Human-in-the-Loop (HITL) Learning

Jump to: navigation, search

YouTube search... ...Google search

Human-in-the-loop (HITL), basically you can say, is the process of leveraging the power of the machine and human intelligence to create machine learning-based AI models. HITL describes the process when the machine or computer system is unable to solve a problem, needs human intervention like involving in both the training and testing stages of building an algorithm, for creating a continuous feedback loop allowing the algorithm to give every time better results. What is Human in the Loop Machine Learning: Why & How Used in AI? | Vikram Singh Bisen - Medium

Example use case:

  • Limited data for use
  • Uncomprehensive data
  • Interpretation required
  • High liability mistakes
  • Rare objectives
  • Uncommon objectives
  • AI functional inexperienced

Creating a virtuous feedback loop from your application to the AI services enabling it can help the system improve automatically without significant investment.Create an AI feedback loop with Continuous Relevancy Training in Watson Discovery | IBM

Human in the loop Workflow Automation
A talk given by Tina Huang from Transposit at the 2019 Platform Summit in Stockholm. Manual workflows are productivity killers. Automation has become our North Star. But many workflows can only be partially automated; they may benefit from or require human intervention. APIs give us the levers we need to build great automation. Applications provide the interfaces that pull humans into the loop at critical junctures. Composition is how we turn those APIs into apps that integrate disparate applications, simplify workflows, notify people, and respond to interactions. In this talk, we’ll demonstrate how we’ve built apps that allow for human-in-the-loop automation and show you how using an API composition platform like Transposit can help you efficiently build apps and bots that automate the tedium and let humans add maximal value.

dabl: Automatic Machine Learning with a Human in the Loop |SciPy 2020| Andreas Mueller
In many real-world applications, data quality and curation and domain knowledge play a much larger role in building successful models than coming up with complex processing techniques and tweaking hyper-parameters. Therefore, a machine learning toolbox should enable users to understand both data and model, and not burden the practitioner with picking preprocessing steps and hyperparameters. The dabl library is a first step in this direction. It provides automatic visualization routines and model inspection capabilities while automating away model selection. This talk will introduce the dabl library and show how to use it to quickly create supervised models and identify modeling and data quality issues.

Easily Implement Human in the Loop into Your Machine Learning Predictions with Amazon A2I
Companies have millions of documents to process along with various types of documents. Often times, these documents are hard to read or have specific data points which are required to complete the business process. Using Amazon Augmented AI, you can now implement human reviews to review your machine learning predictions from Amazon Textract, Amazon Rekognition, Amazon SageMaker and many AWS AI/ ML services. In this tech talk, we walk through how to implement human reviews as well as showcase a use case by DealNet Capital on how they were able to reduce review time by 80% implementing Amazon A2I. Learning Objectives: Learn how to implement human reviews, Understand how Amazon A2I can work with other machine learning services, Learn how DealNet used Amazon Textract and Amazon A2I to process loan applications. To learn more about the services featured in this talk, please visit:

MLOps #15 - Scaling Human in the Loop Machine Learning with Robert Munro
Human In The Loop Machine Learning and how to scale it. This conversation talked about the components of Human-in-the-Loop Machine Learning systems and the challenges when scaling them. Most machine learning applications learn from human examples. For example, autonomous vehicles know what a pedestrian looks like because people have spent 1000s of hours labeling “pedestrians” in videos; your smart device understands you because people have spent 1000s of hours labeling the intent of speech recordings; and machine translation services work because they are trained on 1000s of sentences that have been manually translated between languages. If you have a machine learning system that is learning from human-feedback in real-time, then there are many components to support and scale, from the machine learning models to the human interfaces and the processes for quality control. Robert Munro is an expert in combining Human and Machine Intelligence, working with Machine Learning approaches to Text, Speech, Image and Video Processing. Robert has founded several AI companies, building some of the top teams in Artificial Intelligence. He has worked in many diverse environments, from Sierra Leone, Haiti and the Amazon, to London, Sydney and Silicon Valley, in organizations ranging from startups to the United Nations. He has shipped Machine Learning Products at startups and at/with Amazon, Google, IBM & Microsoft. Robert has published more than 50 papers on Artificial Intelligence and is a regular speaker about technology in an increasingly connected world. He has a PhD from Stanford University. Robert is the author of Human-in-the-Loop Machine Learning (Manning Publications, 2020)

he REAL potential of generative AI
What is a large language model? How can it be used to enhance your business? In this conversation, Ali Rowghani, Managing Director of YC Continuity, talks with Raza Habib, CEO of Humanloop, about the cutting-edge AI powering innovations today—and what the future may hold. They discuss how large language models like Open AI's GPT-3 work, why fine-tuning is important for customizing models to specific use cases, and the challenges involved with building apps using these models. If you're curious about the ethical implications of AI, Raza shares his predictions about the impact of this quickly developing technology on the industry and the world at large.

Thanks to Raza and Humanloop for joining:

MobiSys 2020 - Human-In-The-Loop Reinforcement Learning (RL) with an EEG Wearable Headset
Mohit Agarwal @ ACM SIGMOBILE ONLINE Presented at MobiSys 2020

Humans-in-the-loop: improving artificial intelligence through human intelligence
"A colleague approaches you and requests help with an anomaly detection system on some unstructured data. They tell you the data quantity is sparse but they need to get some sort of solution in place as soon as possible. What now? Often times when solving problems for customers, the initial data is small, imbalanced in class, and there may be little room for error. This is especially true when looking to catch infrequent anomalies. There are many techniques that can aid with these issues such as transfer learning, data augmentation, and data synthesis. However, these techniques may only get you so far initially with a generalized model. In order to get the most out of existing data as well as leverage the expertise internal to the business, human-in-the-loop (HITL) machine learning systems can aid in the effort. In this topic, we'll discuss how HITL systems can be structured, how they help drive more customer engagement, and help deliver more robust solutions from your data team."

Practical Human-in-the-Loop Machine Learning
Learn more about the AWS Partner Webinar Series at - Join us to learn why Human-in-the-Loop training data should be powering your machine learning (ML) projects and how to make it happen. If you’re curious about what human-in-the-loop machine learning actually looks like, join Figure Eight CTO Robert Munro and Amazon AWS machine learning experts to learn how to effectively incorporate active learning and human-in-the-loop practices in your ML projects to achieve better results.

Stanford Seminar - Human in the Loop Reinforcement Learning
Emma Brunskill Stanford University Dynamic professionals sharing their industry experience and cutting edge research within the human-computer interaction (HCI) field will be presented in this seminar. Each week, a unique collection of technologists, artists, designers, and activists will discuss a wide range of current and evolving topics pertaining to HCI. Learn more about Stanford's Human-Computer Interaction Group:

Efficient Deep Learning with Humans in the Loop
Zachary Lipton (Carnegie Mellon University) Emerging Challenges in Deep Learning

Towards ambient intelligence in AI-assisted healthcare spaces - Dr Fei-Fei Li, Stanford University
Artificial intelligence has begun to impact healthcare in areas including electronic health records, medical images, and genomics. But one aspect of healthcare that has been largely left behind thus far is the physical environments in which healthcare delivery takes place: hospitals, clinics, and assisted living facilities, among others. In this talk I will discuss our work on endowing healthcare spaces with ambient intelligence, using computer vision-based human activity understanding in the healthcare environment to assist clinicians with complex care. I will first present pilot implementations of AI-assisted healthcare spaces where we have equipped the environment with visual sensors. I will then discuss our work on human activity understanding, a core problem in computer vision. I will present deep learning methods for dense and detailed recognition of activities, and efficient action detection, important requirements for ambient intelligence, and I will discuss these in the context of several clinical applications. Finally, I will present work and future directions for integrating this new source of healthcare data into the broader clinical data ecosystem. Fei-Fei Li is a Professor in the Computer Science Department at Stanford, and the Director of the Stanford Artificial Intelligence Lab. In 2017, she also joined Google Cloud as Chief Scientist of AI and Machine Learning

"The human-in-the-loop model combines the best of human intelligence with the best of machine intelligence. Machines are good at accurately and efficiently processing large amounts of data, but are not able to make smart decisions if not provided with vast knowledge. On the other hand, humans are inaccurate and slow, but can make decisions with less information. Combining them together omits the disadvantages of the individual models and results in a powerful model that can be applied in various AI applications. Although current research trends claim that fully automatic AI systems will replace human labor in the near future, there are many domains where this is far from reality. One of the first obstacle for this vision is that in many domains near-perfect performance is required. For example, many biomedical applications have near 0% error tolerance, despite datasets full of uncertainty, incompleteness and noise. Furthermore, some problems in the medical domain are quite challenging, making the application of fully automated models difficult, or at least raising questions on the quality of results. Consequently, efficiently including a domain expert as an integral part of the system not only greatly enhances the knowledge discovery process pipeline, but can in certain circumstances be legally or ethically required. Furthermore, machines require vast amounts of consistently labeled data to be able to perform well. However, human annotation tasks intrinsically carry a level of disagreement among annotators, regardless of their level of domain expertise. For example, in the medical domain the inter-annotator agreement can be as low as 66%. Thus, the machines cannot learn effectively from such noisy data. While this issue cannot be solved, a human-in-the-loop model can help the subject matter experts to get to the desired solution more efficiently and effectively. " Dr. Petar Ristoski is a Research Staff Member in the Computer Science Department at the IBM Almaden Research Center. As part of his work, he conducts research in Artificial Intelligence with an emphasis on Neural Networks and Semantic Technologies. His research involves discovering fundamental principles and implementing prototype systems that can be used to understand, analyze, and manage human text as well as collaborating with human experts to train more capable Cognitive Systems.

Active Learning and Annotation
The "active learning" model is motivated by scenarios in which it is easy to amass vast quantities of unlabeled data (images and videos off the web, speech signals from microphone recordings, and so on) but costly to obtain their labels. Like supervised learning, the goal is ultimately to learn a classifier. But the labels of training points are hidden, and each of them can be revealed only at a cost. The idea is to query just a few labels that are especially informative about the decision boundary, and thereby to obtain an accurate classifier at significantly lower cost than regular supervised learning. There are two distinct ways of conceptualizing active learning, which lead to rather different querying strategies. The first treats active learning as an efficient search through a hypothesis space of candidates, while the second has to do with exploiting cluster or neighborhood structure in data. This talk will show how each view leads to active learning algorithms that can be made efficient and practical, and have provable label complexity bounds that are in some cases exponentially lower than for supervised learning.

> Active Learning: Why Smart Labeling is the Future of Data Annotation | Alectio
Today, with always more data at their fingertips, Machine Learning experts seem to have no shortage of opportunities to create always better models. Over and over again, research has proven that both the volume and quality of the training data is what differentiates good models from the highest performing ones. But with an ever-increasing volume of data, and with the constant rise of data-greedy algorithms such as Deep Neural Networks, it is becoming challenging for data scientists to get the volume of labels they need at the speed they need, regardless of their budgetary and time constraints. To address this “Big Data labeling crisis”, most data labeling companies offer solutions based on semi-automation, where a machine learning algorithm predicts labels before this labeled data is sent to an annotator so that he/she can review the results and validate their accuracy. There is a radically different approach to this problem which focuses on labeling “smarter” rather than labeling faster. Instead of labeling all of the data, it is usually possible to reach the same model accuracy by labeling just a fraction of the data, as long as the most informational rows are labeled. Active Learning allows data scientists to train their models and to build and label training sets simultaneously in order to guarantee the best results with the minimum number of labels. Jennifer Prendki is currently the VP of Machine Learning at Figure Eight, the essential human-in-the-loop AI platform for data science and machine learning teams.

PyCon.DE 2018: Building Semi-supervised Classifiers When Labeled Data Is Unavailable - H. Niemeyer
In many situations large datasets are available but unfortunately labeling is expensive and time consuming. Active Learning is a concept for building classifiers by letting the algorithm choose the training data it uses. This achieves greater accuracy than just labeling a random subset of the available dataset. The active learning algorithm selects some unlabeled data instances which are then labeled by a human annotator. Given this information a classifier is trained and new instances for the human annotator to label are selected. This iterative process tries to label as few instances as possible while achieving high classification accuracy. In this talk I will give a general overview of the core concepts and techniques of active learning like algorithms for selecting the queries and convergence criteria. Python

Jan Freyberg: Active learning in the interactive Python environment | PyData London 2019
Many data science and machine learning techniques require labelled data. In many businesses, this means that a lot of time, energy or money goes into acquiring labels. Active learning is a technique to make this process more efficient, by choosing data points to label based on current model performance. Here, I discuss methods of doing so easily and quickly in the interactive Python ecosystem.

Augmented Intelligence

YouTube search... ...Google search

Augmented Intelligence
Cosmo Tech uses Augmented Intelligence to help model and simulate complex systems, allowing decision makers to make the most optimal decisions.

Augmented Intelligence - A Marriage between Machine and Human | Simon Stiebellehner
The marriage of human and machine is commonly referred to as “augmented intelligence”. It is a logical and highly valuable intermediate step on our path to complete automation of significant parts of our lives. Augmented intelligence technologies leverage artificial intelligence to support humans’ decision processes. A concrete case of highly evolved augmented intelligence could be detecting cancer on medical images, computing confidence scores of these predictions, forwarding critical/low-confidence cases to a professional together with an explanation of what the system may have found suspicious, the professional then may return his feedback to the system for it to continue learning. The benefits of such systems are twofold. First, augmented intelligence builds trust through supporting humans without taking away their decision-making power. Trust in machine intelligence is an important prerequisite to more extensive automation. Second, it is important to recognize that both, machines and humans, have different strengths. Whilst machines excel at processing data at a high pace and at recognizing patterns they have frequently seen before, humans are able to learn well based on very few samples and are more flexible in their thinking and perception. Therefore, ideally, these strengths are combined to achieve synergies. However, making this marriage of machine and human a happy one is not trivial. Visit the largest developer playground in Europe! Facebook:

Webinar: Augmented Intelligence- Accelerating Data to Actions
Augmented Intelligence, powered by emerging frameworks and tools in data engineering, analytics and machine learning, is transforming and accelerating the data to action journey at Enterprises. This video discusses: How’s the data value chain evolving? How Augmented Intelligence is accelerating the data to the decision journey? 5 simple ways to leverage Augmented Intelligence in your organization.

AI, Human Augmentation, and the Future of Intelligence on Earth | David Brin | Talks at Google
Futurist, astrophysicist, and best-selling science fiction author David Brin takes us 30 years into the future to explore how developments at companies in fields such as AI and human augmentation will help propel humanity forward… though with some cautions along the way. Brin’s best-selling science fiction novels include The Postman (filmed in 1997), Startide Rising (Nebula Award winner), The Uplift War (Hugo Award winner), and Earth. His non-fiction work, The Transparent Society, won the American Library Association's Freedom of Speech Award for exploring 21st Century concerns about security, secrecy, accountability, and privacy. Brin holds a PhD in Physics from the University of California at San Diego, a masters in optics, and an undergraduate degree in astrophysics from Caltech. Moderated by James Freedman. Get the book here:

Augmented Intelligence: The weapon and shield of the future | David Benigson | TEDxBonnSquare
Artificial Intelligence and its impacts are consistently misunderstood as automation, but the future of AI is the augmentation of human endeavours. David Benigson was named to Forbes’ 30 Under 30 Europe list for 2017, pinpointing leaders who are changing their industries. He founded his startup Signal in 2013 with a simple mission: to use cutting-edge artificial intelligence to cut through the ever-increasing volume of information noise. Led by him, Signal has won numerous awards including, KTP Best Partnership 2015, Bloomberg Business Innovators 2016 and New Business of The Year at the 2017 National Business Awards. Signal uses artificial intelligence to monitor millions of new sources and deliver hyper-relevant news and niche stories to executives in real-time. Last year David was able to complete a nearly $7 million funding round last year for Signal led by MMC Ventures and Hearst Ventures. He speaks regularly on AI, new media, SaaS, startups and fundraising, and fake news and is highly knowledgable in the AI field. This talk was given at a TEDx event using the TED conference format but independently organized by a local community. Learn more at

Augmented Intelligence
How can machines and humans come together to achieve new feats? Subscribe for regular science videos:

Technology is becoming more and more advanced but cannot prosper on its own, the human brain and the experience that humans have is not easily taught, from removing bias to introducing emotional intelligence. Join a panel of experts as they discuss how machines, humans and processes are coming together to create powerful new insights. James Hewitt is a speaker, author & performance scientist. His areas of expertise include the ‘future of work’, human wellbeing & performance in a digitally disrupted world & methods to facilitate more sustainable high-performance for knowledge workers. Karina Vold specializes in Philosophy of Mind and Philosophy of Cognitive Science. She received her bachelor’s degree in Philosophy and Political Science from the University of Toronto and her PhD in Philosophy from McGill University. An award from the Social Sciences and Humanities Research Council of Canada helped support her doctoral research. She has been a visiting scholar at Ruhr University, a fellow at Duke University, and a lecturer at Carleton University. Martha Imprialou is a Principal Data Scientist at QuantumBlack. Watch the Q&A: This event was supported by QuantumBlack and was filmed in the Ri at 16 May 2018.