Learn how to automate your systems, how to build chat bots and the future of deep learning. Explore the applications of machine learning, NLP, and computer vision transferring Neural Network know-how from academia to architects
As machine learning finds a place in high-stakes arenas such as finance and criminal justice, it's critical to make sure the people it affects are treated fairly. Meeting this challenge is a complex and important endeavor that combines computer science, ethics, and mathematics in sometimes surprising ways. We'll talk about recent results on fairness in ML that range from theoretical to practical, including constructive steps developers can take.
Machine learning models, including deep neural networks, were shown to be vulnerable to adversarial examples—subtly (and often humanly indistinguishably) modified malicious inputs crafted to compromise the integrity of their outputs. Adversarial examples thus enable adversaries to manipulate system behaviors. Potential attacks include attempts to control the behavior of vehicles, have spam content identified as legitimate content, or have malware identified as legitimate software. // In fact, the feasibility of misclassification attacks based on adversarial examples has been shown for image, text, and malware classifiers. Furthermore, adversarial examples that affect one model often affect another model, even if the two models are very different. This effectively enables attackers to target remotely hosted victim classifiers with very little adversarial knowledge.
Scheduling meetings, booking travel, managing your receipts, and repetitive sales tasks; these are some of the many chores we must do everyday. But they’re not core to our jobs and often distract us from the high value tasks, like cultivating a lead or sharpening our analysis of our customers. Over the next half decade, as more AI intelligent agents come to market, employees will increasingly deploy a suite of agents to get their job done, and port agents from one job to the next. Much like Bring Your Own Device (BYOD), this new paradigm—Bring Your Own Agent (BYOA)—will likely change the nature of work.
Today, Artificial intelligence is not science fiction anymore. The availability of big data together with big-data analytics platforms, advanced machine learning methods, high-speed internet, and global open-source R&D communities have enabled powerful AI applications such as intelligent web search, machine translation, smart interactive assistants and business intelligence software. In my talk, I will sketch the transdisciplinary applied research at the German Research Center for Artificial Intelligence to illustrate the wide range of tomorrow’s artificial intelligence applications. I will also describe transfer channels from research laboratories to products and their comercialization that have proven effective. But it is not only the channels but also a commercialization-driven methodology that shortens the way from the researcher to the customer. Some central methods will be outlined in the presentation ranging from the integration of “design thinking” into the first steps of the research process all the way to combining research and product use. Then I will zoom into two areas of language technology and explain their applications: 1) big textual data analytics, 2) smart conversational agents.
This workshop will teach you how to use the Recast.AI platform and the basics of chatbot building. You will also learn how to connect your bot to a messaging platform using the bot connector.
Although deep learning is now widely known in the data science community, it still remains underused by big players to solve concrete business cases for three main reasons: deep learning is annotation hungry, it is hard to build data-science teams with the required expertise and because deploying deep learning models on GPUs can be tricky. We intend to change this. In this talk we will see how deep learning can easily be applied to solve image-based problems for businesses. We will demo our self-service/user friendly/accessible ? deep learning platform that allows users to easily annotate and make use of of unannotated data, train the best possible models and deploy them in the cloud or on premise(ça me parait un peu bizarre, je mettrais plutôt « on site » ou « locally », après si c’est un terme technique, oublie ça) . This demo will also be the occasion to unveil mathematical, algorithmic and technical details about how our product works under the hood, so be sure to join us for a comprehensive tour of applied deep learning.
The session will cover NLP and text mining using Python and offer several examples of real world applications. Participants will be introduced to various text processing techniques and learn more about text classification, clustering, and topic modeling. By the end of the workshop, participants will be able to use Python to explore and build their own models on text data.
It is widely accepted that where you live says a lot about who you are, demographically speaking. At the same time, many companies are desperate to find out more about their customers in order to better understand them. By knowing where they live however, many companies are sitting on an extremely rich data set from which they could learn a lot about their customers. Furthermore, this data can be used to optimise their marketing strategy and help them expand their customer base. The technique we have developed enriches a customer data set using UK census data and then applies a novel, tree-based unsupervised learning algorithm to extract differentiating demographic features and identify high-value postcodes. Our algorithm allows us to avoid performing anomaly detection on the entirety of the UK population. Furthermore, the method we have developed is not restricted to the field of marketing; it also applies to rare events. Fires or A&E admissions are relatively rare events where one would like to avoid having to perform anomaly detection on the entire UK population or all UK households. Likewise we look to future applications of our algorithm involving enriching data with data sets other than the census.
Imagine a machine that simultaneously looks at each single developer on a software team, the entire team, the code and all other data sources, and uses this knowledge to make developers smarter, teams more effective and code better. Welcome to the Deckard's world.
The recent introduced theory of practopoiesis offers an account on how adaptive intelligent systems are organized. According to that theory biological agents adapt at three levels of organization and this structure applies also to our brains. This is referred to as tri-traversal theory of the organization of mind or for short, a T3-structure. To implement a similar T3-organization in an artificially intelligent agent, it is necessary to have multiple policies, as usually used as a concept in the theory of reinforcement learning. These policies have to form a hierarchy. We define adaptive practopoietic systems in terms of hierarchy of policies and calculate whether the total variety of behavior required by real-life conditions of an adult human can be satisfactorily accounted for by a traditional approach to artificial intelligence based on T2-agents, or whether a T3-agent is needed instead. We conclude that the complexity of real life can be dealt with appropriately only by a T3-agent. This means that the current approaches to artificial intelligence, such as deep architectures of neural networks, will not suffice with fixed network architectures. Rather, they will need to be equipped with intelligent mechanisms that rapidly alter the architectures of those networks.
Word embeddings have received a lot of attention since some Googlers published word2vec in 2013 and showed that the embeddings that the neural network learned by "reading" a large corpus of text preserved semantic relations between words. As a result, this type of embedding started being studied in more detail and applied to more serious NLP and IR tasks such as summarization, query expansion, etc... In this talk we will cover the implementation and mathematical details underlying tools like word2vec and some of the applications word embeddings have found in various areas. An overview of the emerging field of "<anything>2vec" (phrase2vec, doc2vec, dna2vec, node2vec, etc...) methods that use variations of the word2vec neural network architecture will also be presented.
Chatbots and conversational interfaces are a topic of significant interest, but the space is oddly divided in current thinking, ranging along a gradient from "chatbots are an essential and complex AI component of new tech" to "chatbots are quickly becoming a shallow tech commodity". The truth is, as usual, somewhere in between: it is easy to underestimate the effort and resources needed to create a conversational interface with real value, but the field is exploding with new tools that will help. In this talk I will present practical learnings and advice from my last two years of work, first at babylon health developing one of the leading medical chatbots acting as an interface for what is becoming a true AI doctor, and lately as the CTO of an insurtech startup aiming to improve the insurance experience with AI technology. I will cover when and why conversational interfaces are an appropriate choice, the pitfalls many attempts face when creating their own, thoughts on the best approaches and views on upcoming trends.
Artificial intelligence is more than a game of man vs. machines. For AI to have a valuable impact, it requires a symbiosis with humans. The presentation will explore real-world examples of how core AI ingredients such as knowledge representation and reasoning, learning, and decision-making requires human engagement. Attendees will learn how this synergy can lead business leaders to:1. Utilize the right data (structured and unstructured) from across silos and organizations; 2. Accelerate knowledge discovery within the enterprise; 3.Reduce capital expenditures; 4.Optimize assets and processes through operationalization; 5. Increase profitability The session will conclude with a look at a new category of technology being driven in part by AI algorithms, machine learning, semantic search, and an enterprise knowledge graph, called Knowledge-centric technology.
In a parallel to quantum physics this talk introduces social quantum physics, defining four key principles of social quantum physics that help build collective consciousness of swarms: empathy leading to entanglement, and reflection leading to reboot and refocus. The collective mind is measured through a collaboration scorecard made up of six key variables – “honest signals” – drawn from communication on Twitter and the Web, from e-mail inside large companies and in small teams from smartwatches and sociometric badges. The “six honest signals of collaboration” are strong leadership, balanced contribution, rotating leadership, responsiveness, honest sentiment, and shared context. I will illustrate these “honest signals of collaboration” using numerous examples ranging from biotech startups to innovation teams at the R&D departments of Fortune 500 firms to teams of Healthcare researchers and patients. Read more in the two new books by Peter Gloor: “Sociometrics and Human Relationships: Analyzing Social Networks to Manage Brands, Predict Trends, and Improve Organizational Performance” and “Swarm Leadership and the Collective Mind: Using Collaborative Innovation Networks to Build a Better Business” which will both come out with Emerald Publishers in April 2017.
Deep Learning for Image analysis is now widely spread among academia as well as business use cases. In most cases, the amount of quality labeled data needed as well as the definition itself of the labels is problematic. On the other hand, image data associated with raw text is omnipresent on internet. Using this weak supervision, we will show how we can leverage huge amounts of data for image understanding, and show the pertinence of the method on visual fashion analysis. This work is made and presented by both Charles Ollion as well as Hedi Ben Younes, PhD in Machine Learning at LIP6/Heuritech.
Many open source deep learning frameworks are competing for the position that eases programming the most. This helps researchers with faster iterations. But for the industry, the next question is -- how to run the programs. AI depends on big data, which come from Web servers and in the form of log messages, or from crawlers and in the form of external datasets. In the industry, we need a complete solution that covers the collecting of data, learning from data, and feedback the models to the business. This talk explains lessons we learned from PaddlePaddle, a recently open sourced deep learning platform which has been widely used in Baidu for four years.
To what extent do chatbots use AI? How can it be efficiently used? What’s the right data for chatbots? What kind of machine learning is best suited?
Strong AI is a common goal of many computer scientists. So far, machine learning techniques have created amazing results in narrow fields, but haven’t produced something we could all call “intelligent”. Given recent advances in neuroscience research, we know a lot more about how neurons work together now than we did when ANNs were created. We believe systems with a more realistic neuronal model will be more likely to produce Strong AI. Hierarchical Temporal Memory is a theory of intelligence based upon neuroscience research. The neocortex is the seat of intelligence in the brain, and it is structurally homogeneous throughout. This means a common algorithm is processing all your sensory input, no matter which sense. We believe we have discovered some of the foundational algorithms of the neocortex, and we’ve implemented them in software. I’ll show you how they work with detailed dynamic visualizations of Sparse Distributed Representations, Spatial Pooling, and Temporal Memory.
Join our session on the first application of deep learning to cybersecurity. Dr Eli David, one of the leading global experts on deep learning, co-founder and CTO of Deep Instinct, will cover the evolution of artificial intelligence, from old rule-based systems to conventional machine learning models until current state-of-the-art deep learning models. Deep learning is a novel branch of artificial intelligence inspired by the brain’s ability to learn: once a brain learns to identify an object, its identification becomes second nature. Similarly, as a deep learning-based artificial brain learns to detect any type of cyber threat, its prediction capabilities become instinctive. As a result, the most evasive and unknown cyber-attacks are immediately detected and prevented. In this webinar, Dr Eli David will present the ground-breaking results exhibited by deep learning when applied to computer vision, speech, text understanding and for the first time to cybersecurity.
Supervised deep learning networks require significant computational resources to train. In order to reduce the total time to train, it is advantageous to distribute the workloads across several compute nodes. In this lecture we will discuss the algorithmic challenges of distributed training and methods to alleviate some of these challenges.
As machine learning algorithms become more widely used, it is important to ensure that they provide the privacy and security guarantees. In this talk, I outline some of the kinds of attacks that adversaries can make against machine learning models, and some of the defenses that we can use in response, like adversarial training and differential privacy. This talk is a high-level overview of this area to whet your appetite; AI With the Best also features detailed talks by Nicolas Papernot, Patrick McDanel and Dawn Song zooming into detail on some of these subjects.
Machine learning algorithms are categorized into supervised, unsupervised and semi-supervised. This presentation will discuss how to analyze a given dataset and applying an appropriate model. The steps are: getting and cleaning data, extracting and selecting features and finally developing an appropriate classifier. Some of the popular classifiers such as Naïve Bayes, SVM and Neural Network will be discussed. Generalizing the algorithm on test dataset and calculating error rate is an important part in developing a robust model on any given dataset. The algorithms will be discussed briefly with some practical examples.
I will review some of the opportunities, applications, and challenges of using AI and machine learning for societal good. I will also summarize briefly this year's $1Million Data Science Bowl competition, hosted by Kaggle and sponsored by Booz Allen Hamilton.
In this session, you’ll learn how you can buid a custom image classifier powered by your own CNN architectures in Keras, and also learn to use the trained models and run predictions against them.
Search is an important problem for modern e-commerce platforms such as Etsy. As a result, the task of ranking search results automatically or the so-called learning to rank is a multibillion dollar machine learning problem.In this talk, we first review Etsy's approach to learning to rank using a few hand-constructed features based on the Etsy listing's text-based representation. We then discuss a multimodal learning to rank model that combines these traditional text-based features with visual semantic features transferred from a deep convolutional neural network. We show that a multimodal approach to learning to rank can improve the quality of ranking in an experimental setting. Reference: http://www.kdd.org/kdd2016/subtopic/view/images-dont-lie-transferring-deep-visual-semantic-features-to-large-scale-m
When you think of virtual assistants, what comes to mind? Likely your experiences with Alexa, Google Home or Siri. However, chatbots in the enterprise have often failed to provide a similar quality experience. Yet chat remains the number one way customers want to talk to brands, and they often want to talk when customer service teams are offline. What are the criteria needed to deploy an enterprise-class chatbot, one that yields effective business results? And what is the role that natural language processing technology and machine learning play in the technology powering chatbots? Why is NLP sometimes superior to ML and vice versa? We'll also explore how chatbots platforms like Inbenta are leveraging both technologies to deliver the best conversational experience.
Deep Neural Networks can be viewed as a mechanism for modeling information. In this talk we will share an intuitive view of deep neural network and embedded spaces in terms of the information they hold. The talk will not involve complex mathematical explanations, rather, only intuitions that should help simplify and clarify the process of solving real life problems via neural networks.
Professional opportunities can manifest itself in several ways like finding a new job, enhancing or learning a new skill through an online course, connecting with someone who can help with new professional opportunities in the future, finding insights about a lead to close a deal, sourcing the best candidate for a job opening, consuming the best professional news to stay informed, and many others. LinkedIn is the largest online professional social network that connects talent with opportunity at scale by leveraging and developing novel AI methods. In this talk, I will provide an overview of how AI is used across LinkedIn and the challenges thereof. The talk would mostly emphasize the principles required to bridge the gap between theory and practice of AI, with copious illustrations from the real world.
Computing activation gradients in image space is a basic tool for visualizing individual neuron function in neural nets. First popularized by Erhan et al. (2009), the method without any tweaks usually produces noisy, unrecognizable results. However, with a few tricks, this family of approaches can be made to produce crisp results useful not only for visualizing neural function, but for creating a flexible class of generative models.
Machine learning and data science have taken Silicon Valley by storm with virtually every company creating positions in the field. Currently, it appears to be a general consensus that machine learning as it is being employed in industry is not living up to its promises. I want to take a deeper look into the state of data science in industry. This talk will address some of the problems and challenges that data science has, how it can help industries when it is working properly, and how to help get data science from where it is today to where it has the potential to be more quickly.
Most developers are aware that some algorithms can be run on a GPU, instead of a CPU, and see orders of magnitude speedups. However, many people assume that:
1. Only specialist areas like deep learning are suitable for GPU
2. Learning to program a GPU takes years of developing specialist knowledge.
It turns out that neither assumption is true! Nearly any non-recursive algorithm that operates on datasets of 1000+ items can be accelerated by a GPU. And recent libraries like Pytorch make it nearly as simple to write a GPU accelerated algorithm as a regular CPU algorithm. In this talk we'll write an algorithm first in python (with numpy), and will then show how to port it to Pytorch, and will show how to get a 20x performance improvement in the process. Familiarity with Python and numpy is assumed. No previous pytorch experience is necessary.
Deep Learning and other artificial neural networks in combination with big data have brought a renaissance to AI and led to a torrent of new applications. Continued advances promise to produce autonomous systems that will perceive, learn, decide, and maybe even act on their own. However, the effectiveness of AI systems is still limited in many ways. For example, one of the biggest challenges is that increasingly intelligent machines need to be accountable for their decisions and, ideally, even explain their reasoning to human users. This talk will present an overview of approaches to address these and other current AI problems.
We will look at conversational design tips and tricks to deliver effective learning and engagement through messaging services such as WhatsApp.
Humans are consumers of visual content. Everyday, people watch videos, play digital games and share photos on social media, but there is still an asymmetry, in which not that many of us are creators. In this talk, we aim to build machines capable of visual creativity, and use the “creative machines” as training wheels for visual content creation, with the goal of making people more visually literate. We will present three projects based on Generative Adversarial Networks (GANs). First, we propose to directly model the natural image manifold via GANs, and constrain the output of an image editing tool to lie on this manifold.Then, we present a general image-to-image translation framework, “pix2pix”, where a network is trained to map user inputs directly to the final results. Finally, we present a new algorithm that can learn image-to-image translation even when paired training data is not available. See more details at https://github.com/junyanz.
Explanations have been shown to increase the user’s trust in the recommender system in addition to providing other benefits such as scrutability, meaning the ability to verify the validity of recommendations. Most explanation methods are designed for classical neighborhood-based Collaborative Filtering (CF) or rule-based methods. However, most accurate recommender systems are black-box models, that have difficulty explaining the reasoning behind their recommendations. Therefore, there is a gap between accuracy and transparency or explainability of the models. This talks gives an overview of the main streams of research in the field of explainable models in recommender systems.
The Artificial Intelligence (AI) hype machine has been running at full throttle for the last few years. Machine learning, one of the most rigorously researched AI subfields, has had a few high-profile successes in that period, and that is all it has taken to drive the imaginations of many observers into science fiction. In fact, many of those success stories were built on decades-old techniques that have only now become feasible thanks to the availability of large-scale computation at low costs. The sources of this frenzied perception of machine learning are varied and many; from journalists seeking sensationalist angles, to professors that now make football player salaries, to venture capitalists pouring unprecedented amounts of money into untested, unproven, and often bizarre AI approaches. However, headlines do not make science fiction into fact, neither human nor robotic neurons are amplified by dollars, and general intelligence will not be created in the lifespan of a venture fund. The human involvement needed to develop any basic application that shows a minimum level of intelligence is still huge. When we see a new “AI” beating the best humans at Go or Poker, we rarely get a detailed account of the arduous tasks and enormous amount of grunt work behind the scenes that make these applications really work. Usually, these efforts involve dozens or hundreds of hours collecting and preparing data, meticulous tweaking and fine-tuning of algorithms (that took academia years to invent and perfect), and finally preparing and deploying the infrastructure necessary to transform the data seamlessly into a computer program that can take comprehensible actions. The resulting system often still requires specialized hardware, is useless without significant human interaction, and rarely generalizes beyond the very specific problem it was designed to solve. While the increased attention and investment will help accelerate some research, it will certainly help rediscover that we are further than we believe from producing truly intelligent, general purpose applications at massive scale or with some more general intelligence on them anytime soon. In this talk, I will try to provide a grounded view of what it takes to build an end-to-end machine learning-based application, as well as some evidence on how far AI is from threatening our world.
We will demonstrate how VoiceBase use speech analytics to deliver critical insight into business calls, recordings, and videos. We will show the three layers of insight: speech-to-text transcripts, semantic keywords and topics, and business outcome predictions. We will outline the analytical infrastructure that makes the insight possible at a scale of 10,000 compute cores and growing. Finally, we will dive into key AI lessons of the deep learning, NLP, and ML that powers VoiceBase.
Language modeling is crucial to many NLP tasks. Applications include machine translation and speech recognition. Traditional n-gram and feed-forward neural network language models fail to capture long-range word dependencies in a block of text. Previous work by Mikolov et al. has shown that adding context to a Recurrent Neural Network (RNN) language model solves this dependency problem and yields lower perplexity scores. I will briefly review traditional language models before diving into the more recent contextual RNN-based language models. In particular, I will discuss the TopicRNN model, a RNN-based language model that captures long-range semantic dependencies using latent topics. I will also highlight some results on word prediction and sentiment analysis using the TopicRNN model. This is joint work with Chong Wang, Jianfeng Gao, and John Paisley.
The talk presents an overview of statistical language modeling as applied to real-word problems: speech recognition, machine translation, spelling correction, soft keyboards to name a few prominent ones. We summarize the most successful estimation techniques, and examine how they fare for applications with abundant data, e.g. voice search. We conclude by highlighting a few open problems: getting an accurate estimate for the entropy of text produced by a very specific source, e.g. query stream); optimally leveraging data that is of different degrees of relevance to a given "domain"; does a bound on the size of a "good" model for a given source exist?
An important component of conversational AI is efficient natural language semantic processing and search. Latent Semantic Indexing (LSI) enables the extraction of semantic features but doesn't quite live up to its "Indexing" name. The dense, continuous, high-dimensional topic vectors required to characterize the meaning of natural language documents with LSI are not indexable or searchable with anything other than an "index scan." At Talentpair we used to search for "pairings" within a large database of 200-dimensional topic vectors (for resumes and job descriptions) with a brute force O(N^2) computation (O(N) for a single query). Incremental isometric feature mapping (Isomap) and Incremental Locally Linear Embedding (ILLE) offered the promise of reducing our feature space sufficiently (from 200-D to 3-D or less) to enable indexing using mature GIS database technology. However, mature implementations are not available for our backend technology (Python, PostgreSQL, Elasticsearch). In addition, the run-once T-Distributed Stochastic Neighbor Embedding algorithm offers a higher fidelity embedding that preserves more of the structure relevant to our problem--but TSNE does not allow online (incremental) embedding. We recently discovered a straightforward way to approximate some TSNE mappings with a multivariate polynomial regression using an off-the-shelf open source machine learning package (Scikit-Learn). This is a game changer for our semantic search problem, enabling us to perform semantic queries on a large database of documents (finding the best pairs for a job or candidate) in constant time. It may also be effective for more general natural language semantic search problems, such as those in conversational AI.
With hundreds of thousands of cars available on Turo, how can relevance be defined? What are the best cars to show? This talk will cover the full lifecycle of a Machine Learning model from data collection to its deployment in production.
I'll discuss strategies that let models learn the optimal hyper-parameters for themselves and potential applications
Advances in declarative knowledge modeling can represent the conceptual work of socio-technical systems for rigorous analysis and design. These systems can have complex combinations of multiple users, a variety of computing devices, with information that is used and changed as it flows between activity in the physical world and processing in the digital world. This complexity can overwhelm conventional design methods, which has led to some serious, negative impacts. Clarity and rigor of their design, however, can now be achieved by explicitly representing the products of conceptual work with declarative models, which also enable powerful model checking for design verification. The new techniques will be illustrated with examples chosen from clinical health care, aerospace, and online tech support.
Over $60 billion is spent annually fighting fraud and abuse. Fraudsters are highly motivated, skilled and organized. As the value and volume of online transactions increase, the need to address fraud in a robust, scalable manner becomes indispensable. This talk covers how Sift Science applies large-scale machine learning to protect thousands of web sites from fraud and abuse. By combining models specific to an online business with those learned by mixing data from across its network of customers, Sift Science uses a combination of various ML approaches to prevent fraud.
One of the biggest challenges for research in artificial intelligence is unsupervised learning. Current industrial success with deep learning relies heavily on supervised learning, where humans are needed to categorize data and define high-level abstractions which we want the computer to know about. However, humans are able to discover many aspects of the world without a teacher telling them anything about it, and this ability to autonomously learn to make sense of the world is something that needs to be further developed for computers. The deep learning approach to unsupervised learning centers on the question of learning representations, and different algorithms define an objective function which leads the learner to capture essential aspects of the data distribution along with a new space in which to represent data. Deep generative models can demonstrate their understanding of the data by generating novel examples which nonetheless look like those which were used to train the model. Many of these models are related to the old idea of auto-encoder, with an encoder function mapping data to representation and a decoder (or generator) mapping the abstract representation to the raw data space. The talk will focus in particular on the family of generative adversarial networks (GANs), which question the established approaches based on maximum likelihood and probability function estimation and bring us into the realm of game theory and novel ways of comparing different distributions against each other, as well as with very impressive generation of images.
The next generation of robots will soon get out of the secure and predictable environment of factories and will face the complexity and unpredictability of our daily environments. To avoid that robots fail lamely at the task they are programmed to do, robots will need to adapt on the go. I will present techniques from machine learning to allow robots to learn strategies to enable them to react rapidly and efficiently to changes in the environment. Learning the set of feasible solutions will be preferred over learning optimal controllers. I will review methods we have developed to allow instantaneous reactions to perturbation, leveraging on the multiplicity of feasible solutions. I will present applications of these methods for compliant control during human-robot collaborative tasks and for performing fast motion in sport, such as when playing golf with moving targets. The talk will conclude with examples in which robots achieve super-human capabilities for catching fast moving objects with a dexterity that exceeds that displayed by human beings.
In this talk, I will discuss inventions and patents that are derived from AI. This is not a talk on how machines can invent or write patents. Instead, it will focus on how ideas derived from AI can ultimately lead to new inventions and patents. I will give examples of this creative process. Specifically, I will show how we can think of the problems that surround us in life not just as "nuisances." Instead, problems can be seen as challenges to be resolved with machine learning methods. When the resulting solutions are sufficiently novel, they can lead to new patents.
In order for an AI to be able to assist and help its users, it needs to know them and their world first. For this reason, one of the crucial parts of an AI assistant is acquisition of the knowledge about the world around and the users. In this talk, Luka Bradesko will present a context aware knowledge acquisition system that simultaneously satisfies users’ immediate information needs while extending its own knowledge using crowd-sourcing. The focus is on knowledge acquisition on a mobile device, which makes the approach practical (also partially available as a library to other developers). The viability of the approach was tested experimentally with real users around the world, and an existing large source of common sense background knowledge (Cyc). The experiments show that the approach is promising if it could be successfully brought from a research prototype into a full scale product. The talk will also include some thoughts on why the previous attempt of productivisation attempt was not successful.
After life originated on Earth, the next important transition was the emergence of cognitive life, in which simple organisms self-organized into dynamical networks to compress and express complex information in the environment about their own preservation. Cognition is better understood as the information flow between single agents, implementing a dynamical way to compress relevant information for their own survival, and enabling them to make predictions about their environment on much shorter timescales than Darwinian evolution. In this talk, Dr. Witkowski will present the contribution of artificial life tools, information theory and connectionist machine learning, to our understanding of the transition to cognitive life. Just as life can be formulated computationally as the search for sources of free energy in an environment to maintain its own persistence, cognition is better understood as finding efficient encodings and algorithms to make this search probable to succeed. Cognition then becomes the “abstract computation of life”, with the purpose to make the unlikely likely for the sake of survival.
How is Artificial Intelligence (AI) used in today’s digitization of health, and how will it shape our future of health? AI is becoming increasingly important in digitizing many areas of health care, medicine, and life sciences. Indeed, AI is already key in approaches of digital health, such as digital medicine, digital diagnostics and digital therapeutics. A main driver for this development is that AI can efficiently personalize health services for many people. Key stakeholders in health care, e.g. health insurers, hospitals and doctors, believe that AI is a scalable approach to achieve better health outcomes at lower costs. In short, “automated” personalization improves value in public health. For example, a person can improve long term health when he/she knows the practical meaning of his/her genetic and behavioral background in daily life situations. This presentation focuses on how AI (and related concepts, such as Machine Learning and Data Science) addresses today’s challenges of health, and considers theoretical and technical requirements and limitations. Recent technologies are reviewed in more depth, such as advanced and predictive analytics, and social and mobile health.
Any autonomous agent/system would have to face unseen problems during its lifetime and be able to solve them on its own in order to sustain. Problem solving is an area of artificial intelligence that studies the frameworks and methods related to accomplishing non-trivial tasks, with the given capabilities of an agent. In this talk, we will introduce you to state-space search approach which is a well known problem solving technique in AI and is state-of-the-art for arriving at solutions in various game environments, optimization settings, and robotics. We discuss the combinatorial explosion of states involved, how to handle it, and some applications that guide you on to applying the search techniques in new contexts on your own.
Data is the crucial part of any AI and machine learning applications. the presentation discusses the data in industrial world and how it produces challenges to machine learning algorithms.
The industrial world is changing; - From AI theory to industry changing products; - Engineering AI with KONUX to unlock a new level of asset performance in the rail industry
Big Data and the Internet of Things (IoT) have the potential to fundamentally shift the way we interact with our surroundings. The challenge of deriving insights from the Internet of Things (IoT) has been recognized as one of the most exciting and key opportunities for both academia and industry. Advanced analysis of big data streams from sensors and devices is bound to become a key area of data mining research as the number of applications requiring such processing increases. Dealing with the evolution over time of such data streams, i.e., with concepts that drift or change completely, is one of the core issues in stream mining. In this talk, I will present an overview of data stream mining, and I will introduce some popular open source tools for data stream mining.
From helping doctors prescribe medications to spotting abnormalities in medical images, there are many ways that machine learning can assist in building better precision medicine. However, many people who know machine learning still struggle when working with medical data. This talk will mention some ways to get the most out of your medical data, with a focus on dealing with missing data.
Development of Autonomous driving capabilities through machine and deep learning requires training upon huge annotated data. Obtaining such training data requires a lot of efforts, not to mention the large time required to do so. This talk will explore the possibility of accelerating autonomous driving research by training machine and deep learning models upon objects in a rich virtual world. The talk will briefly comment on how models, trained on simulated data, perform when tested with the real world driving data.
Massively Multi-Player On-Line Role Playing Games provide us with an excellent opportunity to study and implement multi-agent systems. At one hand various in game characters (including mobs, monsters, non-playing characters - NPCs etc.) can be modelled and implemented as (intelligent) agents thus facilitating interesting game-play. On the other hand, the behaviour of players can be analyzed using agent based models based on big data analytics. These models can then be put to use to implement bots (artificial players) for automated testing of such games. Herein results from the ModelMMORPG project will be presented that investigated both aspects mentioned above.
AI has largely been mystified and is commonly viewed as self-thinking “strong” IA (the Ghost in the Shell Fantasy), when operational AI is and remains probability tools and/or classifiers, no matter how complex or bio-inspired they are. The three waves of AI each carry specific assets and liabilities, but endlessly turn out to be “weak”, instructional, predetermined and purpose-oriented. Heuristics, fuzzy logics, expert systems, NN, RNN and their newly- patented architectures are as old in the history of Computer Science as relevant in today’s entrepreneurial world. I shall rather argue for a prosthetic conception of AI – if human cognition is embodied, distributed and non-computational, then AI has to be a tool for human augmentation (not imitation or full substitution), the way an instrument is for a musician. As such, it is an entire constituent of his/her understanding of daily patterns and routines – one may not see the world if he/she glares at his/her glasses, but only perceives it once looking through them. The metaphor extends to VR where feeling immersed rests on the (prosthetic) tools you appeal to for adequate perception. Clay VR is a SDK for gesture recognition on smartphones from any embedded lens (no need for additional hardware). It is designed for uses in VR and distal control to drastically enrich the user’s interaction possibilities – it displays one’s own hands contoured in a virtual environment. Such a touchless experience is control without the hardware, touch, pinpoint or remote-control pains, ensures immersion by preserving self-perception, but implies massive technical difficulties. Because it is constrained by the device limitations, the approach has to be minimal and data, often, has to be rebuilt or inferred. In that, it centrally resorts to Computer Vision and AI. Computer Vision for real-time image interpretation, and AI for learning (from the ever-changing capture environments) and automation (scoring, gesture validation, depth processing, etc.). The overall architecture is, indeed, a multi-image/feature- fed RNN that, in turn, nourishes a heuristics-jugulated expert system, but most of the intelligence provided only serves poise, both technically and for the user’s perceptual loops that the AI-boosted external system provides. It (AI) therefore becomes a constituent of human distributed cognition processes, augments the user’s capacities, but never replaces intellection itself.
Recommendation systems that help users navigate through information by delivering the right content at the right time, are a part of our every day lives. Although a lot of progress has happened regarding the development of recommendation systems for unconstrained offline settings, there are still challenges when deploying such systems in constrained interactive settings. This talk will begin by reviewing the state-of-the-art in offline unconstrained recommendation. Then it will discuss methods for the particular constrained settings of (i) limited screen size of the devices, and (ii) limited capacities of the candidate items for recommendation. The talk will continue with a benchmarking study comparing the proposed methods with the state-of-the-art.
We need to amplify the efficiency of the human experts using Internet of Things with Machine-to-Machine and Machine-to-Human Networks to create intelligent context-aware systems for solving the following three grand challenge problems: N=1 Personalization What if we present information and enable actions relevant to the context (location, role, social circumstance, access level) of the users? What if we can detect user intent and direct the user along personally and commercially valuable paths of action? Can we provide adaptive security to learn from daily behaviors, and detect the unusual through novel signals? Can we minimize loss and optimize usage by predicting when certain patterns are needed? Zero Down Time, Zero Intrusion, Zero Loss Can we provide highly reliable temporally relevant information in people's work context on their mobile devices today? What if we can prevent equipment failure and down-time by predicting maintenance or replacement needs? Can we enable predictive maintenance in mines, factories and optimize equipment usage through stream data analytics and prediction? What if we can give pre-summarized predictions and recommendation with seamless data provenance to technicians? What if we can predict risk for a fire by connecting the condition of a boiler, simulating it with the ambient? Zero Waste, Zero Delay Can sensors in the supply chain prevent food waste, Rx/Dx misuse, detect counterfeiting? Can we predict manufacturing requirements and changes from orders placed and cancelled online? Can we propagate the requirements up the supply chain all the way to suppliers? We will describe how to build smart digital workspaces that know [the context] all the while observing, recording this context of work in episodic memory and generalizations in semantic memory. What am I doing?[Activity Structure, Context, Goals] How am I doing it? [Best Methods] What resources am I using? [Allocation, Discovery] When and where am I doing it? [Time and Place] Who am I, what is my role? [Responsibility, Profile] Who are my collaborators? [Social Network] What is the device doing? Current Action in Business Process, Goals Is it Down or Active? How efficient or inefficient is it? What is the condition of the device? State (Past, Present, Future) Temperature, Pressure, Vibration, Dust, Humidity, Leaks, Fatigue/Stress How does it compare to normal operating ranges? What resources does the device depend on? Device and Human Dependencies Where is it located? When is it needed? Place, Motion What is the function of the device in the business process? Is the current activity expected according to the business process? What is the full downstream cost impact of the device going down (criticality)? What devices are its neighbors? Edge Intelligence The smart workspace will: Let devices solve routine problems automatically depending on risk using edge intelligence Anytime, Anywhere, On the Edge Selected problems that can be automated with very high confidence Let us seamlessly know device states and predictions by presenting information about operational inefficiency, risk of failure, cost of replacement, opportunity costs of switching devices, in-context of our roles To quickly find directly related information and answers to questions based on what we mean, in the context we need it, with access to the source, quality and how the information was derived, connecting us to insights of experts within the organization and beyond Context Semantics Driven Guidance Proactively show us steps others have taken in meaningfully similar situations before, helping us reason and decide faster, with greater confidence. Social Collaboration and Decision Support
The recent surge of interest in brain-based learning algorithms has improved artificial "sensory" capabilities including image recognition and speech processing, and "motor" activities as with robotics and navigation. A focused handle for addressing the "cognitive" capabilities that might intermediate between the two is recommendations, which deals with how to select between numerous available options based on underlying data, with consideration to desired effects. Brain-based learning algorithms are now beginning to meaningfully impact results here, and in turn benefit from original research in this area. We will review major approaches that are improving recommendations, including deep learning, Bayesian methods, reinforcement learning, and how to evaluate and ensemble them, in data-rich and "cold start" conditions alike. This is an exciting time for developing recommendations that really work, and for exploring cognitive technologies in general, when cleaner and more ubiquitous data can begin to inform more meaningful decisions.
In this talk, I will talk about various paths for the arrival of strong AI, and the potential impacts on society.
The field of Creative A.I. is blossoming as many companies, hobbyists and research labs turn their attention to promising new applications in creativity. From novelty search that helps designers explore options, to generative models that empowers artists to produce more artifacts at higher quality, there's a lot going on! In this talk, you'll hear how both classical artificial intelligence and machine learning can benefit creative applications. In practice, deep learning faces many problems when applied in this generative space and what you'll see can be done about it. You'll discover concept of generative pipelines and how they have been used over the years, and why they are now more relevant than ever as we enter the Generative Age.
Networks are a useful data structure for encoding relational information between a set of entities and appear in a variety of fields, from biology to social science. The use of principled statistical, computational and mathematical tools is crucial for the understanding of the structural and functional relationships encoded in the network. In this talk we will summarize 3 important areas of network science, including, 1) link prediction, 2) anomaly detection, and 3) community detection. We will discuss the practical concerns for the implementation of the state-of-the-art tools in each of the 3 areas. Finally, we will discuss the computational challenges in handling large networks.
Everyday, billions of images and videos are uploaded to social media sites, a number that is growing exponentially. It is challenging for brands to reach customers with their content, and hence they are seeking a “viral” message that resonates with their audience, is shared widely and rapidly, and provides audience engagement. Tools for automating assessment of image “virality” based on both content and context hold significant value for marketers. Based on analysis of tens of million of images from social media, we show how deep neural networks can be used to predict image virality. Our analysis shows that image content such as human presence, their emotions, pets as well as objects such as cars, impact the potential virality of an image, as do more abstract concepts such as color, background, theme and composition. For example, images of puppies and babies are more likely to be popular. Our findings further indicate that these attributes apply in a certain context, such as current political, sports and entertainment events. We also find that the social context of the image i.e. the original poster, their network, and their engagement level also impact the potential virality. We used a combination of Deep neural networks and probabilistic models to analyze tens of millions of images from multiple social media sources to identify contextual and content variables that are correlated with a higher image engagement, and to predict the normalized views of images. Using exemplar images, we present a deep-dive into our approach and key findings.
We are evolving, so as our society, lifestyle and the needs. AI has been with us for decades, and now penetrating more in our day-to- day life so as the robots. But, where are all these converging together? Towards creating a smarter eco-system of living, where robots will coexist with us in harmony, for a smarter, healthier, safer and happier life. How? Social Intelligence (SI) of such consumer Robots will be the key technology and the next big R&D challenge. SI will enable such robots to behave in socially expected and accepted manners. The talk will reinforce that robots have a range of potential societal applications, and that as a robotics industry, SoftBank Robotics’ R&D and Innovation is around the centrality of wellbeing of people. The time has arrived, when social robots have started to be deployed, evaluated and available for practical purposes outside automation industry. For example, Pepper robot from SoftBank Robotics, which is mass produced and already being used in thousands of homes, and at public places; the Romeo humanoid robot companion for everyday life of people needing assistance; the Nao robot as teaching assistant. The first part of the talk will illustrate some of the use cases, market analysis and potential applications for such intelligent humanoid robots, grounded with some key European Union Projects. The second part will present the feedback and needs from the real users. This will help to highlight some of the immediate R&D challenges from industrial perspective in the third part of the talk. Hence the young graduates will know the must/should have skills to be the part of this next generation of robotics revolution: the socially intelligent robots. The talk will conclude with some open and grand challenges ahead, including social and ethical issues.
Matrix factorization (MF) collaborative filtering is an effective and widely used method in recommendation systems. However, the problem of finding an optimal trade-off between exploration and exploitation (otherwise known asthe bandit problem), a crucial problem in collaborative filtering from cold-start, has not been previously addressed. I will present a novel algorithm for online MF recommendation that automatically combines finding the most relevant items with exploring new or less-recommended items.
While AI/Machine Learning systems are becoming more “intelligent” and ever-present in today’s products, our approach to interacting with these systems is in its infancy and we have a lot to learn. The value of such systems is directly correlated to the end-user product utility, and the underlying dynamics. This situation beckons new forms of human computer interaction capable of addressing the following questions: How will the end-user interact with the system? How will the system interact with other applications? How should these influence the overall approach to the architecting the solution? Can we formulate the underlying problem and then design the relevant statistical and engineering frameworks to support, and scale, such a solution in the first place? We will explore the evolution of human-computer interaction, it’s influence on product design and architecture, as well as ramifications in the current AI-driven environment.
In this talk I will discuss some of the insights behind the learning algorithms. The selection criterion of machine learning algorithms depending on the nature of problem to be solved. Methods to tune these algorithms and most importantly the methods to infer the errors to direct the algorithm design in the right direction. With this practical guidance, we can customize and tune our learning algorithms to provide more informed results.
In nearly all fields of science and engineering, the amount of data available is growing at unprecedented rates. Applications no longer produce data sizes from megabytes to gigabytes, but rather from terabytes to petabytes (and beyond). Machine learning is one the key tools we use to make sense of these ever-growing quantities of data. We now use machine learning methods every day; they are behind software for e-mail spam filtering, product and advertisement recommendation systems, Microsoft's Kinect, Google Translate, speech recognition on phones, and now self-driving cars. The successes and potential of machine learning are driving the need to develop techniques that can consider even larger datasets and more complicated models. A major challenge is that the "learning" in most machine learning models involves solving a numerical optimization problem, and standard numerical optimization codes are simply not up for the task of fitting very-complicated models to huge data sets. The default way to address this challenge is to use "stochastic gradient" methods. Instead of repeatedly going through your entire dataset between each model update, these methods alternate between looking at small *random* parts of the data and updating the model. These methods have been enormously successful, but they are enormously frustrating to use: it can be very hard to tune their parameters, to decide when to stop running them, and even if you address these issues they still converge very slowly. In this talk I'll give an overview of these methods, and then discuss a revolution that is happening in numerical optimization and machine learning with the development of a new class of stochastic gradient methods in 2012. Not only do the new methods make tuning parameters and deciding when to stop much easier, but these algorithms are dramatically faster than the old algorithms both in theory and practice.
Companies ranging from Google to Facebook to Amazon are on the hunt for scientists to build language intelligence. Learn the essential skills to work in one of the most exciting fields of Artificial Intelligence! This is a sequel of Aerin's first talk "Phrase2Vec in practice"
Cars increasingly are equipped with sensors that can sense their surroundings and insides. These sensors are about to be connected to the cloud in an online fashion, promising to deliver a new era of connectivity and information flow between cars. Cars of the future increasingly would enable autonomous driving, and so need to know and understand more than before. Common sense, such as understanding the meaning of a ball stuck under a parked car, or the possible intentions of a kid hiding behind that parked car, is traditionally easy for people and hard for computers and artificial intelligence. This kind of understanding and commonsense is necessary to ensure safe driving and safe co-existence of humans with those AI cars. In this presentation, I will describe the tools already developed in AI and those that still need to be developed to reach this goal of common sense for cars.
In this work, we apply machine learning to the problem of optimizing data transfer over mobile networks. We developed an adaptive learning framework to optimize TCP parameters of congestion control and concurrent connections. This generates custom TCP strategies for various geographies, origin servers and wireless networks, taking into account change in network quality over time.
This talk will cover algorithmic design principles for intelligent systems exhibiting anticipatory, flexible, autonomous, and sustainable behavior. In particular, you’ll be exposed to anticipatory multi-objective machine learning strategies for automating the resolution of conflicts in sequential decision-making under multiple, noisy, and cost-adjusted optimization criteria. The goal of anticipatory machine learning is to improve decision processes by taking advantage of predictive modeling, data-driven simulation, and prescriptive analytics. You’ll thus realize how anticipating multiple conflicting scenarios contributes for preserving the decision maker future freedom of action, as preferences are learned and refined over time. You’ll then be in a good position to understand how an anticipatory hypervolume-based multi-objective Bayesian metaheuristic can incorporate meta-preferences to improve financial portfolio selection in real and simulated markets. In addition, you’ll learn about connections between conditional future hypervolume maximization and the causal entropic principle proposed by Wissner-Gross and Freer (2013). Finally, you’ll be stimulated to engage in a discussion about the relevance of the anticipatory multi-objective approach to artificial general intelligence. I hope you join us for an exciting discussion!
In this talk I will show how Random Forests is built and optimized for the best results in R.
You've just fine-tuned your strongest model, demonstrated uncanny accuracy on your test set, and the board now wants you to build a product on top of it. Congratulations, now the real work begins! In this talk, I will share lessons learned from building business-critical machine learning systems that reliably meet product requirements in the face of unexpected data distribution changes, rapid code iteration, large infrastructure changes, and loss of shared context within growing teams.
We are going to discuss the status of quantum computing in terms of potential as well as architectures and cloud services that will become available in the near-term, and their usefulness and programmability challenges. This includes an introduction to Quantum Annealing and the D-Wave machines. A good part will be devoted to questions and answers (write to firstname.lastname@example.org)
The purpose of this paper is to present a brief history of the field of cybernetics that is concerned with the simulation of biological brains for the purposes of controlling both machines and industrial processes. This paper will also discuss the current understanding of how biological brains function and learn. It will then discuss the current state of the art in neuromorphic technology design and where it falls short of actually faithfully functioning like a biological brain. It will then discuss how primarily hardware based computational modeling methods can far more rapidly and faithfully model the functionality of biological brains. Finally, it shall conclude with a discussion of the advantages of hardware based computational design over both conventional digital hardware implementation and the ROM-ified code versions of transforming software based ANN’s (Attractor Neural Networks) into hardware. It then discusses some hardware examples of each of the analogs to these biological neural systems that have been discussed and concludes with the advantages of this design approach over the conventional approaches used in neuromorphic hardware design.
With the rise of bots, bot user data has become crucial to extract meaningful insights from. Thus, the automation of analytics systems, will help bot makers to analyze their data by segmenting user conversations, understanding the sentiments of users, and just let the system take the action on behalf of data scientists. As we're experiencing a breakthrough, humans and automation systems will be in charge together unlike it was used to be..
The theme of the presentation is AI strategy for executives and goes into the following:
1. Recent breakthroughs in AI;
2. How it applies to business applications;
3. Things to consider when planning your organization's AI strategy;
4. Common risks and mistakes
In this talk, I will discuss recent advances and key questions and challenges at the intersection of AI and Security: how AI and deep learning can enable better security, and how Security can enable better AI.
Word Embeddings are both a hot research topic and a useful tool for NLP practitioners, as they provide representations that are useful in many intermediate tasks, like part-of-speech tagging, syntactic parsing or named entity recognition, as well as end to end tasks like text classification, sentiment analysis and question answering. The recent attention to the topic started in 2013 when the original word2vec paper was published at NIPS and alongside with an efficient and scalable implementation, but much research was carried out on the topic since the '50s' in fields like computer science, cognitive science, and computational linguistics. The Historical part of the talk will focus on this body of work, with the aim of distilling ideas and learned lessons, of which many practitioners and machine learning researchers are unaware of. The second part of the talk will focus on recent developments and novel methods, highlighting interesting directions that are being explored in the last couple years, like the role of syntax in learning embeddings, the compositionally of meaning and how to learn representations of knowledge graphs.
I will give a talk on the many applications of ML to the rich datasets we have at Quora— touching on recommendation engines, NLP with neural nets, and the difference that scale makes
Deep Learning today is made up of large research body with a large focus on innovation. It's still seen as a complex subject out of the reach of practicioners. In this talk we will cover the requirements for building a production deep learning system as well as some of the problems when running a deep learning application at scale.