5 Lessons Learned from Implementing 40+ Machine Learning Projects – Xiaopeng Li



all right thank you guys for coming my name is Xiao Hong I'm gonna share some lessons learned from us in Metta implementing more than 40 plus machine learning projects during the past four to five years so there we go who is this guy I you might wonder so I'll just say a few words about that I'm working as an AI advisor at in Mehta and what is an AI advisor think about management consultant but less and more mathematics so that's basically the definition of myself at least and then also I'm very passionate about the building tech communities so you might have seen me before in one of the meetup groups so I will say I meet up and one of the organizers the same goes for absolute machine learning and absolu IOT so if you haven't joined any of those meetups I encourage you to search the monk kugel and you can easily find them Mount meetup.com or their own web site and be part of the community and we need your contribution all right so why am I here I wonder so you might have the same question I an AI and machine learning has been so hyped like way too hyped if you look at the the Gardiner hype curve where I at the top so into l2 all 18 the report published by Gartner they have placed different emerging technologies on this curve and deep learning is right on the top what's gonna happen next this is this is like a roller coaster ride you're building up climbing on the top the next step is gonna fall so there will be a falling of actually the gap between the expectation and the actual implementation and there has been so much higher height buzz and crap going on so I just want you have my you know minimal contribution to the community to cut some crap so that's why I'm here just one example showing you the hype of AI right now so only in Europe the top 13 economies there are more than 2,800 startup beginning of this year and now it's probably 3,000 or even more as you can see UK has the the most startups right now focusing on AI and even in Norway being the the smallest within the 30 largest economy in Europe we have more than 30 startups focusing on AI so this gives you a hint on how much hype and also investment and capital is going into AI and version learning right now as we speak but more interestingly there are many companies who claim to be doing machine learning or AI and they're not doing AI at all according to a uk-based research firm and also branch venture capital MMC ventures they have investigated into I think around for 4,000 startups which claim to be working with air and Martian learning across Europe and they have found out more than 40% of them are not doing anything AI at all this is kind of cigar stick I mean basically this is like the trend we have seen with blockchain when you put your net when you put plug in as a keyword in your startup name or your investor deck you have much higher chance of getting money this is kind of sick because what we want to see is real implementation not this type of hype so what am I going to talk about rest assured there will no be no buzzwords or were you know futurism or killer robots or destruct exponential disruption whatsoever I hate those words I really find it difficult to pronouncing them I share you so I will share some hands-on learnings from our implementation add in meta together with clients and also our partners so just a little bit background we have a summation we have done more than 40 plus machine learning projects in the past five years and any meta we have 20-plus data scientist there are engineers and AI advisors I think this is one of the largest data science competency poll in you know it right now and also everything we have learned I have summarized them into five lessons that I want to share with you today in the next five minutes no pressure so just to give you an impression on the different industries we have been working with within those 40 machine learning projects you can see a lot of from maritime healthcare retail etc so this is how we work with merging learning projects add in meta so we normally work with six steps starting with defining the problem defining clear business objectives and also conceptualize a solution and then that's the data preparation you need to explorative lea analyze the data and then also transfer the transform the data to the ideal a format that you want to train your model on and things about constructing the model and training the model before going into evaluation the model post from technical point of view and also business point of view next step putting the model into production and integrating the model into your business process that's what we call operationalization and last but not least continuously optimizing the model if you work with AI and machine learning then this is very familiar to you this is no rocket science but this gives you the basis for my lessons learned later on that's why I'm mentioning them right now alright so first lesson also just to explain how did I come with those conclusions I have conducted an internal survey to the whole department everyone who have been part of the four-year journey working with 40 plus merchant learning project so the conclusions I'm sharing with you are from our point of view so you might be biased due to the type of project we work with but it's from firsthand experience it's not from external research company not gotten or not Microsoft first lesson data preparation takes more time than any single step in the Mersey learning process I think this is not something surprising for you but this is good for you you know because next time when you plan for a merchant learning project making sure you plan enough time for this step and I can easily tell you that our data scientists have to spend more time crunching the data preparing the data then actually beauty mode and training the model this might be a little bit frustrating for some data scientists but this is the fact that we need to admit and we need to have this into consideration when planning and executing for a machine learning project so this is the first learning second from a technical point of view what are the most challenging steps in the six steps we talked about so data exploration transformation as I mentioned this will be an explorative process and you need not only their size you know technical skills but also visualization also domain knowledge to understand what data that you are dealing with and transforming the data set to actually to be suitable and relevant for the business objective that you have defined to solve that also putting models in production is surprisingly challenge challenging compared to actually building and training the model so this is also good to be taken into taking into consideration when you are planning and executing your next merging learning project also good to mention is that actually many of the machine learning project you might have seen or have heard about stop here they don't really go into production that's also very frustrating because many companies don't have the resource or ambition to move things into production but only when you move this into production that's when you start to realize the value from merging learning on a daily basis in your business process so that's my comment so next ensuring clearly defined business objectives is the single most important success factor in an emerging learning project I've seen some projects also within meta when we started four years ago where we skip this step and go straight into working with the data and also beauty model and an training mode etc and oftentimes what we need to go back and actually clarify what are we trying to do what are we trying to solve who is our target user and and what pain points are we addressing and how we are improving the current business process using machine learning and AI or actually there are simpler solutions that we can use statistical models etc not merging learning at all so this need to be defined upfront so that you know you are solving the right problem which will create business impact and value and also machine learning is actually suitable for solving this problem so if you remember only one thing from today's presentation this will be that clearly defined business objective will be the single most important factor for success immersion learning project so since we have been working with so many different organizations across different industries we have also learned a little bit perspectives that is broader than just a merging learning project so basically the challenges and also struggles organizations my face when they embrace AI in general so one thing we have learned is those five are the most common challenges that organizations are struggling with when they implement merchant learning we already talked about this part you need a clear objective in order to succeed and data AI machine learning is all about data without data you will not be able to create a value I mean people say you know junk in junk out when you train a model so making sure you not only have the volume of the data you need but also the quality of data you need and here also comes in the perspective of AI effects and responsibilities because to a very large extent when you have bias the results of your mercy learning solution is normally introduced from the data set that you are using so they so it's good to keep in mind when we talk about the quality of data not only from a technical point of view but also from an ethical point of view and also as I mentioned already the inability to move things into production really prevent a lot of organization of organizations from realizing value of a and the merchant learning so next time when you are involved in a emerging learning project in your company or in your clients organization making sure to convince them to move things into production and this is when you start to reap the benefits of merchant learning for real last but not the least many companies and organizations organizations trade AI as experiment experiments : and that means they don't have real commitments of implementing this or scaling this across the organization they just want to you know have one or two data scientists play with some merchant learning magic and maybe sprinkle a little bit mercy learning our existing business that doesn't work you really have to you know making sure you're committed to conducting this merging learning project into production and also measure your results maybe not the same way as you measure software development but still you need to define success criteria and then you will have the commitment from the management to actually sustain your AI efforts going forward so defining clear success criteria and convincing your business stakeholder will be a very important step within version learning implementation process the last lesson we have learned working very broadly with different companies of course Industries is that the key reason for many organizations to struggle with the implementing or adopting AI machine learning in general is a lack of overall data and AI strategy this is important for you to know as well because your manager might be on the business side were your stakeholders might be on the business side they might see emerging learning AI as a only as a tool but not as an element in their business strategy and that's kind of risky because then they are not taking this into consideration from its impact or from its implications and both from business and technical point of view so making sure you have data and AI as one key elements in your digitalization program or in your corporate strategy is crucial you to succeed as an organization when you embrace AI and immersion learning so here I just want to quickly show you this model actually published by Microsoft together with PwC they have identified the key gaps in AI and the merchant learning related the competency across Europe and as you can see advanced analytics when it comes to data science they're engineering that's one gap ail leadership that's another data amendment that means your governance process and infrastructure around the hand and they immersion learning so I'm showing this because when you have an data and AI strategy in place and that means you will have all those elements covered then you have but much better chance in mitigating the gaps in your organization and your future State and actually start to realize value of machine learning and AI in your organization if you forget to take photos here's a summary just for your information so now it's the time for you to capture it if you want you that's the five lessons I want to share with you today and if you want to get in touch we can talk later Thanks [Applause]

Leave a Reply

Your email address will not be published. Required fields are marked *