Random forests are a combination of tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest. The generalization error for forests converges a.s. to a limit as the number of trees in the forest becomes large. The generalization error of a forest of tree classifiers depends on the strength of the individual trees in the forest and the correlation between them. Using a random selection of features to split each node yields error rates that compare favorably to Adaboost (Y. Freund & R. Schapire, Machine Learning: Proceedings of the Thirteenth International conference, aaa, 148–156), but are more robust with respect to noise. Internal estimates monitor error, strength, and correlation and these are used to show the response to increasing the number of features used in the splitting. Internal estimates are also used to measure variable importance. These ideas are also applicable to regression.

Random Forests

Multilayer neural networks trained with the back-propagation algorithm constitute the best example of a successful gradient based learning technique. Given an appropriate network architecture, gradient-based learning algorithms can be used to synthesize a complex decision surface that can classify high-dimensional patterns, such as handwritten characters, with minimal preprocessing. This paper reviews various methods applied to handwritten character recognition and compares them on a standard handwritten digit recognition task. Convolutional neural networks, which are specifically designed to deal with the variability of 2D shapes, are shown to outperform all other techniques. Real-life document recognition systems are composed of multiple modules including field extraction, segmentation recognition, and language modeling. A new learning paradigm, called graph transformer networks (GTN), allows such multimodule systems to be trained globally using gradient-based methods so as to minimize an overall performance measure. Two systems for online handwriting recognition are described. Experiments demonstrate the advantage of global training, and the flexibility of graph transformer networks. A graph transformer network for reading a bank cheque is also described. It uses convolutional neural network character recognizers combined with global training techniques to provide record accuracy on business and personal cheques. It is deployed commercially and reads several million cheques per day.

Gradient-based learning applied to document recognition

We propose a deep convolutional neural network architecture codenamed Inception that achieves the new state of the art for classification and detection in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC14). The main hallmark of this architecture is the improved utilization of the computing resources inside the network. By a carefully crafted design, we increased the depth and width of the network while keeping the computational budget constant. To optimize quality, the architectural decisions were based on the Hebbian principle and the intuition of multi-scale processing. One particular incarnation used in our submission for ILSVRC14 is called GoogLeNet, a 22 layers deep network, the quality of which is assessed in the context of classification and detection.

/pdf/going-deeper-with-convolutions-1yobw2o2ds.pdf

Going deeper with convolutions

Deep learning is a form of machine learning that enables computers to learn from experience and understand the world in terms of a hierarchy of concepts. Because the computer gathers knowledge from experience, there is no need for a human computer operator to formally specify all the knowledge that the computer needs. The hierarchy of concepts allows the computer to learn complicated concepts by building them out of simpler ones; a graph of these hierarchies would be many layers deep. This book introduces a broad range of topics in deep learning. The text offers mathematical and conceptual background, covering relevant concepts in linear algebra, probability theory and information theory, numerical computation, and machine learning. It describes deep learning techniques used by practitioners in industry, including deep feedforward networks, regularization, optimization algorithms, convolutional networks, sequence modeling, and practical methodology; and it surveys such applications as natural language processing, speech recognition, computer vision, online recommendation systems, bioinformatics, and videogames. Finally, the book offers research perspectives, covering such theoretical topics as linear factor models, autoencoders, representation learning, structured probabilistic models, Monte Carlo methods, the partition function, approximate inference, and deep generative models. Deep Learning can be used by undergraduate or graduate students planning careers in either industry or research, and by software engineers who want to begin using deep learning in their products or platforms. A website offers supplementary material for both readers and instructors.

Deep Learning

Reinforcement learning, one of the most active research areas in artificial intelligence, is a computational approach to learning whereby an agent tries to maximize the total amount of reward it receives when interacting with a complex, uncertain environment. In Reinforcement Learning, Richard Sutton and Andrew Barto provide a clear and simple account of the key ideas and algorithms of reinforcement learning. Their discussion ranges from the history of the field's intellectual foundations to the most recent developments and applications. The only necessary mathematical background is familiarity with elementary concepts of probability. The book is divided into three parts. Part I defines the reinforcement learning problem in terms of Markov decision processes. Part II provides basic solution methods: dynamic programming, Monte Carlo methods, and temporal-difference learning. Part III presents a unified view of the solution methods and incorporates artificial neural networks, eligibility traces, and planning; the two final chapters present case studies and consider the future of reinforcement learning.

/pdf/reinforcement-learning-an-introduction-rzxgej9p17.pdf

Reinforcement Learning: An Introduction

Tasks have been identified as playing an important role to knowledge workers as high-level units for organizing their information. TaskTracer is a task-aware desktop system that leverages the semi-automatic association between tasks and information to provide intelligent assistance to the user. TaskPredictor is a component that attempts to automatically predict the current task of the user. Our experience with TaskPredictor tells us that there still remain challenges to increasing acceptance by users. We describe our approaches to reduce the number of interruptions to the user, and to improve the accuracy of predictions. Projects, Activities, and Tasks Projects, tasks, activities, and to-do items have been identified as playing an important role to knowledge workers, i.e. computer users who create and process information as part of their work [2, 5, 7, 9]. These activities are conceptual structuring devices – they are the units into which a person divides their work, and the names associated with those units. Some tasks correspond to concrete projects like "Conduct user study" or "Create end of year report". Other tasks may be ongoing, with no start date, end date, or deliverables, such as "Advise students". More intuitively, tasks might correspond to the folders that you would find in a physical filing cabinet, or to a bulleted list that a person might write describing the projects and tasks for which they are responsible. To take advantage of tasks, a few recent task-centric solutions have been developed for email [4], or for desktop activity [12, 3, 16]. However, these systems have so far not exploited the support that could be provided by intelligent assistants to increase the productivity of knowledge workers. Task Prediction for Desktop Activities TaskTracer [8] is a system that allows users to organize their information by labeling them with high-level units that make sense to them (which we call tasks), such as Copyright © 2006, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. "teach CS534", "prepare NSF proposal B", "make travel arrangements for conference C". We assume that the behavior of the user at the desktop is an interleaved timeline of different tasks and each task is associated with a set of resources (e.g. documents, folders, web pages, emails, contacts, etc.) relevant to that task. Once tasks have been defined, the user can indicate to the system the name of the current task and switch to new tasks. TaskTracer records all resources that are accessed by the user and automatically associates them with the current declared task. This data can be used by machine learning components to provide intelligent assistance to the user to make finding and refinding information easier [6, 13]. These benefits assume that the data is not too noisy and therefore the user must let the system know when they have changed tasks. To reduce the cost of task switches, we have been working to supplement the user's manual task declaration with a TaskPredictor (Figure 1) that attempts to automatically predict the current task of the user [14, 15]. Our experience with TaskPredictor tells us that there still remain two main challenges to increasing acceptance by users: 1) reducing the number of interruptions to the user, and 2) improving the accuracy of predictions. 1 The types of tasks that we typically encounter have no or only loosely fixed structure and are highly distinctive of the individual knowledge worker. A plan often has flexibility but the user is executing a specific structured activity. We are not suggesting that these two approaches are mutually exclusive. Figure 1. The TaskPredictor icon makes a prediction and informs the user through a balloon notification. Also shown are the TaskSelector to indicate tasks (on left) and the main TaskTracer icon (on right). Reducing the number of interruptions In order to reduce the number of interruptions to the user, a cost-sensitive approach to task prediction and notification in the user interface is needed [10]. Currently, we already reduce the number of possible notifications made by the TaskPredictor through two main ways. First, we assume that in an unbroken segment of time in which a particular window has focus and the name of the document in that window does not change, the user is still working on the same task and hence we do not have to make a prediction. Second, to keep user costs within reason for TaskPredictor, we do not make a prediction for every window/document. We are willing to accept lower coverage in order to maintain high precision, because we would like to minimize the number of interruptions, particularly if these interruptions may provide wrong predictions. Thus, the TaskPredictor employs a probabilistic threshold for predicting tasks for windows/documents; if the threshold is not reached then notification is not made. We are currently exploring further options to better address user cost issues by exploiting task boundaries [1, 11]. Improving the accuracy of predictions One way of increasing the accuracy of predictions is by incorporating user feedback. TaskPredictor at the moment only allows the user to indicate that a prediction was wrong and to specify what the correct prediction should have been. This is just a glimpse of the rich knowledge users have about the correct prediction, and we would like to better harness the user’s knowledge to improve learning. To investigate the feasibility of rich user feedback, we conducted a formative think-aloud study with 13 email users using 122 messages from the publicly available Enron dataset (user farmer-d). In this study [17], machine learning algorithms (Ripper and Naive Bayes) sorted email messages into folders and explained their reasoning using three different simplified but faithful explanation paradigms: Rule-based, Keyword-based, and Similaritybased. The participants were asked to provide feedback to improve the pre-dictions. No restrictions were placed upon the form or con-tent of participants’ feedback. We observed and video-taped their activities and comments throughout the experiment, as well as collecting their work products, and questionnaires investigating their preferences and behavior. From a user perspective, we assessed the participants’ willingness to provide feedback, accuracy in doing so, and ability to understand the different explanations. From an algorithm perspective, we analyzed the participants’ feedback to determine how easily its types of advice could be understood and assimilated by machine learning algorithms. The system explains to the user We provided simplified but faithful explanations of the learning algorithms, which were concrete in terms of specific features that were visible in the current email message. Using these design principles, we found that Rule-based explanations were the most understandable and preferred but that a substantial number of the participants chose one of the other explanation paradigms as their favorite. This implies that machine learning systems may need to support multiple explanation paradigms in order to effectively reach all of their users. The factors that contributed to participants' overall preference of an explanation were approval of reasoning soundness, clear communication of reasoning, perceived accuracy, and a less technical style of expression. Similarity-based explanations had serious understandability problems in our experiment. The user gives rich feedback to the system The participant corrections brought the accuracy rates for all paradigms to almost identical levels (71-72%). When the participants disagreed with the machine, participants were usually right, but not always (introducing some noise into the data). Table 1 shows the results of our analysis in terms of what type of feedback participants made and the knowledge upon which it was based. Among the feedback were requests for reweighting features, feature combinations, relational features, and even wholesale changes to the algorithms. Almost a third of the participations’ suggestions relied on knowledge of English, or on some knowledge that could be encoded once and then reused. Roughly half of the suggestions for improvement appear to be amenable to automated assimilation with existing methods. The results open new questions for research on methods for assimilating complex user suggestions for feature extraction, relational features, and incorporating KBEnglish KBcommon sense KBdomain KBother Total % 1. Adjust weight 11 11 4 13 39 12% 2. Select different features (words) 70 64 25 16 175 53% 3. Parse or extract in a different way 7 17 10 0 34 10% 4. Employ feature combinations 9 5 2 1 17 5% 5. Relational features 0 9 5 0 14 4% 6. Other 3 12 4 33 52 16% Total 100 118 50 63 331

/pdf/improving-intelligent-assistants-for-desktop-activities-1m0ljdpphw.pdf

Improving Intelligent Assistants for Desktop Activities.

Emergence of Motion-sensitive Neurons Properties by Learning Sparse Code for Natural Moving Images

Identification and population counts of soil mesofauna can be an important tool for soil ecologists to determine soil biodiversity. The process of performing population counts, which includes classifying and sorting specimens, is very time consuming because of the large diversity and quantities of specimens in soil samples. A mechanical system was designed and developed to automate this process. The system featured automated imaging, positioning, and sorting. Images acquired with system were used by pattern recognition software to classify individual specimens from a soil sample. Positioning was a required function of the system for automation and was accomplished using a pair of linear stages mounted orthogonally to each other. The system also featured a six-axis robot attached with a pipette end effector for sorting classified specimens, a required function for validation. The final system was successful in creating high-quality images of 19 species of soil mesofauna and sorting them.

Design of an Automated System for Imaging and Sorting Soil Mesofauna

Induction: Weak but essential

Automated recognition of object categories in images is a critical step for many real-world computer vision applications. Interest region detectors and region descriptors have been widely employed to tackle the variability of objects in pose, scale, lighting, texture, color, and so on. Different types of object recognition problems usually require different image features and corresponding learning algorithms. This dissertation focuses on the design, evaluation and application of new image features and learning algorithms for the recognition of biological, generic and social objects. The first part of the dissertation introduces a new structure-based interest region detector called the principal curvature-based region detector (PCBR) which detects stable watershed regions that are robust to local intensity perturbations. This detector is specifically designed for region detection for biological objects. Several recognition architectures are then developed that fuse visual information from disparate types of image features for the categorization of complex objects. The described image features and learning algorithms achieve excellent performance on the difficult stonefly larvae dataset. The second part of the dissertation presents studies of methods for visual codebook learning and their application to object recognition. The dissertation first introduces the methodology and application of generative visual codebooks for stonefly recognition and introduces a discriminative evaluation methodology based on a maximum mutual information criterion. Then a new generative/discriminative visual codebook learning algorithm, called iterative discriminative clustering (IDC), is presented that refines the centers and the shapes of the generative codewords for improved discriminative power. It is followed by a novel codebook learning algorithm that builds multiple codebooks that are non-redundant in discriminative power. All these visual codebook learning algorithms achieve high performance on both biological and generic object recognition tasks. The final part of the dissertation describes a socially-driven clothes recognition system for an intelligent fitting-room system. The dissertation presents the results of a user study to identify the key factors for clothes recognition. It then describes learning algorithms for recognizing these key factors from clothes images using various image features. The clothes recognition system successfully enables automated social fashion information retrieval for an enhanced clothes shopping experience.

/pdf/image-features-and-learning-algorithms-for-biological-14o6hgj2xt.pdf

Thomas G. Dietterich

Papers

Improving Intelligent Assistants for Desktop Activities.

Emergence of Motion-sensitive Neurons Properties by Learning Sparse Code for Natural Moving Images

Design of an Automated System for Imaging and Sorting Soil Mesofauna

Induction: Weak but essential

Image features and learning algorithms for biological, generic and social object recognition