Reward functions, learned or manually specified, are rarely perfect. Instead of accurately expressing human goals, these reward functions are often distorted by human beliefs about how best to achieve those goals. Specifically, these reward functions often express a combination of the human's terminal goals -- those which are ends in themselves -- and the human's instrumental goals -- those which are means to an end. We formulate a simple example in which even slight conflation of instrumental and terminal goals results in severe misalignment: optimizing the misspecified reward function results in poor performance when measured by the true reward function. This example distills the essential properties of environments that make reinforcement learning highly sensitive to conflation of instrumental and terminal goals. We discuss how this issue can arise with a common approach to reward learning and how it can manifest in real environments.

We introduce model editing with canonical examples, a setting in which (1) a single learning example is provided per desired behavior, (2) evaluation is performed exclusively out-of-distribution, and (3) deviation from an initial model is strictly limited. A canonical example is a simple instance of good behavior, e.g., The capital of Mauritius is Port Louis) or bad behavior, e.g., An aspect of researchers is coldhearted). The evaluation set contains more complex examples of each behavior (like a paragraph in which the capital of Mauritius is called for.) We create three datasets and modify three more for model editing with canonical examples, covering knowledge-intensive improvements, social bias mitigation, and syntactic edge cases. In our experiments on Pythia language models, we find that LoRA outperforms full finetuning and MEMIT. We then turn to the Backpack language model architecture because it is intended to enable targeted improvement. The Backpack defines a large bank of sense vectors--a decomposition of the different uses of each word--which are weighted and summed to form the output logits of the model. We propose sense finetuning, which selects and finetunes a few ($\approx$ 10) sense vectors for each canonical example, and find that it outperforms other finetuning methods, e.g., 4.8% improvement vs 0.3%. Finally, we improve GPT-J-6B by an inference-time ensemble with just the changes from sense finetuning of a 35x smaller Backpack, in one setting outperforming editing GPT-J itself (4.1% vs 1.0%).

Learning in small sample regimes is among the most remarkable features of the human perceptual system. This ability is related to robustness to transformations, which is acquired through visual experience in the form of weak- or self-supervision during development. We explore the idea of allowing artificial systems to learn representations of visual stimuli through weak supervision prior to downstream supervised tasks. We introduce a novel loss function for representation learning using unlabeled image sets and video sequences, and experimentally demonstrate that these representations support one-shot learning and reduce the sample complexity of multiple recognition tasks. We establish the existence of a trade-off between the sizes of weakly supervised, automatically obtained from video sequences, and fully supervised data sets. Our results suggest that equivalence sets other than class labels, which are abundant in unlabeled visual experience, can be used for self-supervised learning of semantically relevant image embeddings.

Deep neural networks are vulnerable to adversarial examples. Prior defenses attempted to make deep networks more robust by either improving the network architecture or adding adversarial examples into the training set, with their respective limitations. We propose a new direction. Motivated by recent research that shows that outliers in the training set have a high negative influence on the trained model, our approach makes the model more robust by detecting and removing outliers in the training set without modifying the network architecture or requiring adversarial examples. We propose two methods for detecting outliers based on canonical examples and on training errors, respectively. After removing the outliers, we train the classifier with the remaining examples to obtain a sanitized model. Our evaluation shows that the sanitized model improves classification accuracy and forces the attacks to generate adversarial examples with higher distortions. Moreover, the Kullback-Leibler divergence from the output of the original model to that of the sanitized model allows us to distinguish between normal and adversarial examples reliably.

Machine learning has already extended into so many aspects of daily life that it can be handy for us to memorize a set of go-to examples of its impact on certain industries. For instance, we might think of fraud detection as the canonical example of machine learning in the financial sector. Or we might think of Watson's cognitive approach to oncology as the canonical example of machine learning in healthcare. Or, yet again, we might point to recommendation engines at Netflix and Amazon as canonical examples of machine learning in retail. If you are interested to try out new IBM's Watson Machine Learning service - click here To try the Data Science Experience - subscribe for the free trial Certainly, those are tremendous demonstrations of the power of the technology -- and in aggregate, they give a sense of machine learning's pervasive presence in our lives.

Welcome to the first of a series of posts where we dive into use cases that are shaping and expanding what's possible with machine learning. Each month, we'll post the Top 10 use cases from the IBM Machine Learning Hub -- where IBM hosts data scientists to collaborate on prototypes with our in-house ML specialists. We anticipate the Top 10 will change as industries adopt machine learning at different rates and as entirely new use cases emerge -- so check back. We kick off the Top 10 with a look at three off-the-radar ways that machine learning is driving vital improvements to government agencies. Next month, we explore health care use cases.

EXAMINATION OF SHAFER'S CANONICAL EXAMPLES Kathryn Blackmond Laskey Decision Science Consortium, Inc. 7700 Leesburg Pike, Suite 421 Falls Church, VA 22043 1 Abstract In the canonical examples underlying Shafer-Dempster theory, beliefs over the hypotheses of interest are derived from a probability model for a set of auxiliary hypotheses. Beliefs are derived via a compatibility relation connecting the auxiliary hypotheses to subsets of the primary hypotheses. A belief function differs from a Bayesian probability model in that one does not condition on those parts of the evidence for which no probabilities are specified. The significance of this difference in conditioning assumptions is illustrated with two examples giving rise to identical belief functions but different Bayesian probability distributions. Introduction The artificial intelligence community is in the midst of a lively debate over the representation and manipulation of uncertainty.

This paper is concerned with two theories of probability judgment: the Bayesian theory and the theory of belief functions. It illustrates these theories with some simple examples and discusses some of the issues that arise when we try to implement them in expert systems. The Bayesian theory is well known; its main ideas go back to the work of Thomas Bayes (1702-1761). The theory of belief functions, often called the Dempster-Shafer theory in the artificial intelligence community, is less well known, but it has even older antecedents; belief-function arguments appear in the work of George Hooper (16401723) and James Bernoulli (1654-1705). For elementary expositions of the theory of belief functions, see Shafer (1976, 1985).