Audioburst Search

Coming up next

Tuomo, Black Sea, India discussed on This Week in Machine Learning & AI

This Week in Machine Learning & AI
|
2 months ago

Trump says Navy ship Comfort will be used for COVID-19 patients

WTOP 24 Hour News
|
3 hrs ago

Maeve Kennedy Townsend McKean's body recovered

WTOP 24 Hour News
|
3 hrs ago

Wisconsin Supreme Court rules that governor cannot postpone Tuesday's presidential primary, despite virus outbreak fears

KCBS Radio Afternoon News
|
7 hrs ago

Thanks to Trump, There’s a Hydroxychloroquine Shortage

Markley and Van Camp
|
29 min ago

Boris Johnson spent the night in intensive care receiving oxygen treatment after his coronavirus symptoms worsened

Bloomberg Daybreak: Europe
|
28 min ago

Ex-Vatican treasurer Cardinal Pell freed after winning appeal over child abuse

All Things Considered
|
14 min ago

Washington state, California to send 400 ventilators to other states

WBZ Afternoon News
|
12 hrs ago

Governor Cuomo: "Now is not the time" to ease up on social distancing for coronavirus

Terry Meiners and Company
|
12 hrs ago

Abortion in Oklahoma can resume, despite governor's ban

Sean Hannity
|
12 hrs ago

A rush to medical face masks could hurt health workers, global officials say

Sean Hannity
|
12 hrs ago

New York City dog parks, runs closed over coronavirus

Michael Wallace and Steve Scott
|
14 hrs ago

Dallas - North Texas Food Bank Calls on National Guard to Help Distribute Food

Jim Bohannon
|
14 hrs ago

Automatic TRANSCRIPT

Producing a lot more weed to export to India or Canada and other countries will who are employed will have to import from other bases. So it it's has. It's it's sort of like this big iceberg Hiding under the water at could have a big impact so so all that to say is that the first part of modeling process is. Is it an interesting interesting thing to model from a business point of view is it a and fundamentally and so you know that. Know that and there's a lot a lot of them but it's important that we work on the right ones and secondly is then we we like I said we don't come out problem with a solution which which I think distinguishes us from many companies. Where would we didn't start out as an AI? Companies saying well you know. Let's just find the problem. Lover saying had to work on and stumbled into agriculture. was more a joint thing. Let's figuring that you know. This particular domain needs needs these answers of needs better decisions that we don't care if we end up using neural networks or gradient descent Algorithm X G Bluestar list on a random forest algorithm where where agnostic to technology but. We want to solve the problem so I think that the second in part of our approaches were every single new problem where prepares to use different approaches but third Third piece is that when we build a particular model we do try to build a framework that helps us experiment in all in res things in in different situations uh-huh so for example even though each country in crop is different for a yield modeling. We do have a basic framework that applies across the board and some of the input signals will change but there is a two or three key algorithms that we we will always use and then the specific inputs will be different but the The back testing and how we evaluate our results in on will we'll be. We'll be following that process in a now we try to look at Finally when we were getting close to to into having a something that we like and that we were ready to launch We look at performance in in very unique ways so for example we don't look at just the average value of Arab or or just the signal. But you know even when we're predicting a single number unlisted Yield of Russia are the are the Black Sea region. It might be a single number at the end but underneath it. We're actually making that same Shane prediction in a much more granular way. So we look at all. How's the era distributed spatially? Does that make sense of the era distributed in the back testing like what. What are that when we run at when me back to historic doesn't you know what are the years where our model would have performed better or worse than why isn't them? Does that just random noise or is it like does give us features that we should model so we have a very iterative process where we look at spatial temporal distribution distribution for far rents and bring in a lot of domain expertise to sort of figure out the future engineering and antiques. That we need to do you too. Have a good model. Is it hard in your case to know when to stop when it's good enough I know I think we're lucky because again from the first point I started with. We usually have a very clear idea of why this is interesting and to gather with that comes in idea of what's were what's reasonable to move the needle right so it turns to add value to the world so like let's say if you're trying to model something in the United States than there is typically going to be a lot of really good inputs a lot of historical data and better than most host other places in the world so the so we will our expectation would be like well okay to make a difference in this case we really make GonNa make sure that our error is less than zero point five percent right because anybody can get to within two percent or whatever right like in the data of good it's yeah exactly and and conversely wrestling you go to a country where part of the world where is very little data and there's no or maybe maybe no truth at all on on and then you're like wow okay everybody's the best anybody in the world can do ten percent or twenty percent are they have or maybe fifty percent because they have absolutely no idea what's going on and we were like well okay can we consistently robust league fruit make. A forecast has an era of tempers on. That's Rabi that's awesome. That would be terrible for for something that You know that everybody else has one percent error on but if it's ever as twenty percent and we'll put it. They will be confident that we're adding value so so so we always know when when we are adding valued we're not about and so that gives us an easy way to say okay. Let's launch this even though there's a million things that we can do to improve or were already ahead of where value well Nemo. Thanks so much for the so generous with your time and sharing with us What you're up to their at grow very much appreciated thanks? I yea enjoyed it. I hope it was informative and inspiring. Absolutely thanks so much. Thank you all right. Everyone that's our show for today for more information about today's guests or any of the topics mentioned in the interview visit Tuomo. Ai Dot com slash shows was to learn more about the IBM AI. Enterprise workflow study group. I'll be leading visit dot com slash. Ai workflow of of course if you like what you hear on the podcast please subscribe to rate and review the show on your favorite podcast. Her thanks so much for listening and catch you next it's time..