Abstract In people, the ability to solve analogies such as “body: feet:: table: ?” emerges in childhood, and appears to transfer easily to other domains, such as the visual domain “(: ) :: < : ?”. Recent research shows that large language models (LLMs) can solve various forms of analogies. However, can LLMs generalize analogy solving to other domains like people can? To investigate this, we had children, adults, and LLMs solve a series of letter-string analogies (e.g., a b : a c :: j k : ?) in the Latin alphabet, in a near transfer domain (Greek alphabet), and a far transfer domain (list of symbols). Children and adults easily generalized their knowledge to unfamiliar domains, whereas LLMs did not. This key difference between human and AI performance is evidence that these LLMs still struggle with robust human-like analogical transfer.

We characterise the problem of abstraction in the context of deep reinforcement learning. Various well established approaches to analogical reasoning and associative memory might be brought to bear on this issue, but they present difficulties because of the need for end-to-end differentiability. We review developments in AI and machine learning that could facilitate their adoption.

Adversarial examples for neural network image classifiers are known to be transferable: examples optimized to be misclassified by a source classifier are often misclassified as well by classifiers with different architectures. However, targeted adversarial examples—optimized to be classified as a chosen target class—tend to be less transferable between architectures. While prior research on constructing transferable targeted attacks has focused on improving the optimization procedure, in this work we examine the role of the source classifier. Here, we show that training the source classifier to be "slightly robust"—that is, robust to small-magnitude adversarial examples—substantially improves the transferability of class-targeted and representation-targeted adversarial attacks, even between architectures as different as convolutional neural networks and transformers. The results we present provide insight into the nature of adversarial examples as well as the mechanisms underlying so-called "robust" classifiers.

We devise an algorithm using a Bayesian optimization framework in conjunction with contextual visual data for the efficient localization of objects in still images. Recent research has demonstrated substantial progress in object localization and related tasks for computer vision. However, many current state-of-the-art object localization procedures still suffer from inaccuracy and inefficiency, in addition to failing to provide a principled and interpretable system amenable to high-level vision tasks. We address these issues with the current research. Our method encompasses an active search procedure that uses contextual data to generate initial bounding-box proposals for a target object. We train a convolutional neural network to approximate an offset distance from the target object. Next, we use a Gaussian Process to model this offset response signal over the search space of the target. We then employ a Bayesian active search for accurate localization of the target. In experiments, we compare our approach to a state-of-the-art bounding-box regression method for a challenging pedestrian localization task. Our method exhibits a substantial improvement over this baseline regression method.

A major goal of computer vision is to enable computers to interpret visual situations - abstract concepts (e.g., “a person walking a dog,” “a crowd waiting for a bus,” “a picnic”) whose image instantiations are linked more by their common spatial and semantic structure than by low-level visual similarity. In this paper, we propose a novel method for prior learning and active object localization for this kind of knowledge-driven search in static images. In our system, prior situation knowledge is captured by a set of flexible, kernel-based density estimations - a situation model - that represent the expected spatial structure of the given situation. These estimations are updated by information gained as the system searches for relevant objects, allowing the system to use context as it is discovered to narrow the search. More specifically, at any given time in a run on a test image, our system uses image features plus contextual information it has discovered to identify a small subset of training images - an importance cluster - that is deemed most similar to the given test image, given the context. This subset is used to generate an updated situation model in an on-line fashion, using an efficient multipole expansion technique. As a proof of concept, we apply our algorithm to a highly varied and challenging dataset consisting of instances of a “dogwalking” situation. Our results support the hypothesis that dynamically-rendered, context-based probabilistic models can support efficient object localization in visual situations. Moreover, our approach is general enough to be applied to diverse machine learning paradigms requiring interpretable, probabilistic representations generated from partially observed data.

Hierarchical networks are known to achieve high classification accuracy on difficult machine-learning tasks. For many applications, a clear explanation of why the data was classified a certain way is just as important as the classification itself. However, the complexity of hierarchical networks makes them ill-suited for existing explanation methods. We propose a new method, contribution propagation, that gives per-instance explanations of a trained network's classifications. We give theoretical foundations for the proposed method, and evaluate its correctness empirically. Finally, we use the resulting explanations to reveal unexpected behavior of networks that achieve high accuracy on visual object-recognition tasks using well-known data sets.

We investigate the role of learned shape-prototypes in an influential family of hierarchical neural-network models of vision. Central to these networks' design is a dictionary of learned shapes, which are meant to respond to discriminative visual patterns in the input. While higher-level features based on such learned prototypes have been cited as key for viewpoint-invariant object-recognition in these models [1], [2], we show that high performance on invariant object-recognition tasks can be obtained by using a simple set of unlearned, “shape-free” features. This behavior is robust to the size of the network. These results call into question the roles of learning and shape-specificity in the success of such models on difficult vision tasks, and suggest that randomly constructed prototypes may provide a useful “universal” dictionary.

In this note we argue that biological computation is a process that occurs in nature, not merely in computer simulations of nature.

This paper presents a genetic algorithm (GA) for combining representations of learned priors such as shape, regional properties and relative location of organs into a single framework in order to perform automated segmentation of the prostate. Prostate segmentation is typically performed manually by an expert physician and is used to determine the locations for radioactive seed placement during radiotherapy treatment planning. The GA accounts for the uncertainty in the definitions of tumor margins by combining known representations of shape, texture and relative location of organs to perform automatic segmentation in two (2D) as well as three dimensions (3D).

Segmentation of medical images is challenging due to poor image contrast and artifacts that result in missing or diffuse organ/tissue boundaries. Consequently, this task involves incorporating as much prior information as possible (e.g., texture, shape, and spatial location of organs) into a single framework. In this paper, we present a genetic algorithm for automating the segmentation of the prostate on two-dimensional slices of pelvic computed tomography (CT) images. In this approach the segmenting curve is represented using a level set function, which is evolved using a genetic algorithm (GA). Shape and textural priors derived from manually segmented images are used to constrain the evolution of the segmenting curve over successive generations.We review some of the existing medical image segmentation techniques. We also compare the results of our algorithm with those of a simple texture extraction algorithm (Laws' texture measures) as well as with another GA-based segmentation tool called GENIE. Our preliminary tests on a small population of segmenting contours show promise by converging on the prostate region. We expect that further improvements can be achieved by incorporating spatial relationships between anatomical landmarks, and extending the methodology to three dimensions.

We investigate the results of coevolution of spatially distributed populations. In particular, we describe work in which a simple function approximation problem is used to compare different spatial evolutionary methods. Our work shows that, on this problem, spatial coevolution is dramatically more successful than any other spatial evolutionary scheme we tested. Our results support two hypotheses about the source of spatial coevolution's superior performance: (1) spatial coevolution allows population diversity to persist over many generations; and (2) spatial coevolution produces training examples ("parasites") that specifically target weaknesses in models ("hosts"). The precise mechanisms by which the combination of spatial embedding and coevolution produces these results are still not well understood.

Previous work on coevolutionary search has demonstrated both successful and unsuccessful applications. As a step in explaining what factors lead to success or failure, we present a comparative study of an evolutionary and a coevolutionary search model. In the latter model, strategies for solving a problem coevolve with training cases. We find that the coevolutionary model has a relatively large efficacy: 86 out of 100 (86%) of the simulations produce high quality strategies. In contrast, the evolutionary model has a very low efficacy: a high quality strategy is found in only two out of 100 runs (2%). We show that the increased efficacy in the coevolutionary model results from the direct exploitation of low quality strategies by the population of training cases. We also present evidence that the generality of the high-quality strategies can suffer as a result of this same exploitation.

Coevolution, between a population of candidate solutions and a population of test cases, has received increasing attention as a promising biologically inspired method for improving the performance of evolutionary computation techniques. However, the results of studies of coevolution have been mixed. One of the seemingly more impressive results to date was the improvement via coevolution demonstrated by Juille and Pollack (1998) on evolving cellular automata to perform a classification task. Their study, however, like most other studies on coevolution, did not investigate the mechanisms giving rise to the observed improvements. In this paper, we probe more deeply into the reasons for these observed improvements and present empirical evidence that, in contrast to what was claimed by Juille and Pollack, much of the improvement seen was due to their "resource sharing" technique rather than to coevolution. We also present empirical evidence that resource sharing works, at least in part, by preserving diversity in the population.

Genetic algorithms are computational models of evolution that play a central role in many artificial-life models. We review the history and current scope of research on genetic algorithms in artificial life, giving illustrative examples in which the genetic algorithm is used to study how learning and evolution interact, and to model ecosystems, immune system, cognitive systems, and social systems. We also outline a number of open questions and future directions for genetic algorithms in artificial-life research.