About the talk
Retailers might find it challenging to start with AI and machine learning – but the job gets much simpler as their efforts progress, according to Rashed Haq, global head of AI, Robotics and Data at Publicis Sapient, a digital transformation consultancy owned by French advertising giant Publicis.
At the AI Summit in San Francisco, Haq presented a session on applications of AI technologies in retail. He noted that machine learning projects are easier to scale if the same models and data are constantly re-purposed across the organization.
Keep up-to-date with all the latest AI news and insight at https://aibusiness.com
Welcome to add to the city here at the anthem in San Francisco 2019. I'll be very pleased to have with us today or should have you think we will head 00:05 of AI and data at Publix today in Sussex Welcome to our presentation yesterday. 00:14 It was around. Can you talk to a few points? Annex key 00:23 takeaways musician business everything that's engaging with the customer and 00:33 operations in the supply chain and some of the points I was making his giving examples of how businesses are using it across the board 00:43
and the advantage of using it more broadly across the business rather than within each of the department because 00:53 if you look at what's going on in Marketing and then what happens in the e-commerce Channel some of the data and many of the algorithms 01:03 can be reused SoDo integrating those on a single platform makes sense. And now we're also seeing some of 01:12 that data being used inside of the supply chain. So if you build the supply chain demand forecasting models, if you look up the 01:22
demand forecasting models within the supply chain, they are built on historical historical sales data 01:32 stream data and the search data from the e-commerce site and what the customers are doing the Acura signal improve significantly, 01:39 so being able to use both improves the quality of the results the performance of the business 01:48 and the advantages you have shared data platform in a shared AI platform, which reduce Can you feed an example the 01:58 best song, you know in real life how this is being implemented today? And he use cases are around the world 02:08
15% of retail businesses are doing this in some form or other 02:12 there had various stages of off the transformation for themselves. So if you look up for exempt some companies where they're 02:22 using data that they're collecting from The Clique stream on their eCommerce + data, they collected from adding fractions that the copy over 02:32 to the to the same connect to database and also data that they're connecting from in-store the for example, if your if you 02:42 have Wi-Fi in a customer connect with a prospect connect to the Wi-Fi, you have their device ID and you know where they're what their 02:52
behaviors are in store. So if you collect all of those together We were working with the company where we increase their 03:01 customer acquisition from digital ads from 2% conversion to over 6% conversion. And the 03:10 cost of the marketing dollars went down by 43% because they didn't have to blanket Target people that could 03:19 very specific examples in the supply chain, like the example of using the historical 03:28 research data for another company. We improve the accuracy Buy close to 03:38
82% for the demand forecast for reduce the error. So, 03:47 how is seeing a are being included? 2019 03:57 thinking about digital transformation. What we're seeing is one of the big Fellers 04:06 underneath that is most the leading companies today at transforming themselves into the data driven companies 04:16 and the main way that data is utilized. So that's a big component and they're doing doing that across not just 04:25 almost every every industry kind of skills acquisition keeping skills within the company or Outsource a 04:35
T1 versus Diaz has a big conversation around the pros and cons 04:45 between the two 04:54 How many smart people you have inside the company? They're more spark smart people outside of your company. And if you try to choose one or the other 05:05 their big downsides to to both so it's ideal to do both at the same time. So you do have to 05:13 build your team internally, but particularly the decision making components you want to keep in-house you want to bring Partners from outside 05:23
whether those are Consulting companies like ourselves for startups that are startups are other companies that have some of the products 05:32 that you can use an integrated into your API and you can also build your models on on your own but there's no reason to build 05:42 something if somebody or has already build it and you can use it as an API. 05:51 Tilting Point push the 06:05 importance of both the the transfer transparency 06:12 and reducing bias in AI systems and the interpretability. So I'll give you a couple couple of examples 06:21
we're working with one company where we were trying to automate the loan approval process using machine learning algorithm. 06:31 And if you put them older historical data and then ask it to predict them historical data that it hasn't seen we we were getting 06:40 extremely high accuracy, but then if you've tested for bias, it's also very biased and wait. We're particularly looking at gender bias. And then we 06:50 went back to the source data and we looked at just the date itself. And we said if you're if we remove the gender would just the remaining data 06:59
off their application information on whether they were approved or rejected. We could predict the gender of the applicants with 82% accuracy. And the 07:09 reason is not 50 + 82 is because there is bison and then so then you have to do data transformation and the documentation 07:18 even before you start modeling to remove them by otherwise. We'll get embedded throughout the process and then when you're doing the modeling itself 07:28 usually machine learning and deep-learning what are optimizing for is the lowest are so that's what the algorithm is looking for. What's the best data 07:38
set with the lowest are in these situations? You have to say. I want lowest airfare plus lowest price so that those two together 07:46 as minimize otherwise by Supreme back in my benefits not in the data because just increase accuracy and then you still have 07:56 to test it. That's one example, or the other issue is interpretability. And what went when an algorithm makes a 08:06 decision. It's usually a black box. Sometimes it's interpretable, but often it's not there's 08:16
no agreement in in the apply to a community or even in the research Community about what interpretability means because if you 08:26 ask Adidas scientists, they can say well it came up with this answer because these nodes in the in the neural network have these 08:36 weights and these weights game because of this data, but that's not useful from a business decision-making perspective. And 08:45 for the way we look at it is how to make interpretability is how to make the outcome intuitive to the user to businesses and 08:55
inform their intuition so that they can and 09:05 obviously the regulatory implications. Also Beyond just building a business intuition that like gdpr in 09:13 California City. Yeah, that'll come on by next year does require interpretability 09:23 AI hype reality. Where are we on that Spectrum at the 09:30 moment? We still in the hype curve and how do we separate but the hype from reality that people are Xanax and and kind of feel I feel down with 09:40 that exist simultaneously right now 09:50
and may for some time. So if you looked at it one of the things I was talking in my house yesterday is how much has 09:59 improved over the last seven years between 2012 and now the capability of the eye has improved 10:09 300,000. Next remorse law more slow looks like that. pipeline so that kind of capability 10:18 needs to be acknowledged that now you can do a lot lot with that but on the other hand, there are also many 10:27 money talks and discussions about the about the idea that are completely height and one of the things 10:36
the first to be able to do is separate I've been reality is just ramped up on how it 10:46 works and I actually have a book coming out for Enterprise AI 10:54 transformation where I spend quite a bit of time explaining that and how how decision-making should work if using AI 11:04 gets used some of that is around affirmation of actions. And those are mostly possible today the majority of 11:14 the applications around learning so you collect. Patterns from data so that when new data comes then you can make 11:24
accurate predictions. So that's the burning after that that part works very well. But when you start to see a eye that looks like it's doing reasoning 11:33 or Can a causal inference that this happened? So that's going to happen. There is no AI today that has 11:42 that capability Smart Systems where does rule ceramic coated but if you look at Deep learning 11:52 machine learning or the other other types, they I know we're close to reasoning so that that's another very quick way to see 12:02
what I think of you. 12:11
Buy this talk
Buy this video
With ConferenceCast.tv, you get access to our library of the world's best conference talks.