14.09.14 - 19.09.14, Seminar 14381

Neural-Symbolic Learning and Reasoning

The following text appeared on our web pages prior to the seminar, and was included as part of the invitation.

Motivation

The goal of this Dagstuhl seminar is to build bridges between symbolic and sub-symbolic reasoning and learning representations using computer vision as a catalyst application. This will require an ability to handle big data and lifelong learning in computer vision, and enable the systematic evaluation of neural-symbolic frameworks and systems. The seminar's main focus is the integration of methods and techniques from neural computation and machine learning, cognitive science and applied logic, and visual information processing. The expected outcomes include a neural-symbolic manifesto with an in-depth re-thinking of learning models capable of incorporating symbolic representations and the definition of a neural-symbolic big-data challenge and evaluation framework.

Computational systems are having to deal with huge data sets, e.g. billions of videos on the web. A major research challenge, therefore, is how computing will respond to the needs of society in such a world where data grows exponentially. Real-world data is noisy, often incomplete and high-dimensional. The computational systems of the 21st century will have to be robust, flexible, modular and expressive in order to face up to this challenge. The techniques already being applied to tackle the big data problem can be divided into symbolic (using e.g. spatial logic, inductive logic programming) and sub-symbolic (e.g. neural networks, statistical/graphical models), supervised (using labelled data), unsupervised (unlabelled data) and semi-supervised, shallow (e.g. support vector machines) and deep-learning techniques (e.g. deep belief networks). The deep network representation allows the sharing of common features and the selective composition of features that can bring about efficient computation and good generalisation. While results in computer vision indicate that the deep learning approach is promising, considerable research is needed to realise the goals of modularity, robustness and flexibility. In particular, a better understanding of the processes underlying concept formation (like e.g. the concept of a shape) is required to enable a more systematic network set-up, robust learning and validation (i.e. explanation), and the integration of deep networks with state-of-the-art systems that are symbolic. To this end, neural-symbolic systems seek to make use of the reasoning capacities of logic, and the learning capacities of network models. In a neural-symbolic system, neural networks provide the machinery for parallel computation and robust learning, while logic provides knowledge representation and reasoning, and explanation capability to the neural models, facilitating transfer learning and the interaction between the models and the world. In this integrated model, no conflict arises between a continuous and a discrete component of the system. Instead, a tightly-coupled hybrid system exists that is continuous by nature, but that has a high-level, discrete interpretation. Neural-symbolic systems have application in knowledge acquisition areas such as visual intelligence, where a system needs to learn to adapt to changes in the environment, and to reason about what has been learned in order to respond to a new situation.

Several challenges arise in this context, which can be categorized into the following three major topics for discussion: Symbolic knowledge representation, reasoning and learning by connectionist systems (includes comparisons with purely-symbolic and purely-connectionist models, representation of temporal, modal, commonsense, relational, first-order and higher-order reasoning, emergence, connectivity, abstraction, causality and analogy); Extraction of high-level concepts and knowledge from complex networks (deals with issues of efficient and effective knowledge extraction from very large networks, comprehensibility, explanation, validation, maintenance and transfer learning); Applications in vision, robotics, simulation and the web (includes learning and description of actions and other high-level concepts from large videos, sound and sensor data, noise robustness, gap-filling and anomaly detection in surveillance data, etc.). From a more practical perspective, complex systems engineering questions also arise. Most models in science rely on numerical models while the high-level concepts referred to above tend to be feature-based. How and to which degree should we integrate geometric approaches into representation languages and still preserve tractability as a major criterion for large scale reasoning and robust learning? How do we assist users in modelling and how can we semiautomatically transform this modelled knowledge for use in analogous domains? Humans categorize and reason based on analogies and similarities. Traditional (classical) logic-based reasoning is rigid in the sense that it either produces correct answers or no answers. Our goal of flexibility requires support for gradual correctness, and handling of incomplete or contradictory information. Approximate and non-classical reasoning is needed so that learning from big data can incorporate analogy, similarity and commonsense reasoning.

The emergence of symbolic representations is natural whenever one wants to tackle complex problems that are inherently associated with huge collections of data. The workshop will promote the idea of intelligent agents that interact with the environment by life-long learning. This is particularly relevant when machine learning meets computer vision. As part of the turn of neural-symbolic integration towards the practical, the seminar will also include the evaluation of algorithms and methods. This is another fundamental issue; one of the most serious drawbacks in our field is the lack of relevant and systematic evaluation mechanisms of the research. New ideas in this direction will be discussed, with challenges posed and demonstrations given and evaluated.

The seminar will mark the 10th anniversary of the workshop series on neural-symbolic learning and reasoning (NeSy). NeSy has been gathering members of the Artificial Intelligence, Neural Computation and Cognitive Science communities since 2005, but for only one day. At the last NeSy workshop it became clear that this was not enough, given that these communities share many common goals and aspirations, but are still largely disconnected in the organization, publication and sharing of research results and systems. The desire of many at NeSy to go deeper into the understanding of the main positions and issues, and to collaborate in a truly multidisciplinary way, using computer vision as a catalyst towards achieving specific objectives, has prompted us to put together this Dagstuhl seminar marking the 10th anniversary of the workshop. We hope you will be able to attend and contribute to the seminar.