Slashdot Mirror
A Christmas Menu Dreamed Up by a Robot (bbc.com)
For most of us, using up the Christmas leftovers means endless rounds of turkey sandwiches and lashings of Brussel sprout curry in the days leading up to New Year. So, to help inject some creativity into this year's leftover eat-up, BBC turned to artificial intelligence for some culinary assistance. From a report: A number of research teams around the world have been developing AI systems that are capable of learning from existing recipes and then coming up with some of their own. We asked researchers behind two innovative algorithms to see what their AI's take would be on Christmas food. One, developed by computer scientists at Stanford University, can turn whatever food is left in your fridge into a unique recipe based on those ingredients. The other, created by AI researchers at the University of Illinois, puts a cultural twist on a meal by creating dishes from one country in the style of another cuisine.
The first algorithm, called Forage, uses a type of AI known as deep neural networks, which attempts to replicate the way the human brain works. Networks like these are able to handle problems involving complex data and are increasingly being used to tackle tasks as diverse as controlling self-driving cars and recognising the early signs of cancer in health scans. [...] The second algorithm we used was developed by Lav Varshney and his team at the University of Illinois. It was trained on nearly 40,000 recipes from 20 different countries using a system that can apply semantic reasoning to replace certain ingredients with those it considers to be equivalent from a different cuisine.
The first algorithm, called Forage, uses a type of AI known as deep neural networks, which attempts to replicate the way the human brain works. Networks like these are able to handle problems involving complex data and are increasingly being used to tackle tasks as diverse as controlling self-driving cars and recognising the early signs of cancer in health scans. [...] The second algorithm we used was developed by Lav Varshney and his team at the University of Illinois. It was trained on nearly 40,000 recipes from 20 different countries using a system that can apply semantic reasoning to replace certain ingredients with those it considers to be equivalent from a different cuisine.
https://xkcd.com/720/
How can I believe you when you tell me what I don't want to hear?
If AI systems really worked, why would you apply them to creating recipes, playing Go, and playing Chess? I mean, these morons inputted 40,000 recipes in some form and trained a NN against it. Why? Is there no practical use?
Because go and chess have measurable complexity, so they can better gauge their progress.
Also, there are no externalities, so they can better measure their progress.
You want to put it to work doing something practical, you have to get some PHB to agree, and then that PHB will add externalities, and change both the instructions and the externalities as the results come in. That's all well and good for whatever they use they want to put it to, but it doesn't work for building the theoretical framework. And you want the programming libraries to be based on a theoretical framework.
Obviously, 1 + 1 = the programming libraries are still too difficult to work with for PHB-led projects. This is why it is used for practical stuff, but at the level of trade secrets; only the people with high quality teams and certain classes of problems are going to be doing this in practice already, and it still so hard to put together (from management perspective, using a big team that includes turnover) that it is just siloed. Once there is enough competition in those uses, then they'll start extracting their toolsets into libraries and leveraging that to become big players. Then it will become accessible to the masses.
Just correlating stuff without understanding does not work and can only succeed by chance. Understanding, however, remains firmly in the hands of humans, machines have not even demonstrated they may potentially one day far in the future have any say in that.
Well, in this case we're not really giving the robot a chance because it's denied access to the underlying data, all it gets is ingredients and recipes. All you can get is a "turn this summer photo into a winter photo" without any idea of the physical process behind it. If you gave it access to the chemical composition of the ingredients and the transformations caused by cooking, roasting etc. and it was eaten by sensors that could detect flavors like sweet, sour, salt, bitter and umami, smell, temperature and texture maybe it could design food that pleased our actual palate. As it is it's like sitting someone who's never had lobster down with a cookbook and ask them to figure out what lobster goes well with. Of course the answer will be nothing but a bad guess even with intelligence.
Conversely, if we managed to make AI work in a realistic simulation and avoid the problem of infinite degrees of freedom I think it could be a lot less guesswork than we believe. A bit like our vision boils down to rods and three types of cones (four for some) taste might eventually boil down to something pretty simple on the tongue too. That the creativity is more in the number of different ways we can reach roughly the same destination. In the beginning it wouldn't even have to be a dish, it could just learn what eating apples "taste" like compared to oranges. And then if apples and oranges go together...
Live today, because you never know what tomorrow brings