements in tablets if neural network tells him so.

Answer 1

This neural network doesn't seem to have access to any more advanced biological knowledge than the rest of the world. It applies existing knowledge very well to its subject, but he starts out with very average athletic ability.

No, this will not produce a world-beating marathon runner. Current winners of marathons are selected for outstanding talent in the field, and then spend more time than PJ training. Their coaches are trying hard to create individually optimised training programs, and they are very skilled coaches.

The advantage of a slightly better optimised training program (and the advantage isn't going to be very large) will not make up for the much better talent of the conventionally selected athletes.

Answer 2

Olympic athletes in track and field have started by winning a genetic lottery. I'm not talking about the West-African fast-twitch muscle fibre theory. Nor the East African endurance runners bias. However sprinters are tall and have the best physical dimensions and ratios of muscle mass to bone density to limb length and balance (plus more fast-twitch muscle fibre).

Endurance runners have advantageous anaerobic endurance systems. They are also smaller than sprinters. Neuroscience research suggests that elite athletes also have atypical brains. So that may be a genetic lottery factor too.

So phenotype is important. Assuming your genius programmer has an appropriate phenotype to become a world class marathon runner then yes he can win. Even if he is currently out of shape.

In addition to that perfect training and perfect discipline, the greatest advantage of the neural net will be in reinforcing and growing the biological neural net of the athlete. By developing atypically superior biological and psychological responses to the biological stress of the marathon the athlete will have a better chance of winning the race.

Training the athlete to deal with Central Nervous System fatigue intrinsically (via genetics) and through physical training & therapy, diet, supplements, legal drugs and neuroscience.

If the artificial neural net can train the athlete's brain to produce less serotonin and more dopamine relative to other athletes then that is an immediate advantage. See the wikipedia article for a summary of the current neuroscience https://en.wikipedia.org/wiki/Central_nervous_system_fatigue

The reality is that White European athletes can and have been competitive in endurance races, ultra-marathons and walking events.

I missed this point

we start with average human

in which case the answer is no he will never win the race. Average humans with superior knowledge and training only win when the training or knowledge is the determining factor.

It's like elite special forces troops. All types and branches reject far more candidates than they accept. The candidates are all in superior physical shape. They look for an X Factor of mental toughness and then train the soldier.

Answer 3

This is not possible because current artificial neural network technology will not be able to solve this problem and also because you have biological limits over which you cannot optimize yourself any further.

Current artificial neural network technology will not be able to solve this problem

Simplifying, a neural network is nothing more than a huge statistical correlations calculator, regardless of how complex its design might be. During its training it ingests large amounts of data about the problem you are trying to solve, and if properly configured is able to find subtle and complex correlations in your data. However these insights will quite likely be useless outside the reality that is portrayed with your training data.

For instance, let's suppose we manage to collect a huge dataset on professional runners, with info on their diet, training plans, medical data, race performances and so on. A neural network trained with this might be useful for inferring what kind of diets or training plans lead to better performance for certain groups of people. If the network is very good at it, it might be able to generate useful advices for specific individuals. It will work like "OK, so you are PJ and this is your data, which kind of resembles the data of these other 5 runners, and they managed to improve this marks by doing THIS and THAT". It's really not so different from Amazon's "other similar clients bought also THIS".

But these advices will only be useful to improve your current marks, not to take you to a champion level. You might try to ask the network, "OK, so I'm this data point and I want you to take me to the data point where this champion is". This could be possible if you were a 100% configurable entity, such an universal mathematical function, but you are human, which takes us to the next point. But even if you were a superhuman with such capability, there might not be any single example in your data of an individual who managed to go from average human to champion runner, so the network will be at a loss.

You have biological limits over which you cannot optimize yourself any further

Following paulzag answer, you have a lot of biological conditioners: genetic biases, past and current medical conditions, psychological conditioners, ... even though your body and brain are somewhat flexible there are limits you won't be able to break through. Even if the network finds that the answer is "you'd better start running 40km every day before breakfast" it might be impossible for you to do so.

About Deep Dream

Let me also clarify what Deep Dream does. Again simplifying things, it is just a gigantic neural network, specially designed for recognizing different classes of objects in images and trained with millions of them. The Deep Dream trick is then to feed a random image to the network, and tell it to transform it gradually to things the network can identify as objects seen in the training photos. That's why images of clouds suddenly transform into all kinds of animals parts, because the network was trained to identify animals, not clouds.

There are however some techniques to get some insight into what each individual neuron in the network is doing. But the same is true for biological brains, where it's easy to isolate a neuron and see how it works, yet we still barely get the whole picture. It is the joint work of all neurons in a network what is hard to give an interpretation, because this joint activity exceeds the complexity of the sum of individual behaviors.

Answer 4

Let's see each point of your list...

---

• We start with an average human

So basically you are asking if training and knowledge can convert the average human into the best at a physical skill.

I can ignore the rest of the bullet points and tell you here that the answer is: NO.

Why?

Because for any sport there is selection bias on the athletes: only the people that has an inclination or some form of proficiency will try the sport at a proffesional level, and only those who perform well will be selected for world wide competitions.

This selection will translate to populations with biological adventages being overrepresented. For intance the average height of males in the USA is 175.7 cm (5 ft 9 in) [2014 data] but the average height of NBA players is 201.3 cm (6 ft 7 in) [2004 data].

So, it is expected that the average human will underperform compared to the best biologically fit individuals. Edit: Note that is given that such individual had decent training.

---

• The Neural Network in question is trained perfectly to understand the human body and its goal is to run marathon as fast as possible (minimize marathon running time)

I'll take the AI knows as well it is possible to know the body of the person it is training. Each body is different, perhaps the AI has considered the available information on all known athletes...

And I get you are using the work "understand" loosely here.

The AI you want probably isn't a simple neural network. You have stated that it has a goal, so, if you are using a neural network for this, the "smart" part is not only in the network itself but on the algorithm you are using to trai it.

Just saying....

---

• At the end of this training session, this human is so good in running marathon that he can win

Sure, he can win. But it will depend on who is this human running against...

Consider that there will be competitors that are have a better biologically fit body, and they have gone in exhaustive training too. Even if such training is inferior to the one provided by the AI, it doesn't need to be perfect training... just good enough.

Of course that is asumming all competitors had a similar preparation time and they are using equivlanet equipment. Edit: by "equivalent equipment" I mean, during the competition. You may disregard this - because it is standard - but clothing and footwear may mean a huge difference.

---

• This human does exactly as told by the Neural Network
• This human obeys everything the neural network tells him to do for a maximum of 5 years in row

5 years may not be enough. The training for the best performers in some sports may take 10 or even 20 years.

In other words, given a usual human we can determine what that human should eat, drink and how the subject should generally behave in order to run marathon perfectly.

The execution of the actions may be limited by the resources of this average human, such as access to equipment for intance. I wonder if this AI gives financial advice so that this human can buy the best equipment and best food.

And to think it knows what is the best equipment and food, would require perfect knowledge!

The AI needs to be trained for all possible food items, that would have to include all possible recipes, even that one your neighboor just invented yesteday and that new brevage that is coming to the market next month.

The AI needs to be trained for all possible training equipments and exercises. Again, this requires to know all that is in the market. Also consider that the best exercise may be one that requires apparatus yet to be invented.

The best we can realistically aspire to is an AI that is constantly learning.

---

Addendum:

Does this AI care for what is legal? It may tell this average human to steal the equipment, to use illegal drugs, and to undergo surgical body enhancements. And such action may not only break the law but also disqualify him from the competition.

And of course, killing all potential competitors with a better fit body would also "help".

Answer 5

It could be plausible. As many people have commented, varying the type/intensity of training and diet alone will not make up the difference. But who knows what could happen with a wholistic lifestyle change that is radically different (but perhaps superior to) that of other athletes.

Maybe the Neural Network realizes the ideal training cycle for his body is 15 hours per cycle. Not many people live on anything other than a 24 hour cycle, and while it's true that life has generally adapted towards this, it may not be the optimum for transforming into a perfect Olympian.

Or, for easier math, maybe he does 12 hour cycles of 30 min eat -> 2 hours workout -> 30 min eat -> 2 hours workout -> 30 min eat -> 2 hours workout -> 4.5 hours sleep. In 24 hours he'll work out 12 hours, eat 6 times, and sleep 9 hours.

The demands placed on the mind and body by this training regimen, long term, would be extremely high. The Neural Network would need to be able to motivate him highly, or induce a long term trancelike state where he is highly susceptible to suggestion and follows orders when any sane human would quit from exhaustion.

Also he'd likely need access to training equipment beyond just running. Run too much and your body breaks down. Combination of low-impact water-resistance activities (water cycling, water running, swimming) and direct electrical stimulation of muscles in Neural Network determined patterns.

He'd accomplish no work outside of his training, no time for socializing of any form, and the cumulative stress might push his body to the breaking point right after the marathon (collapse with a heart attack crossing the finish line). Maybe the Neural Network knew this but didn't tell him because it would impact his motivation for training.

Even then, maybe it requires special circumstances for him to win. Like moderate rain during the marathon. Most competitors start out slower, and no one chases the underdog who goes out fast. No one can draft off of him, and he unexpectedly holds on to win with a time 15 minutes slower than the world record marathon, but still faster than what the other competitors did that race. Again, Neural Network might even know that him winning is a long shot, but lie to bolster his confidence, and in your story rainy conditions just so happen to occur.

Even with such an advanced AI, I don't think he could remain at the top of the Marathoning field for say, 4 years (enough to win 2 Olympic gold medals). You generally need to win the genetic lottery in a big way to get near the top, as others have stated. You might fudge your way in with absolute perfect preparation as an average joe, for one performance under ideal circumstances, but sooner or later your "averageness" is going to catch up with you. Most human bodies cannot handle the stress of performing at an elite level over a sustained period of time.

Of course, this AI has already gone far beyond just proper diet/training, but I'd be willing to believe this one, if the author met me half way and acknowledged that in the end, even with all this (near-suicidal) prep, it was a 50/50 gamble.

Answer 6

If You Want Him To

The best chance of winning is not to have a neural network for marathon running, per se, but rather one that understands human biology at an unprecedented level. For instance, if the NN understood the proteomics of PJ's body at a fine-grained level, as well as his microbiome and the methylation patterns of his DNA, it may be able to tailor a program which alters all of these in ways which current humans cannot even conceive.

Genes Aren't All That...

Everyone is hung up on having the perfect DNA, but the biggest surprise of the Human Genome Project is how few genes we have, and how little genetic variation we have. The reality is that genes get turned on and off all the time, as part of regular processes, or in response to environmental conditions. A gene can become enabled/disabled long-term (say, over several generations) via DNA methylation. The pattern of methylation forms what we call epigenetics.

...But They Are!

Now, if humans have so few genes, where does all the complexity of human biology come from? Well, it turns out that humans are mostly alien. Only 10% of your cells have a genetic pattern which can be recognized as "you". The rest? Parasites and symbiotes (hopefully mostly the latter), primarily in your gut: your microbiome. Only 1% of the genes floating around the fluid bag which is you are "your genes".

Part of the reason we are the ultimate chemical factories, transforming everything from plants to animals into energy and building materials is that we are host to thousands and millions of highly specialized chemical processors. And what we feed those microbes changes their relative balance.

Conclusion

So, while you pretty much cannot change your human genes, you can affect the methylation patterns, which can affect all kinds of internal body systems. And you can definitely change your microbiome, to the extent of wiping it out almost entirely and starting from scratch (which is what strong antibiotics and diarrhea do, basically).

If your NN can then use these tools to remake PJ into a super-runner, I personally wouldn't question it too much. Oh, and there's the boring things like using the NN to create the perfect stride form for PJ's biomechanics. But everyone probably believes that is easy, more or less. What might be more interesting is if it also taught him how to alter the stride during the course of the race as his hydration level and other dynamic factors change his weight balance. For instance, there can likely be fine-tuning when climbing and descending hills, taking sharp turns, etc. Or, it is even possible that he can shift which muscles are used in order to rest some of them during the race by altering his stride, though I have no idea if that is a net benefit or not (think lactic acid buildup, etc.).

Answer 7

Neural Networks work just like a human brain - they get some data and learn how to output other data. We usually teach them to output something we won't - e.g. an input is a photo, an output - information whether or not the photo contains tank.

See, no magic. This won't tell you the purpose of the universe. It's just a tool that if trained properly can be really good at something(e.g., game called GO).

Creating a proper exercise plan and diet is all about biology, chemistry and keeping an eye on the progress, so that you can change the plan if the need comes(e.g. injury, better or worse results then expected, etc.). What does it mean for us?

It means that this neural network could become a cheaper alternative for personal trainer. This wouldn't revolutionize the field, certainly wouldn't instantly start producing ubersoldiers capable. Sorry, that's not how Neural Networks work.

Answer 8

To answer your question, no, we cannot expect him to win the race given substantial competition, for instance if there is a substantial cash award for winning. What might be more practical is to train a neural network to optimally train a runner using a particular training philosophy - say Lydiard or Daniels. Feedback to the neural network would be the runner's behavior - details of workouts, weight, injuries and aches and pains, diet, etc.

Answer 9

No.

Consider your godly neural network as just a black box. It consumes whatever information, and emits commands for your human.

This is no different from having the best marathon coach in the world to train you for 5 years. The coach also is a black box who consumes information and spits out commands.

If your body starts out average, it will stay average. You will, at the end of the day, be an average person trained perfectly to run marathons. Everybody that started out with a non-average body will beat you easily (assuming they also train for a few years, of course), even if their training is slightly less optimal than yours, because there are things influencing your running that cannot be changed by training (like the length of bones, the absolute upper limit for certain structures in your body to grow with training (i.e. cardiovascular capabilities etc.) and so on).