Machine Learning fürs Business: So wird es ein Erfolg - Sparx S2/E7, Ralf-Dieter Wagner

Machine Learning fürs Business: So wird es ein Erfolg - Sparx S2/E7, Ralf-Dieter Wagner

Show Video

Machine learning is on everyone's lips. But how can many more projects actually be made productive? My name is Ralf-Dieter Wagner, I am responsible for the go-to-market for our AI and ML services in the German-speaking region at Amazon Webservices. Machine learning is on everyone's lips. But what is it actually? Machine learning is a subcategory of the broad term artificial intelligence. And if you ask ten people for a definition of artificial intelligence, you will get eleven answers. My preferred answer goes as follows: Artificial intelligence enables computers to imitate, to mimic human intelligence. 

Machine learning is a subcategory, as already mentioned, that uses patterns in data to recognise correlations from which to develop its own logics and its own programmes. Deep learning is another subcategory of machine learning and uses neural networks, deep technologies that mimic the way the brain works, to generate solutions to customer problems and business challenges. Why is the topic so relevant today? If you think about it, the term artificial intelligence was coined in the 1950s and has since experienced some high phases, but also some low phases. Why is the topic so relevant today? Well, we are currently seeing a convergence of monumental technologies. On the one hand, we have many different machine learning frameworks and technologies that are evolving very quickly at the moment. On the other hand, the computing power available for teams to even apply machine learning is evolving dramatically.

And digitalisation is making it possible to produce more and more actually usable data. And data is definitely an important ingredient for machine learning. On top of that, all these topics are also available in the cloud – the cloud, which enables users to quickly, cost-effectively and very flexibly implement machine learning in reality. What types of machine learning do we know? On the one hand, there is the whole issue of supervised learning – supervised because you don't just give the algorithm the input data, but also the desired result.

So it's like when a teacher already gives you the answer to the corresponding question about initial problems. And what problems does it solve? Typically, supervised learning is very well suited to support classifications. Is it red or yellow? Is it fraud or is it not fraud? These kinds of questions, these classifications are a prime example of supervised learning, but also demand predictions.

How long is my waiting time outside the restaurant? Or in other words: predictions of concrete values. The second topic is unsupervised learning. Here, the algorithm receives the data, but not yet the answer as to what is right or wrong.

And the algorithm and the machine then try to recognise patterns in the data and find out: what is similar? Subsequently, the algorithm can, for example, merge these topics into customer clusters, i.e. identify all customers who have similar needs or who react similarly. Or it can recognise: what is not similar? What are anomalies? Which is very important, for example, in the area of production quality control, predictive maintenance, to find out when machines may need maintenance or will break down. The third category is re-enforcement learning: In this case, I don't need a concrete answer, as in the previous two cases, but rather a strategy. How can I teach a robot to walk? How can I learn to play chess? This is a quite complex type of machine learning, geared towards the development of strategies. We see that machine learning is currently already driving transformations for our customers across every industry, every business function and every digital product. And we are not alone in this.

All the major industry analysts are predicting that the use of machine learning by companies will increase dramatically, that the corresponding revenue targets, the business profits that the various companies in the market want to achieve by using machine learning will increase dramatically. So, all is well? No. From our point of view, all is not well yet and we definitely see clients struggling with the application of machine learning.

On the one hand, it's the sheer amount of data that's being generated – actually that's good news, because machine learning really thrives on being fed lots of data. But many clients are struggling to cope with this mass of data in the first place. The data comes from very different sources: social media, video formats, log data from web applications, etc. This means that you have a multitude of data from very different sources that now suddenly have to be processed. Unfortunately, these often accrue and are kept in silos which makes data access for data analysis very difficult.

In addition, many companies are not yet ready to actually use machine learning because of their mentality and organisational structure. Projects are often not yet carried out in interdisciplinary teams, which makes it incredibly difficult to harness the full power of data. And we haven't even spoken about the question of the lack of skills, the ability of the teams to master the various technologies. Data security, data governance, secure access to data are of course also topics that concern all customers. As AWS, we have opportunities to accompany our customers on this path and support them in the implementation of their machine learning projects.

If you look at the history of Amazon itself, you can see that Amazon is actually a prime example of a company that has gone through a machine learning journey over the years. Machine learning is used in every area of Amazon's business and is the true core of the company's value creation. The personalisation of the customer experience on the web, the corresponding prediction of demand – quite detailed and granular. So that the supply chain can be controlled accordingly and the entire logistics processes can be neatly balanced. These are very fundamental functions of Amazon.

But also how to interact with customers in a whole new way: via chatbots. And here we might mention Alexa as a voice control system using natural language processing, i.e. corresponding machine learning methods. This has put the way customers interact with businesses on a completely new footing. Groundbreaking technologies in the field of image recognition, video recognition and autonomous flight, such as drones for the delivery of packages, complete the picture of a journey that Amazon has made to really use machine learning across the board.

Did this happen overnight? Of course not. Is this a journey that was ultimately anchored culturally in the company? Definitely. Where do we see the core issues that are present in today's market? I said at the beginning that the term machine learning was coined back in the 1950s.

So is it just new maths on faster machines? No, it's not. We see an incredible amount of investment and innovation happening in the market right now. All designed to reduce the barriers to entry for customers to use machine learning. For example, we see that there is a vibrant ecosystem to push machine learning frameworks further and minimise the implementation there accordingly. We see that there are many approaches to either develop new algorithmic methods or to address the weaknesses that are ultimately inherent in machine learning in terms of systems engineering.

To give an example: The question of trusted AI explainability, i.e. how I can explain how an algorithm ultimately arrived at a result, is at the forefront of many clients' minds. Here, too, we as AWS have invested very heavily to support the explainability of algorithms in terms of systems technology.

To check in advance whether the data used in algorithmic procedures for training already has structures that are alien to reality. We also see that the whole topic of data strategy and data governance is very important and that end-to-end data structures are being created in order to be able to use machine learning in a truly scalable way. Being able to do machine learning not only in the cloud, but also on the shop floor, in the production line, at the edge, is important to many of our customers and that's why we will continue to innovate in the market to make that possible. Two big topics to conclude the look at the market today: On the one hand, how can I really scale machine learning? How can I actually transfer the themes anchored in software development today, such as DevOps, to machine learning? Here, we see that the term ML-Ops is evolving in the market.

We support this with our software suite to help ensure that automation and scalability of machine learning can be implemented. The second issue is: how can I enable our customers to use machine learning without them having to develop it themselves? So, how can I provide API based machine learning to customers? We as AWS are investing very intensively in individual areas to make exactly that possible. Examples of these areas are personalisation of the end customer experience, forecasting of demand, image recognition processes, text recognition processes, which we make available to customers as API-based services. So that you don't have to develop these services again yourself. We have divided the technology offering for our customers into three areas, none of which are mutually exclusive. And we also have customers who demand all three areas, depending on their needs.

The first area is an area of infrastructure where customers ask us for fast, low-cost infrastructure to run frameworks, services of their choice on. This area addresses customers who ultimately want to manage every aspect of the ML workflow independently. The second area is our ML managed service called SageMaker. It addresses all the necessary stages end-to-end of the ML workflow – from data import to monitoring the models in production. And helps our customers to automate as much as possible the things that do not directly influence the added value in the use case, but are necessary to do ML at all.

Examples include managing the infrastructure – ramp up, ramp down, dynamically as needed, preparing data and the appropriate tools to make this as efficient and effective as possible. And the aforementioned monitoring of the models in production to ensure that the models do what they are designed to do. The third area is AI services, which in many cases are already pre-trained. They can be accessed by our customers via simple API calls and allow them to incorporate ML capabilities into their processes and programmes without having to know ML themselves. These are, for example, services around the topic of personalising the end customer experience or forecasting demand at different points in the supply chain. These are methods for automated image recognition, methods of speech recognition, also natural language processing methods, based on the technology of Alexa.

This means that with this offer, our customers can choose from these three areas, depending on the use case and the individual customer situation. The use cases we see on the market are diverse and very customer-specific and cover the entire range of our customers' business processes. At BMW, for example, we support the customer in using sensor data, telemetry data, to predict the susceptibility to maintenance, the service life of parts in the car, and in order to make suggestions, if necessary, as to when the appropriate maintenance should be carried out again for parts or for the entire car. Zalando uses our technology to personalise the end customer experience to provide them with an optimised customer experience when visiting the Zalando website. The German Football League uses our services to produce real-time analyses from match data in order to make predictions.

Will the attack be successful? What is the probability that a goal will be scored on one side or the other? As I said, there are basically no limits to the possible applications. What do we see as success factors for anchoring ML in the company in a truly scaled way? One very important thing is certainly that machine learning is not just seen as a technology, but that there is real sponsorship on the part of the business units. You also have to remember that machine learning is often used to try things out, to drive innovation. This means that not all projects will be successful. Innovations are tried out, some work, some don't.

This presupposes a culture that allows things to be unsuccessful at times. It is essential to embed this culture in the company. The second important area is the question: what do I use machine learning for and what do I not use it for? Machine learning is not the solution to all problems in business.

If business problems, challenges can be solved simply by static business rules, "if than else" – if this happens, then do that – machine learning should not necessarily be used. The same applies if you have to be 100 per cent right all the time. When using machine learning you should at least keep the human in the loop. Because machine learning involves statistical procedures and is therefore not necessarily always 100 per cent correct.

Data availability – secured, scalable to the right business problems – is essential for machine learning. We need to make sure that we implement a data strategy, data governance that allows only the appropriate staff who absolutely need to have access to certain data to ultimately have that access. Skills, capabilities, ML techniques are a huge topic for many of our clients. We see that often only one type of skill is built up: the classic data scientist. Projects become successful when different skills that are needed throughout the ML environment are brought together to form teams.

I need the product owner, the business owner, who is ultimately responsible for this requirement. I need the researcher who knows, and can perhaps do, deep analyses about the relevant implications and opportunities that ML offers for this business problem, and bring these issues to the table. We need the data scientist to develop the models. We need data engineers who prepare the data and we need ML engineers who ultimately anchor the ML code in the technical infrastructure and integrate it into the corresponding development process in such a way that the models can be put into production in an auditable, repeatable and scalable manner. And last but not least: how can I make it possible for the teams to concentrate on the essentials? How can I enable teams to automate away the things we call at AWS the "undifferentiated heavy lifting" – the things that don't add value – so that the teams are given the freedom to do what is relevant to the business.

How do we support our customers on AWS besides the technology, the different offerings to different customer needs that I have outlined? In order to approach the topic of ML technologically, we naturally also offer all other measures that are necessary to position ML successfully. On the one hand, training and enablement: We have online training, our own ML Universities, training courses, deep-dive sessions, certifications of all kinds to enable our clients' employees to introduce and use machine learning in the company themselves. But we also help our clients to develop the idea of what the first, best, most effective areas of application of ML could be. We offer executive and working level workshops to define precise use cases, develop business potentials and evaluate them accordingly to help our clients take the first steps, which we then together turn into prototypes, proof of concepts and minimum buyable products. Then, together with our customers, we can also implement these initial solutions productively via our professional services teams and implement further use cases for our customers.

Thank you very much for your interest. 

2022-02-23 17:19

Show Video

Other news