How to Build Ar­ti­fi­cial In­tel­li­gence That’s Smarter Than a Farmer

The Star (St. Lucia) - - BUSINESS - By Emma Cos­grove

The first ar­ti­fi­cial in­tel­li­gence (AI) en­abled aug­mented re­al­ity crop man­age­ment sys­tem may be com­ing to an in­door farm near you very soon. Hux­ley com­bines ma­chine learn­ing, com­puter vi­sion and an aug­mented re­al­ity in­ter­face to es­sen­tially al­low any­one to be a mas­ter farmer.

With the help of a wear­able tech­nol­ogy like Google Glass, the user is pre­sented with in­for­ma­tion about the plants in any in­door farm. Hux­ley's AI aims to de­tect and di­ag­nose vis­ual anom­alies and then sug­gest an ac­tion to mit­i­gate the is­sue while cor­re­lat­ing it with en­vi­ron­men­tal data to de­ter­mine the cause. And since AI keeps get­ting smarter with ev­ery har­vest, founder Ryan Hooks says it won't be long un­til the world's grat­est ex­pert on hy­dro­pon­ics isn't a hu­man.

Hooks has had a var­ied ca­reer work­ing in the me­dia space for large tech com­pa­nies like Google and Vevo, as well as work­ing on food is­sues with Food Inc and the G8 Sum­mit. In 2014, he founded Is­abel, a smart grow sys­tem for the growth and trans­porta­tion of pro­duce in­doors, and de­buted Hux­ley's Plant Vi­sion plat­form in 2016.

We got into the weeds with Hooks about how Hux­ley will work, how much it will cost, and how quickly it could get smarter than to­day's mas­ter farm­ers.

What is the sta­tus of the prod­uct right now?

I've been de­vel­op­ing this for the last two years. We have 22 pi­lots ready from cannabis to big green­houses and ver­ti­cal farms. Peo­ple al­ready want the sys­tem, we just have to con­nect the cap­i­tal to get our team in place, so we can con­quer these pi­lots.

A lot of these data-based solutions need a cer­tain amount of scale to be ef­fec­tive. What kind of scale does your prod­uct re­quire?

We in­stall in­frared and RGB cam­eras in a fa­cil­ity. That can be mon­i­tor­ing 1,000 square feet of veg­eta­bles or a cou­ple of cannabis plants, or it could mon­i­tor a $15,000 or­chid. If you want, think of it as ar­ti­fi­cial in­tel­li­gence (AI) plant in­sur­ance. On the plant level, the eco­nomics of it aside, if you're grow­ing a plant, let's say this or­chid, and you have this AI plant in­sur­ance from Hux­ley and we're tak­ing a photo ev­ery minute and scan­ning it for dis­eases or any anom­alies and then we are pulling in en­vi­ron­men­tal data from the sen­sors. It can be lit­er­ally one plant. So if I have a cou­ple of plants or a hun­dred or a thou­sand or ten thou­sand, ev­ery time that crop grows and you find the best flavour and the best yield, you can vis­ually and en­vi­ron­men­tally know what the con­di­tions were that made that.

What will the pric­ing struc­ture be?

De­pend­ing on which crop and the spec­i­fi­ca­tion of the green­house, we're go­ing to charge per square foot for the AI and then for the aug­mented re­al­ity there will be a monthly ser­vice fee that will tie in with our main sys­tem. Just the AI it­self is go­ing to self op­ti­mize over time so the more green­houses and types of crops we're grow­ing, the more data we'll be able to share.

Where are you sourc­ing the ini­tial recipes that will give the AI a base of plant knowl­edge to build upon?

We're uti­liz­ing data sets from aca­demic in­sti­tu­tions that train our AI to know what to look for. One of the in­ter­est­ing things with Hux­ley is that we've been de­vel­op­ing a back-end for what's called ‘su­per­vised learn­ing'. So mas­ter grow­ers and aca­demics in the world can train the AI in what the op­ti­mal sce­nario looks like; what dis­eases look like and anom­alies too, so as we train the sys­tem, it will get smarter.

This is all done through vis­ual learn­ing, cor­rect?

What we're do­ing is vis­ual, but our data­base is track­ing from the seed to ship­ping and shows the air, light, and wa­ter con­di­tions, and then we cor­re­late that with our vi­sion.

How would it per­ceive some­thing like taste or flavour anom­alies?

If the crop is not in ideal con­di­tion, it will know that. If the let­tuce is be­com­ing more yel­low or light green, or off the op­ti­mal vis­ual path, even­tu­ally you'll just know how to self­cor­rect that. So if it's a nu­tri­ent deficiency or an en­vi­ron­men­tal sce­nario, it will know. You tell Hux­ley that this was bad let­tuce and then Hux­ley knows that all those im­ages in that whole data set be­comes a good ref­er­ence for what not to grow.

Doesn’t that leave the door open for a false cor­re­la­tion?

As it is su­per­vised by aca­demics and mas­ter grow­ers, these cor­re­la­tions will start to make more sense. The au­topi­lot of the green­house is con­trol­ling the air, light and wa­ter con­di­tions. If the nu­tri­ent pH is fine and the en­vi­ron­ment is at op­ti­mal set­tings, you're not go­ing to run into that. But if it did hap­pen, we would be able to cor­re­late why it hap­pened and it would get bet­ter over time.

Does Hux­ley have the po­ten­tial to get smarter than the smartest farmer out there?

With any com­puter vi­sion sys­tem or ma­chine learn­ing sys­tem, the data that goes into it is very im­por­tant. It needs to be su­per­vised. It needs to be catalogued in an ap­pro­pri­ate way. As you train it or [for ex­am­ple] as more self driv­ing cars go down the road, they're cre­at­ing a bet­ter map for the world around them. What Hux­ley is do­ing is that the in­for­ma­tion, as the grower is grow­ing, is tracked through ma­chine learn­ing and the com­puter vi­sion. Then even­tu­ally the con­fi­dence level of that goes up and it be­comes bet­ter than the best grower. The end goal is to cre­ate an AI that can take the stress off of all these vari­ables and sim­plify the process for the farmer.

Can you es­ti­mate a time­line for that?

I would say one-to-two years per plant species. It will never be 100%, but for scout­ing in a green­house, the best hu­man might be at an 80% con­fi­dence level, be­cause even the best hu­man go­ing through the fa­cil­ity is go­ing to miss what might be go­ing on and can't do it 24 hours a day. Once you pass that 80% thresh­old, then it's do­ing a bet­ter job than the best per­son you could hire.

It sounds like fo­cus­ing on one plant species would be the fastest way to see what it can re­ally do. Is there a rea­son your pi­lot pro­gramme is so var­ied?

The sys­tem that we're mak­ing is called plant vi­sion so what­ever grow­ers are work­ing with, they'll be able to start train­ing their own data sets. We're start­ing with cannabis be­cause it's the most eco­nom­i­cally vi­able, but if it's an or­chid or a high-mar­gin, high­value item, those are go­ing to be the best plants to start with. And also for R&D, the cy­cles are go­ing to be more im­por­tant. You're go­ing to learn a lot quicker on 30 day cy­cles than you are on 90.

Is there re­ally a $15,000 or­chid out there some­where?

Yes, if you look up high­priced flow­ers there's an or­chid that sold for £160,000 ($207k). So we could do AI or­chid in­sur­ance. Saf­fron is $900 per pound! The sec­ond a bug comes in and tries to dam­age your saf­fron crop, we'll zap it with a laser.

MedMen cannabis cul­ti­va­tion fa­cil­ity fea­tur­ing LED lights from Flu­ence Bio­engi­neer­ing with AR. Im­age: Hux­ley

Newspapers in English

Newspapers from Saint Lucia

© PressReader. All rights reserved.