Repairing Learned Knowledge Using Experience

Explanation-based learning occurs when something useful is retained from an explanation, usually an account of how some particular problem can be solved given a sound theory. Many real-world explanations are not based on sound theory, however, and wrong things may be learned accidentally, as subsequent failures will likely demonstrate. In this paper, we describe ways to isolate the facts that cause failures, ways to explain why those facts cause problems, and ways to repair learning mistakes. In particular, our program learns to distinguish pails from cups after making a few mistakes.

Battling Reality

In the four years that the MIT Mobile Robot Project has benn in existence, we have built ten robots that focus research in various areas concerned with building intelligent systems. Towards this end, we have embarked on trying to build useful autonomous creatures that live and work in the real world. Many of the preconceived notions entertained before we started building our robots turned out to be misguided. Some issues we thought would be hard have worked successfully from day one and subsystems we imagined to be trivial have become tremendous time sinks. Oddly enough, one of our biggest failures has led to some of our favorite successes. This paper describes the changing paths our research has taken due to the lessons learned from the practical realities of building robots.

Linear Object Classes and Image Synthesis from a Single Example Image

The need to generate new views of a 3D object from a single real image arises in several fields, including graphics and object recognition. While the traditional approach relies on the use of 3D models, we have recently introduced techniques that are applicable under restricted conditions but simpler. The approach exploits image transformations that are specific to the relevant object class and learnable from example views of other "prototypical" objects of the same class. In this paper, we introduce such a new technique by extending the notion of linear class first proposed by Poggio and Vetter. For linear object classes it is shown that linear transformations can be learned exactly from a basis set of 2D prototypical views. We demonstrate the approach on artificial objects and then show preliminary evidence that the technique can effectively "rotate" high- resolution face images from a single 2D view.

Spatial Reference Frames for Object Recognition: Tuning for Rotations in Depth

The inferior temporal cortex (IT) of monkeys is thought to play an essential role in visual object recognition. Inferotemporal neurons are known to respond to complex visual stimuli, including patterns like faces, hands, or other body parts. What is the role of such neurons in object recognition? The present study examines this question in combined psychophysical and electrophysiological experiments, in which monkeys learned to classify and recognize novel visual 3D objects. A population of neurons in IT were found to respond selectively to such objects that the monkeys had recently learned to recognize. A large majority of these cells discharged maximally for one view of the object, while their response fell off gradually as the object was rotated away from the neuron"s preferred view. Most neurons exhibited orientation-dependent responses also during view-plane rotations. Some neurons were found tuned around two views of the same object, while a very small number of cells responded in a view- invariant manner. For five different objects that were extensively used during the training of the animals, and for which behavioral performance became view-independent, multiple cells were found that were tuned around different views of the same object. No selective responses were ever encountered for views that the animal systematically failed to recognize. The results of our experiments suggest that neurons in this area can develop a complex receptive field organization as a consequence of extensive training in the discrimination and recognition of objects. Simple geometric features did not appear to account for the neurons" selective responses. These findings support the idea that a population of neurons -- each tuned to a different object aspect, and each showing a certain degree of invariance to image transformations -- may, as an assembly, encode complex 3D objects. In such a system, several neurons may be active for any given vantage point, with a single unit acting like a blurred template for a limited neighborhood of a single view.

Learning Object-Independent Modes of Variation with Feature Flow Fields

We present a unifying framework in which "object-independent" modes of variation are learned from continuous-time data such as video sequences. These modes of variation can be used as "generators" to produce a manifold of images of a new object from a single example of that object. We develop the framework in the context of a well-known example: analyzing the modes of spatial deformations of a scene under camera movement. Our method learns a close approximation to the standard affine deformations that are expected from the geometry of the situation, and does so in a completely unsupervised (i.e. ignorant of the geometry of the situation) fashion. We stress that it is learning a "parameterization", not just the parameter values, of the data. We then demonstrate how we have used the same framework to derive a novel data-driven model of joint color change in images due to common lighting variations. The model is superior to previous models of color change in describing non-linear color changes due to lighting.

On the Dirichlet Prior and Bayesian Regularization

A common objective in learning a model from data is to recover its network structure, while the model parameters are of minor interest. For example, we may wish to recover regulatory networks from high-throughput data sources. In this paper we examine how Bayesian regularization using a Dirichlet prior over the model parameters affects the learned model structure in a domain with discrete variables. Surprisingly, a weak prior in the sense of smaller equivalent sample size leads to a strong regularization of the model structure (sparse graph) given a sufficiently large data set. In particular, the empty graph is obtained in the limit of a vanishing strength of prior belief. This is diametrically opposite to what one may expect in this limit, namely the complete graph from an (unregularized) maximum likelihood estimate. Since the prior affects the parameters as expected, the prior strength balances a "trade-off" between regularizing the parameters or the structure of the model. We demonstrate the benefits of optimizing this trade-off in the sense of predictive accuracy.

Activity Zones for Context-Aware Computing

Location is a primary cue in many context-aware computing systems, and is often represented as a global coordinate, room number, or Euclidean distance various landmarks. A user?s concept of location, however, is often defined in terms of regions in which common activities occur. We show how to partition a space into such regions based on patterns of observed user location and motion. These regions, which we call activity zones, represent regions of similar user activity, and can be used to trigger application actions, retrieve information based on previous context, and present information to users. We suggest that context-aware applications can benefit from a location representation learned from observing users. We describe an implementation of our system and present two example applications whose behavior is controlled by users? entry, exit, and presence in the zones.

Mobilized ad-hoc networks: A reinforcement learning approach

Research in mobile ad-hoc networks has focused on situations in which nodes have no control over their movements. We investigate an important but overlooked domain in which nodes do have control over their movements. Reinforcement learning methods can be used to control both packet routing decisions and node mobility, dramatically improving the connectivity of the network. We first motivate the problem by presenting theoretical bounds for the connectivity improvement of partially mobile networks and then present superior empirical results under a variety of different scenarios in which the mobile nodes in our ad-hoc network are embedded with adaptive routing policies and learned movement policies.

Contextual models for object detection using boosted random fields

We seek to both detect and segment objects in images. To exploit both local image data as well as contextual information, we introduce Boosted Random Fields (BRFs), which uses Boosting to learn the graph structure and local evidence of a conditional random field (CRF). The graph structure is learned by assembling graph fragments in an additive model. The connections between individual pixels are not very informative, but by using dense graphs, we can pool information from large regions of the image; dense models also support efficient inference. We show how contextual information from other objects can improve detection performance, both in terms of accuracy and speed, by using a computational cascade. We apply our system to detect stuff and things in office and street scenes.

Internet Fish

I have invented "Internet Fish," a novel class of resource-discovery tools designed to help users extract useful information from the Internet. Internet Fish (IFish) are semi-autonomous, persistent information brokers; users deploy individual IFish to gather and refine information related to a particular topic. An IFish will initiate research, continue to discover new sources of information, and keep tabs on new developments in that topic. As part of the information-gathering process the user interacts with his IFish to find out what it has learned, answer questions it has posed, and make suggestions for guidance. Internet Fish differ from other Internet resource discovery systems in that they are persistent, personal and dynamic. As part of the information-gathering process IFish conduct extended, long-term conversations with users as they explore. They incorporate deep structural knowledge of the organization and services of the net, and are also capable of on-the-fly reconfiguration, modification and expansion. Human users may dynamically change the IFish in response to changes in the environment, or IFish may initiate such changes itself. IFish maintain internal state, including models of its own structure, behavior, information environment and its user; these models permit an IFish to perform meta-level reasoning about its own structure. To facilitate rapid assembly of particular IFish I have created the Internet Fish Construction Kit. This system provides enabling technology for the entire class of Internet Fish tools; it facilitates both creation of new IFish as well as additions of new capabilities to existing ones. The Construction Kit includes a collection of encapsulated heuristic knowledge modules that may be combined in mix-and-match fashion to create a particular IFish; interfaces to new services written with the Construction Kit may be immediately added to "live" IFish. Using the Construction Kit I have created a demonstration IFish specialized for finding World-Wide Web documents related to a given group of documents. This "Finder" IFish includes heuristics that describe how to interact with the Web in general, explain how to take advantage of various public indexes and classification schemes, and provide a method for discovering similarity relationships among documents.

Pose-Invariant Face Recognition Using Real and Virtual Views

The problem of automatic face recognition is to visually identify a person in an input image. This task is performed by matching the input face against the faces of known people in a database of faces. Most existing work in face recognition has limited the scope of the problem, however, by dealing primarily with frontal views, neutral expressions, and fixed lighting conditions. To help generalize existing face recognition systems, we look at the problem of recognizing faces under a range of viewpoints. In particular, we consider two cases of this problem: (i) many example views are available of each person, and (ii) only one view is available per person, perhaps a driver's license or passport photograph. Ideally, we would like to address these two cases using a simple view-based approach, where a person is represented in the database by using a number of views on the viewing sphere. While the view-based approach is consistent with case (i), for case (ii) we need to augment the single real view of each person with synthetic views from other viewpoints, views we call 'virtual views'. Virtual views are generated using prior knowledge of face rotation, knowledge that is 'learned' from images of prototype faces. This prior knowledge is used to effectively rotate in depth the single real view available of each person. In this thesis, I present the view-based face recognizer, techniques for synthesizing virtual views, and experimental results using real and virtual views in the recognizer.