Adoption of conversational experiences (chatbots, voice assistants) is directly related to the quality of the experiences that users have with the technology. Even those who have adopted the latest wave of conversational technologies regularly report having poor experiences. Yet, conversation is built into the human experience, and we need to learn how to design experiences that people find useful, safe, and functional.
In this workshop, you’ll learn principles of conversation design and how to apply them into your own conversational experiences. We’ll do exercises to see how an actual design process works, such as building personas, scripting conversations, testing with users, and prototyping. Anyone is welcome, designers, developers, and business people who want to get more insight into exactly how to think through building quality experiences and understand what it takes to engage users.
In this workshop you will discover how machines can learn complex behaviors and anticipatory actions. Using this approach autonomous helicopters fly aerobatic maneuvers and even the GO world champion was beaten with it. A training dataset containing the “right” answers is not needed, nor is “hard-coded” knowledge. The approach is called “reinforcement learning” and is almost magical. Using TF-Agents on top of TensorFlow 2.0 we will see how a real-life problem can be turned into a reinforcement learning task. In an accompanying Python notebook, we implement – step by step – all solution elements, highlight the design of Google’s newest reinforcement learning library, point out the role of neural networks and look at optimization opportunities. The Python notebooks are hosted on Colab. All you need is a laptop with a current Chrome browser and a Google account. We also gladly discuss application ideas you – as an attendee – might bring along.
Most experienced data scientists would agree that data processing takes most of the time when undertaking machine learning projects. Both data pre-processing and feature engineering quality is crucial for model performance. However, it is not typically an easy thing to do. Dealing with real data, you are likely to encounter such problems as noise, missing values, excessive information, etc. Building a good feature vector turns out to be just as hard. In this workshop, you will learn some simple but effective ways of handling these problems using a public Google Play Store dataset as an example.
Since the middle of the 20th century, robots have found their way into more and more areas of human life. From industry and its first robot Unimate over military and warfare to service, care, medicine and household, robots are used either today or in the foreseeable future. The moral questions raised in the construction and use of robots are the subject of the philosophical discipline of robot ethics. Because with robots as specific technologies, the values implemented in them and the (social) consequences resulting from them, moral questions always go hand in hand. In a first part, this lecture gives a critical overview of some fields of robotics and the ethical questions that arise here. In a second step, the fields of work of the philosophical discipline of robotics ethics will be presented, before what has been said is concluded with practical implications and a plea.
Developing intelligent applications is becoming easier based on huge amounts of freely accessible documentation, tutorials, and software frameworks. But what does it actually take to bring these applications into production environments, run them scalable and in high quality and improve them continuously?
We will present how we have developed a chatbot for HR services with rasa.ai, TensorFlow and Keras and organized the development process with mixed teams of product owners, data scientists, dialog designers, data engineers, and machine learning developers with a continuous delivery approach. By using pipelines and version control mechanisms for the different artifacts (code, data, models, parameters), quality gates, and continuous delivery orchestration, we were able to automate the process of continuously developing and improving the chatbot. This enabled us to bring models continuously from the data scientists notebook into production, improving the chatbot and experimenting with live users. The audience will learn how to apply Continuous Delivery for Machine Learning and the great benefits they will get. In the end, we will give an outlook for future developments in software engineering practices for machine learning applications.
Predictive maintenance predicts the future of machines. Using data science we establish the machine’s unique life cycle and increase efficiency. In a world full of machines, we need to be the bridge connecting the methods of the past to the opportunities of the future.
It's not a secret, that deep learning already made a revolution in several perception fields as vision, language and speech understanding and keeps pushing the frontiers. Meanwhile, one important data type which includes time series, digital signals and any sequential observations is still mainly processed with rather standard mathematical and algorithmic routines. In this talk, we will review, what are the main sources of time series in the world, what are the "basic" algorithms and how exactly they might be improved and replaced with different neural network architectures. Apart of the models' details, we will also study the typical tasks that have to be solved while working with time series: classification, prediction, anomaly detection, simulation and others and exactly deep learning can be leveraged to solve them on the state-of-the-art level. Some previous experience with time series/signal processing is useful, but not required.
Image/Video processing is the most popular area in Artificial Intelligence world. Since AI capabilities are developing so fast, some impossible ideas in the past became feasible to implement. With vision of Digital Operator, Turkcell investing so much in Data Analytics where AI is the main focus. In the last 2 years, we have implemented many image/video processing projects in order to increase revenue or cost saving. Examples include but not limited with Passport Fraud Detection, Video Voting, Face Authentication, Car Park Slot Suggestion, Photo Social Media Tag Recommendation, Emotion Analysis, Recruitment Scoring on Candidate Video Resumes, etc. The most important thing regarding these projects is that they are already developed and in production. Also they are measured that they are feasible enough to increase revenue or decrease current cost.
Arguably one of the most beautiful ideas in the Deep Learning revolution of the past decade has been the invention of Generative Adversarial Networks (GANs). This new architecture allows a neural network to learn the distribution of the training data, allowing for the generation of new data samples, entirely unsupervised. While this idea was initially shown to work on small, gray-scale images, the past few years have brought tremendous progress to the domain of GANs both in terms of training algorithms, model architectures and computational scale. The sample quality of modern networks has now reached the point where generated samples are almost indistinguishable from real ones, opening an entirely new era of digital media creation. In this talk I will shed light on: how a GAN actually works and how GANs relate to similar ideas like auto-encoders and auto-regressive models. I will also take a look at the current state of the art of sample generation in both images, audio and other data types, and take an overview of the emerging industry surrounding this novel tool for digital media creation.
Accidents at sea happenall the time. Their costs – in terms of lives, money and environmental destruction – are huge. Wouldn’t it be great if they could be predicted and perhaps prevented?
With more than 350 years of history, the marine insurance industry is the first data science profession to try to predict accidents and estimate future risk. Yet the old ways no longer work, new waves of data and algorithms can offer significant improvements and are going to revolutionise the industry.
In my talk, I will show that it is now possible to predict accidents, and how data on a ship’s behaviour such as location, speed, maps and weather can help. I will show how fragments of information on ship movements can be gathered and taken all the way to machine learning models. I will discuss the challenges, including introducing machine learning to an industry that still uses paper and quills (yes, really!) and explaining the models using SHAP.
The high-level expert group on artificial intelligence set up by the European Commission has been visionary in putting together a set of guidelines for the human-centric, ethical use of AI in our society. Trustworthy AI has three components, which should be met throughout the system's entire life cycle: 1) it should be lawful, complying with all applicable laws and regulations 2) it should be ethical, ensuring adherence to ethical principles and values and 3) it should be robust, both from a technical and social perspective. I am a member of the European AI Alliance, a forum that engages European citizens and stakeholders in a dialogue on the future of AI in Europe. In my presentation I will share the latest thoughts and developments on making sure that AI as represented by machine learning solutions will by be tested and certified and make itself aware to humans as such.
In this talk you will get an introduction into neural networks from first principals using low level TensorFlow 2. From there we will work our way up to TensorFlow's high level Keras API solving the same simple challenge. Along the way you will understand how matrix multiplication is the basis of neural networks, how loss functions work, how partial derivatives can be computed and used to bring down the loss of a model architecture.
The way developers collaborate inside and particularly across teams often escapes management’s attention, despite a formal organization with designated teams being defined. Observability of the actual, organically formed engineering structure would provide decision makers additional tools to manage their talent pool. What is the best engineering team capable of migrating this part of the stack from language X to language Y? What are the most efficient funnels of coding collaborations? On which developers your codebase is relying on? During this talk, not only we aim to identify existing inter- and intra-team interactions but also suggest relevant opportunities for suitable collaborations. To do so, we will rely on contributors’ commit activity, usage of programming languages, and code identifier topics by embedding and clustering them. We will evaluate our approach analyzing codebases of several open source companies. The findings will show that only looking at a codebase, we are able to restore the engineering organization behind, and also reveal hidden coding collaborations as well as justify in-house technical decisions.
Chatbots have arrived in business reality: they offer a simple but powerful self-service approach for customers to handle business transactions. Chatbots are available 24*7, are scalable and can reduce cost significantly.
However: a lot of today's chatbots are not well trained (aka "dumb") and/or offer little benefit to the users. We discuss and demonstrate our approach to address these issues:
Subject matter experts can train the bots themselves, using familiar tools
Use separate bots for specific information domains and integrate them using a concierge service
Have the chatbot handle real business transactions for the user – acting as the conversational interface to complex backend systems
Hand over chats to human agents
Provide an abstraction layer to integrate front end channels, multiple chatbot platforms as well as backend systems
In many companies, the enterprise architecture is very complex and includes a variety of infrastructures, individual software and off-the-shelf solutions, interfaces between these and external systems, as well as numerous documents and other data sources. Here too – in the IT itself – a large amount of data is generated.
Using a business intelligence approach, many companies nowadays own sophisticated mechanisms for analyzing business data to generate knowledge or decision recommendations. For technical data, these mechanisms are only partially being respected and in IT itself, one hardly utilizes data science methods – "The cobbler wears the worst shoes".
In our talk, we show how Lufthansa uses Artificial Intelligence (AI) for architecture management (for instance, to identify obsolete services or optimization opportunities in processes), we discuss the methods in detail, and provide hands-on experience with interesting cases.
We see a major shift coming in the machine learning industry with the transition from Machine Learning (ML) – learning by example – to Machine Teaching (MT) – learning by explanations. Three problems that ML has today will be solved with this transition: (1) Explainability of ML Solutions (2) Data Access and Labeled Data and (3) Accurate control of the algorithms reasoning. Though the talk, Murat will explain the foundational concepts of Machine Teaching along with an account of how this technology will transform and improve many industries with specific use cases and examples of how the technology is being used today.
As long as mankind exists, diseases are also a constant companion, stress often is a cause or at least a co-trigger. In the age of permanent connectivity and IoT, not only our phones and cars are getting smart. Also, the health sector is transforming a lot and gets more and more digital. People now have started to measure their vital systems with smart watches and chest straps and thereby produce plenty of data which can be analysed. This leads to complete new ways of diagnosing diseases and even their causes. Thus, I am going to show you how we used machine learning to implement a model capable of detecting stress in physiological data gathered with the use of a commodity chest strap. Beside that I am going to show you how we put that model in use by detecting stress among our colleagues from Berlin.
Every day at Outfittery, hundreds of stylists use personalised styling systems to create personalised outfits for men all around Europe. They do so by talking to and understanding their customers’ needs, rather than a customer browsing an e-commerce catalogue.
These tools heavily rely on machine learning to perform things such as
Personalisation: Will a customer like a particular type of clothing, will it fit them, do they identify with the brand?
Decision Support: Whats a good shirt to go with these shoes? Does this outfit make sense for the event the customer is about to attend?
Logistics: Is this item in stock? Where in the world is it at the moment, will we have it in time to send it to the customer or should we replace it automatically?
Customer Support: Our stylists talk directly to the customers to understand their needs. Should we do something different based on the feedback they give us?
To make this possible, we have machine learning embedded everywhere to help our stylists make the right decisions and be more efficient. This requires unique data collection, tooling and innovation. In this talk, Steven will explain the various people, products, and system processes to bring this all together.
Honey bee colony assessment is usually carried out via the laborious manual task of counting and classifying comb cells. Beekeepers perform this task many times throughout the year to asses the colony's strength and to track its development. As you can imagine, this is an extremely time-consuming and error-prone task. We will share our experience with the development of a tool for automatic honeybee colony assessment, the DeepBee.
DeepBee is a tool that encapsulates an image classification pipeline using classical image processing methods and state-of-the-art Deep Neural Networks (DNN) for image segmentation and classification. To get to the final solution, we have compared 13 distinct DNN architectures and chosen the best model based on several metrics. We discuss the steps taken from image collection to the delivery of the final solution, highlighting the mistakes we have done during the process, the hurdles we overtook, and the lessons learned. The project has been developed at the Polytechnic Institute of Bragança.
AI and intelligent machines are becoming an ubiquitous part of our everyday lives. However, while emotions play a key role in every human interaction, machines today significantly lack this emotional aspect. Machines that are capable of understanding emotions based on facial expression, voice interactions, body language etc. have the potential to take human-machine interactions to another level that is currently little explored but has immense possibilities, This talk will explore the recent advancements in emotional machine learning, as well as delve into the pros and cons of building emotionally intelligent machines.
This talk is based on a real data science project of mine. The used dataset will have a target column, that is going to be predicted. With the use of several packages like SHAP it will be possible to interpret the predictions. SHAP values are giving us the opportunity to see which features the algorithm took into account to get to the prediction. This very useful package by Scott Lundberg plots also insightful visualizations. As a data science consultant I find this approach very effective, as some of my clients were looking for a machine learning solution that provides interpretable results.
Privacy-preserving machine learning is a subfield of machine learning in which the training of the model happens in such a way that the privacy of the data is preserved. Various approaches already exist but are not well established. At the same time, privacy considerations become more important. Among the approaches is federated learning for a decentralized training, whereby the data can stay at the place of origin and only learning updates or gradient updates are exchanged. Another approach is differential privacy – stochastic gradient descent whereby the learning algorithm of the neural network is modified so that single training examples do not affect the model too much. Thus, limited inference can be made from the model to the data it was trained on. In this talk we will understand both approaches and have a look on how to implement them with the help of TensorFlow.
A person's speech reveals much more than just the content of what they said. With the help of intelligent speech analysis, artificial intelligence can understand how humans communicate.Thanks to machine learning methods, the audio signal can be used to identify human demographic features such as gender or age as well as emotions, personality traits and health conditions.
Going even further - Emotion AI recognizes demographic features such as gender, age, personality traits, and health conditions. AI's engagement with human emotions is especially essential because emotions influence our associations, our capacity for abstraction and our intuition. Emotions affect our well-being, they direct our attention, and they influence our decision-making. Multidimensional emotion modelsare able to recognize more than 50 emotions in real time.AI models also incorporate social aspects and behavioural patterns in their evaluation via acoustic scene recognition.
This provides the basis for a deeply meaningful interaction between man and machine. In her lecture, Dagmar Schuller will shed light on how exactly the analysis works and what potential the technology creates for a wide variety of industries such as health, mobility and devices. The question is not: Will intelligent machines understand emotion? The questions is: Can you call a machine intelligent if it can't understand emotions?
As machine learning (ML) based approaches continue to achieve great results and their use becomes more widespread, it becomes increasingly more important to examine their behavior in adversarial settings. In this talk, we will take a look at everything an ML practitioner should know when it comes to security issues in ML systems. At the end of the talk, you will know what is and what isn’t possible, what you should and what you shouldn’t worry about. We will start with a general overview of security issues in ML systems (eg. poisoning, evasion, inversion attacks), and then focus on vulnerabilities at test time (adversarial examples). We will see what adversarial examples are, what negative consequences they might cause, and take a look at existing attacks on ML systems. We will cover attacks on ML as a service (Google Cloud, AWS), attacks on state of the art face recognition systems, attacks on autonomous vehicles, attacks on voice assistants (Apple Siri, Google Now, and Amazon Echo) and more.
The voice itself lays in the ambivalence of persistence and transition; its sound is gone as soon as it is produced, but despite its inherent fluidity, it is the sonic marker of the identity of the speaker. Who is the person speaking? What can you tell about their personality, their background, their physicality, their gender? Our voices tell a lot about who we are.
The voice is a carrier of meaning. It projects that specific meaning through the sound thus produced and expresses it beyond our control because of its emotional aspect. The sound itself is the medium and acts as a carrier of messages that we consciously or unconsciously decode. The voice is part of an interwoven interpersonal and social system of sounds that creates interaction – a permanent exchange from which one can hardly escape.
The human being encounters the voice every day in different forms. We interact privately and professionally with it, listen to music or radio, make phone calls and use artificial voices, like Siri, which is now part of our everyday lives.
The voice is our acoustic business card – the first impression of a person.
How is the voice perceived and evaluated? Are there differences between genders? What is the meaning of the natural voice and the artificial voice? How can the voice be used authentically as a performative act? What could be the social impact of shaping your own voice?
Modern ML teams face a variety of issues when increasing the team size, scaling computation or trying to comply with different laws and legislation. Tools and workflows that enable version control, traceability, and repeatability for machine learning can greatly reduce risks and help teams reach their goals.
Human / AI interaction loop training as a new approach for interactive learning with reinforcement-learning: Reinforcement-Learning (RL) in various decision-making tasks of Machine-Learning (ML) provides effective results with an agent learning from a stand-alone reward function. However, it presents unique challenges with large amounts of environment states and action spaces, as well as in the determination of rewards. This complexity, coming from high dimensionality and continuousness of the environments considered herein, calls for a large number of learning trials to learn about the environment through RL. Imitation-Learning (IL) offers a promising solution for those challenges, using a teacher’s feedback. In IL, the learning process can take advantage of human-sourced assistance and/or control over the agent and environment. In this study, we considered a human teacher, and an agent learner. The teacher takes part in the agent’s training towards dealing with the environment, tackling a specific objective, and achieving a predefined goal. Within that paradigm, however, existing IL approaches have the drawback of expecting extensive demonstration information in long-horizon problems. With this work, we propose a novel approach combining IL with different types of RL methods, namely State-action-reward-state-action (SARSA) and Proximal Policy Optimization (PPO), to take advantage of both IL and RL methods. We address how to effectively leverage the teacher’s feedback – be it direct binary or indirect detailed – for the agent learner to learn sequential decision-making policies. The results of this study on various OpenAI-Gym environments show that this algorithmic method can be incorporated with different RL-IL combinations at different respective levels, leading to significant reductions in both teacher effort and exploration costs.
Neda has completed her PhD in autonomous driving field from École de Technologie Supérieure (ÉTS), and postdoctoral studies from HEC Montréal, McGill University and Polytechnique Montréal. She has been machine learning (ML) researcher, applied research scientist and data scientist in different research teams. She is also an expert in deep learning, reinforcement learning, supervised / unsupervised learning, natural language processing, computer vision, and time series data. She now works in AI research and development at AI Redefined Inc.
Automated Machine Learning is rapidly becoming a pervasive tool for data scientists and machine learning practitioners to quickly build accurate machine learning models. Recent AutoML products from Google, Microsoft, AutoSKLearn, Auger.AI and others emphasize a programmatic API approach (versus a visual leaderboard) to applying AutoML. All of these products have a similar processing pipeline to achieve a deployed prediction capability: data importing, configuring training, executing training, evaluating winning models, deploying a model for predictions, and reviewing on-going accuracy. With AutoML, ML practitioners can automatically retrain those models based on changing business conditions and discovery of new algorithms. But they are often practically locked into a single AutoML product due to the work necessary to program that particular AutoML product’s API. We propose a standardized automated machine learning pipeline: PREDIT (Prediction, Review, Evaluation, Deploy, Import, and Train). And we walk through a multi-vendor open source project called A2ML (http://github.com/deeplearninc/a2ml) that implements this pipeline for Google Cloud AutoML, Microsoft Azure AutoML, AutoSKLearn, H20 and Auger.AI. We then show building an application and trained model with multiple AutoML products simultaneously using this standard API.
Some projects’ challenges are simply not meant to be solved – just because anything that can go wrong will go wrong. Miscommunications, dataset issues, unreasonably high expectations to name a few – everything will happen. More than that, usually you won't see it coming until it's too late. The question is what to do if you are in such a situation, especially when you have a bunch of competitors.
In this talk, we will take a thorough look at such project to cover all the problems the development team came across and discuss what actually matters when the goal seems unreachable.
Big Data ist der Treibstoff für Deep Learning. Aber was kann ich tun, wenn meine vorhandene Datenmenge zu klein ist, um die Parameter meines Machine-Learning-Modells ausreichend zu trainieren? Data Augmentation ist hier oft die Lösung. Aber wie kann ich Data Augmentation sinnvoll in meine bestehende Deep-Learning-Pipeline einbauen? Warum brauche ich überhaupt eine Pipeline, wenn ich doch Jupyter-Notebooks auf meinem Rechner ausführen kann? In diesem Talk werde ich für Deep-Learning-Anfänger und Machine-Learning-Praktiker Vorteile, Möglichkeiten und Tooling von Pipelines für Deep Learning mit Small Data vorstellen. Dabei wird gezeigt, wie ich Werkzeuge und Prinzipien von Continuous Delivery im Machine-Learning-Umfeld anwenden kann, um mit meinem Machine-Learning-Raumschiff in die Produktion zu starten.
Grundlegende Kenntnisse der Funktionsweise von Machine-Learning-Verfahren
Grundverständnis neuronaler Netzwerke
Schwierigkeiten beim Einsatz von Machine Learning mit Small Data
Continuous Integration Prozesse für Deep Learning und Data Augmentation
Werkzeuge für Versionierung, Deployment und Monitoring von Machine-Learning-Modellen
Traditionally, AI models are trained and evaluated in specific high-performance or cloud environments. Currently, there is a trend towards edge computing which is primarily caused by better processors and hardware acceleration, especially on mobile devices. Furthermore, data protection is becoming ever more important and running AI on the client-side, without data leaving the device, becomes increasingly relevant. With Tensorflow.js models can be trained and evaluated in the browser or in the backend. This enables web applications to use AI online as well as offline. Moreover, there is vendor independent hardware acceleration with WebGL when using the browser. That means no lock-in on CUDA and better performance than running solely on the CPU. In a practical example the complete workflow for developing a gesture classifier will be presented. Participants will gain insights into the Tensorflow.js API and learn how to apply transfer learning. For demonstrative purposes, the classifier will be integrated into a vertical scrolling airplane game. Training and evaluation will be done completely in the browser, so that no data leaves the device. Learning objectives: Build models that can be used in the frontend and backend, understand the power and limitations of transfer learning and understand concepts of modern computer vision models.
OLX is a platform for online classifieds, and millions of users visit it every day to buy and sell from each other. Attractive thumbnail images play a crucial role in making the transactions successful, especially in categories like fashion, where having a clear visual impression of the item is extremely important for buyers. In this talk, we present a system for automatic image cropping and creating attractive thumbnail images. The system is based on a deep learning model that identifies the salient regions of the image, which is used to perform the cropping operation. The new image focuses on the most important aspects of the original, and this leads to higher buyer engagement.We present our journey, starting from a research project done by a master student to a production system. We describe the cropping model itself, the architectural design of the service and the decisions we made along the way. We also cover implementation details and discuss how we utilize AWS, Kubernetes, Python, and Tensorflow to make it possible.
Development of machine learning models happens mostly in Jupyter Notebooks these days. To bring them into production, Data scientists struggle sometimes. In this talk, I will show how to transfer your great Jupyter notebook into a docker image that allows you to train your model locally. This local model will also let you predict locally as a service. The nice benefit of this docker image is its scalability: You can, for instance, upload it to AWS Sagemaker and run it on any instance type you want. I will also show you how to implement a prediction endpoint as a callable API.
BERT is a state-of-the-art natural language processing (NLP) model that allows pretraining on unlabelled text data and later transfer training to a variety of NLP tasks. Due to its promising novel ideas and impressive performance we chose it as a core component for a new natural language generation product. Reading a paper, maybe following a tutorial with example code and putting a working piece of software into production are, however, two totally different things.
In this session, we will tell you how we trained a custom version of the BERT network and included it into a natural language generation (NLG) application. You will hear how we arrived at the decision to use BERT and what other approaches we tried. We will tell you about the failures and the mistakes we made so you do not have to repeat them, but also about the surprises, successes and lessons learned.
Video, audio (multimodal) mobile and edge use cases that utilize machine learning models (e.g. Tiktok, Shazam, Google Home Hub) are becoming more common. However, creating these multimodal ML applications is challenging as developers need to deal with real-time synchronization of time series data during model inference and doing it cross-platform on mobile and edge devices.
Google open sourced MediaPipe in June 2019, a cross-platform applied machine learning pipeline framework that simplifies the development process. My talk will introduce the open source MediaPipe framework, walking through mobile and edge (EdgeTPU coral) demos and getting developers started on building multimodal ML applications.
Deep Reinforcement Learning, also often referred to as (Deep) Q-Learning, has hit the news after winning against human teams in multi-player gaming environments. In this session, I outline the key elements that are the foundation of reinforcement learning and explain how the journey progressed to current successes. Live demonstrations offer insights into understanding Q-learning principles down to the numbers. The presentation will conclude with an outlook into future developments and applications in areas such as logistics and pharmacy.
AI is not new. It has existed for decades now, showing gradual progress over the years. With the consistent growth of the on-demand industry, changing consumer behaviour, and desire for personalization, the logistics industry is constantly solving a problem of delivering products anytime, anywhere. Major trends that are converging AI into logistics are:
Rising SaaS with mobile availability
Self-learning systems resulting in informed decisions
Jugnoo, in particular, is amplifying human ingenuity with intelligent technology by solving complex problems like:
Demand prediction to reduce waste, and meet the service level agreement
Route and Batch Optimization to optimize operational efficiency and reducing costs
Merchant profiling by understanding merchant behaviour to meet the demand and reduce delivery time
Estimated time of arrival by utilizing data and providing faster services by adjusting supply based on the predicted demand