Machine Learning is often hyped, but how does it work? We will show you hands-on how you can do data inspection, prediction, build a simple recommender system and so on. Using realistic datasets and partially programmed code we will make you accustomed to machine learning concepts such as regression, classification, over-fitting, cross-validation and many more. This tutorial is accessible for anyone with some basic Python knowledge who's eager to learn the core concepts of Machine Learning. We make use of an iPython/Jupyter Notebook running on a dedicated server, so nothing but a laptop with an internet connection is required to participate.
This workshop teaches how to break away from direct manipulation in experience design. The materials are based on real-world case studies of products in which a person’s actions are displaced over space and time, illustrating a range of techniques for designing with Machine Learning technologies. Starting with an artificial intelligence story writing workshop, participants will create a product pitch that concludes in a design brief through service design methods consisting of context, human impact, technologies involved, system design, data considerations, visual design and usability, including ethics. Previous experience in user experience or service design methods is not required. Participants attending the workshop should expect to leave the course with a basic working knowledge of designing systems using dynamic data ready for further prototyping and evaluation.
In this workshop, we will dive into the world of Deep Learning. After introducing the basics of Google’s Deep Learning library ‘TensorFlow’ we will continue with hands on coding exercises. You will learn how to create a neural network that is able to classify images of handwritten digits (MNIST) into their respective classes (Image Classification). While starting with a very basic, shallow network, we gradually add depth and introduce convolutional layers and regularization to improve the performance of our model.
Through my own YouTube channel on Machine Learning (https://www.youtube.com/c/ArxivInsights) I recently committed to staying on track with the latest developments in AI and Machine Learning and find the most interesting new trends that will greately impact the future of the field. One of my current focusses is on "How Neural Nets Learn". Here I engage with things like feature visualisation and various ways to fool neural networks into making obvious mistakes called 'adversarial attacks'. These topics are already crucial for many existing ML applications in place today, but will become even more relevant in future domains like self-driving cars, AI-assisted medical care, drug discovery and many more. Large amounts of research groups are currently focussing on tackling the 'black box problem' which is that neural nets are currently uninterpretable: there is no way of knowing why they make certain predictions. In this talk I want to shed some light upon this black box of neural nets and what future progress we can expect!
This presentation shows how to quickly build Machine Learning applications with Python and how we can understand what is happening ‘under the hood’ using Python modules as well. Two examples will be presented: unsupervised and supervised learning for text classification. It is fascinating how fast you can build a text analyzer with Python and Scikit to then apply unsupervised learning. A common approach is to first build numerical representations of the text, and then to apply standard statistical (or machine learning) techniques. I also wanted to know how the intermediate data looks like. Therefore, I built a little example that writes the internal data into an Excel file to better visualize and understand (down to the numerical values) how the feature extraction and the cluster building work together. Another big benefit offered by Python are Deep Learning packages like Keras, which we use for a supervised learning examples. You can quickly set up a complex neural network and have its construction, training and testing in less than 20 lines of code.
The dominant programming language for deep learning is Python. It has a wide variety of frameworks and data scientists love it due to its ecosystem and the workflows it allows. Yet when it comes to actually taking models to production, it is usually met with resistance, as in many enterprise environments Java is still king of the hill – and rightly so. It is the underpinning of big data infrastructure, provides better tooling for production monitoring and scales better to larger teams. Deeplearning4J is both the name of a deep learning library, but also the umbrella for a set of libraries aimed at the production usage of deep learning. Those libraries cover everything from loading data from a variety of data sources, over defining and training your model on a single node with a CPU or GPU or in a distributed environment on a Cluster, to running it in production, and even importing already trained models from Keras and Tensorflow. This talk will introduce the Deeplearning4J ecosystem with a brief overview of the most important libraries, including ND4J, DataVec, Deeplearning4J itself, RL4J (reinforcement learning) and Arbiter (hyperparameter optimization). The overview of the ecosystem is followed by a short history of Deeplearning4J and a look into the near future. Finally, a demonstration of how all of them are used together is given.
Humanity is more than ever confronted with artificial intelligence (AI), yet it is still challenging to find a common ground. By adopting the term intelligence, it has inherited a myriad of issues from the history of psychological intelligence research. AI research inevitably requires an interdisciplinary approach. Our research project aims to understand, measure, compare and track changes of AI capabilities. Therefore, an Artificial Intelligence Model is derived, understood as a system of capabilities to solve problems. It integrates seven categories: explicit knowledge, language aptitude, numerical and verbal reasoning, working memory, divergent and critical thinking. Forms of thinking are classified following Bloom’s Taxonomy. The Artificial Intelligence Scale is developed, reflecting academic IQ-testing procedures. The evaluation through multiple question and answer categories and individual weighting of both is unique, resulting in the A-IQ score.
Tests are executed with digital assistants, independent of their ecosystem: Google Now, Siri (Apple), Cortana (Microsoft) and Alexa (Amazon). Results indicate best overall performance by Siri due to strong working memory skills. Among other results on category level, Cortana leads at explicit knowledge. None exhibit critical or creative thinking skills. A solution to automate A-IQ testing is proposed. Weaknesses and implications for future research and practice are discussed.
Whether you are a data scientist, ML researcher, or developer, Amazon Web Services offers Machine Learning services and tools that are tailored to meet your needs and level of expertise. In this session, Christian will show you a diverse set of capabilities that allow you to easily add intelligence to any application to solve your business challenges.
The Malmo project is a Microsoft-initiated platform for using Minecraft as an AI test environment. This is not about Minecraft as a game, but about an experimental AI platform. In my presentation I deal with the topic Reinforcement Learning and show how to solve individual problems with project Malmo using (Deep) Reinforcement Learning and Minecraft.
In 2014 China’s government announced the implementation of big data based social credit systems (SCS). The SCS will rate online and offline behavior to create a score for each individual and company. Functioning on gamification, they combine powerful tools to influence its users. Today governmental and commercial SCS exist, the government was cited to become world leader of SCS and in 2020 at least one of these systems will be mandatory. While the official goal of the SCS is to level economic development and to bring harmony, sincerity and trust to the whole country, the question is what the “side effects” might be. It displays the ramification of huge amounts of information provided via ICTs and so-called “social media”. They demonstrate possible consequences of the combination of big data and nearly endless storage on the one hand and evaluation by algorithms, artificial intelligence and deep learning on the other. China's SCS are seen as examples of a tendency that exists in most countries: the attempt to solve social problems with technology. This leads to the question how critical thinking and societies themselves can develop in a reality that is constantly rating behavior to create a score that will be defining vast parts of our life.
At Gini we strive to develop a product which will free people from unpleasant paperwork and AI is being of great help on this mission. In my talk I will walk you through how we have implemented Deep Learning technology into our information extraction system to extract information from document images. The talk will shed light onto the network architecture including its convolutional (for image processing) and recurrent (for text classification) components and how they are built into the backend system. I will also share Gini’s experience of bringing DL prototypes developed in the research phase as POCs to actual live product.
Machine Learning on Source Code (MLoSC) is an emerging and exciting domain of research which stands at the sweet spot between deep learning, natural language processing, social science and programming. We've accumulated petabytes of source code data that is open, yet there have been few attempts to fully leverage the knowledge that is sealed inside. This talk gives an introduction into the current trends in MLoSC and presents the tools and some of the applications, such as deep code suggestions and structural embeddings for fuzzy deduplication.
Getting Machine Learning to work let alone turn it into a sustainable business is a real pain in the ass. It really sucks ... But you can do amazing stuff with it! As a Machine Learning consultant I train a lot of people to build their own Machine Learning algorithms and turn them into a customer benefit. And although it is not that hard in and of itself, it is really easy to make mistakes, even for the best of us. In this talk I will highlight some of the most common mistakes and how to avoid them. But if you think you will be able to stop making mistakes if you do everything right, you are wrong! Because Machine Learning sucks!
This session reviews building a Machine Learning pipeline for detecting anomalies of sales point transactions. It is a tricky job for a company like Superonline, that has over 3000 sales points all over the country, distributed across metropolitan and rural areas. An aim is to detect anomalies in transactions because these can lead us to detect fraudulent transactions. The problem splits into two branches, the first is to detect outlying sales points; the second to detect outlying days of a sales point. Detecting the latter is not as complex as finding outlier sales points. Every sales point has a medial number of transactions and a trend. Building a trend line out of past data and calculating a point estimation using the prediction interval method of statistics help us draw boundaries for each day. When a sales point goes beyond these boundaries we raise a flag. For detecting outlier sales points, we need to group them by their type and location and then calculate quartiles and IQR in these groups. When we compare every sales point to these metrics, we see whether they are outlier or not.
Machine Learning (ML) often feels a lot harder than it should to most developers because the process of building, training, and then deploying models into production is too complicated and slow. Amazon SageMaker is a fully-managed service that enables developers and data scientists to quickly and easily build, train, and deploy Machine Learning models at any scale. It removes all the barriers that typically slow down developers who want to use Machine Learning. In this talk, we will walk through a typical ML workflow in Amazon SageMaker, from initial data exploration through training to a running model in production. We’ll also take a look at more than 10 highly optimized algorithms that SageMaker provides out-of-the-box as well as how to bring in your own TensorFlow or MXNet based Deep Learning model or any other custom developed algorithm into SageMaker's toolset. This talk will include both slides and interactive demos.
Data Science: it's actually easier than you think. In this session I'll be your guide through a data science project using Microsoft's team data science process. You'll finally see the parallels between different technologies, both on-premises with SQL Server and in the cloud with Azure Workbench or ML Studio. And you'll learn how easy it is to build and integrate predictive models into day-to-day workflows. To top it all off, you'll experience why data science can actually be very hard, even with tools that aim to make it easier. This all will help you get started easily on that idea or project that's been lying on the shelve!
Deep Reinforcement Learning leverages the strenghts of Neural Networks to enable robots and machines to learn how to operate optimally in complex, dynamic environments. This session will provide a general introduction to the field, look at some of the most promising applications in industrial process optimization and intelligent robotics as well as the most challenging R&D problems we are facing today.
Many companies are exploring paths to become a player in AI business. If there is no experience and expertise in the company already, this means a decision on the right way to go:
Building up a new AI team inside the company. Experiences are that this can take a very long time.
Buying a smaller company with already proven expertise. This is becoming a more and more costly approach.
Founding a start-up outside of the main company and later integrating it. This can evolve to a risky approach.
We propose an embedded approach that allows a quick start. From our experience, the availability of training data on the identified business case is the most challenging bottleneck. Therefore, the combination of good performances on a small amount of data of one's own business case and the ability to quickly increase data volume are highly important for a quick start. We address the steps of management involvement, definition of business ideas, collection and growths of data for training, and roll-out. We also explain the relation of data volume and capability of the AI model and provide insights on how to address this fact in the process of business development .
Deep down ML is a pure numbers game. With very few exceptions, the actual input to an ML model is always a collection of float values. This is straightforward for numerical, spreadsheet-like input, images where pixels are just numerical color values or audio samples, but how do ML algorithms work on words and letters? As proper preprocessing is often the most crucial part in a successful ML project, it is important to understand how to handle textual input properly. We will have a look at the two most important jobs when handling text in ML: preprocessing/normalization and vector representations of text. We will first navigate the minefield of correct Unicode normalization of our input and then - after we have tamed our strings - how to convert normalized and sanitized strings into various vector representations, from simple one-hot encodings to embeddings produced by algorithms like Word2Vec.
Conversational UIs is an important concept for future apps. Yet, adding support for natural language understanding is hard. Microsoft offers the Language Understanding Intelligent Service (LUIS) that should make natural language support feasible for enterprise apps. The Machine Learning-based service offers a cross-platform Web API so that it can be used from any platform and technology. In this session, long-term Microsoft MVP and Regional Director Rainer Stropek starts by introducing the fundamental concepts of LUIS. Based on this, he shows live-coding demos where LUIS is used to build Bots.
Telecom is one of the domains that have the richest available data. CDR (call detail records), usage patterns, customer documents and call center messaging etc. are all part of the large volume of data which is sometimes hard to predict or to find a meaning in. In this session, we'll describe two production ready real-world use cases in a telecommunication domain. One of them is verifying a port-out document based on image processing. Port-out documents can be misleading due to frauders or faulty scannings. In this task, we'll provide a solution on how to preprocess these images and then build a convolutional neural network model. Another one is building a retrieval-based chatbot using text preprocessing, word embedding, convolutional neural network and sequence-to-sequence learning.
While analyzing structured data (even tremendous amounts of it) is a solved mystery nowadays, retrieving actionable insights from unstructured data (i.e. text) is the new challenge to be met. This talk even goes one step further and places this challenge in a streaming data setting. A reference architecture that works across industries will be shown to illustrate how to process text immediately after being written, how to analyze it, how to gather its meaning, and eventually visualize the results to provide actionable insights. This architecture will be composed of several open source projects. When combined, these tools are capable of accomplishing this ambitious task of analyzing streaming unstructured data. The talk will be completed by a live demo that showcases how real-life customer reviews can be processed in real-time to do sentiment analysis on unstructured data and display the results on a dashboard to provide actionable insights.
What about making your app smarter without any knowledge in AI? Thanks to pre-trained models, Machine Learning APIs can automatically analyze your data. In this session, you’ll see how to transform or extract information from text, image, audio and video, and you’ll be an active player of a live demo. Do not put your smartphone in airplane mode!
In this interactive panel, all participants of the conference have the opportunity to ask their questions and make their comments. An expert panel will try to answer the questions and, if necessary, moderate a discussion on the future of Machine Learning. It’s your turn!
Artificial Intelligence has an impact on all areas of society, across a broad range of applications. NVIDIA invests both in internal research and platform development to enable its diverse customer base across gaming, VR, AR, AI, robotics, graphics, rendering, visualisation, HPC, healthcare and more. Working with developers in academia, enterprise and startups, Alison's talk will cover the hardware and software that comprise NVIDIA's GPU computing platform, across PC to data centre, cloud to edge, training to inference. The talk will also detail current state-of-the-art research and recent internal work combining robotics with VR and AI in an end-to-end simulator.
In this talk, we will present three different strategies for finding concrete use cases at medium to large companies. These strategies are: finding problems and see if they can be solved with Machine Learning, mapping known solutions to your company, and targeting processes that are currently driven by business rules. These use cases should be evaluated by their inherent potential, the risk involved and the estimated work needed. Typically you lose 75% of the potential use cases after simply checking basic requirements like availability of data, potential of business process integration, technology readiness ... Next we will present the Machine Learning Maturity Assessment: an evaluation framework for evaluating an organisation on their ML use. In that framework, four dimensions are important: strategy, people, data and legal. These dimensions are then scored on objective subcriteria according to fixed maturity levels. This tool helps companies to find their weaknesses, strengths and source of uniqueness. Finally, we will show you why you need to integrate AI and Machine Learning consistently as a process, instead of just as a one off. We will show real life examples of how these techniques were applied to large companies that are active in telco, automotive and finance industries.
In the "Augmenting Business Intelligence at the Edge" sesson, I will talk about how AI and Machine Learning can unlock unprecedented business value from IIoT (Industrial Internet of Things) sensor data. The cycle of exploratory visual analytics, numerical encoding and embedded rules/models enables real-time surveillance and statistical process control. In combination with digital twins, this cycle supports the understanding and fine tuning on software representations of physical assets. Combining sensor data with visual analytics tools lets operators see patterns of equipment productivity, degradation and stoppages. The embedding of numerical models, encoding such patterns enables continuous surveillance, decision support and automated interventions when degradation is detected. Such technologies augment business intelligence at the edge, driving operational efficiency and ROI.
Attention mechanisms were the key to improving the state of the art in many sequence modelling tasks, like machine translation and image captioning. Besides that, they also allow to make a model more interpretable by showing which sequence elements were the most important when making a decision. This talk will give you an overview of the different kinds of attention and how they are used. The complex math is skipped in favor of a more intuitive explanation, which should provide you with an understanding of how the different concepts fit together. Finally, a working example of using Deeplearning4J with attention is demonstrated and explained.