Enjoy these six tech articles as we come down the homestretch into the weekend. Greg Otto at fedscoop wrote this interesting piece about IBM's Watson-as-a-Service. IBM announced last week it has moved its cognitive computing system into the cloud to form the Watson Discovery Advisor, allowing researchers, academics and anyone else trying to leverage big data the ability to test programs and hypotheses at speeds never before seen. Since Watson is built to understand the nuance of natural language, this new service allows researchers to process millions of data points normally impossible for humans to handle. This can reduce project timelines from years to weeks or days.

AIs that can work out how to hack self driving cars and other vehicles to turn them into killers are coming - and sooner than many people think, a leading expert has warned. 'Such attacks, which seem like science fiction today, might become reality in the next few years,' Guy Caspi, CEO of cybersecurity start-up Deep Instinct, told CNBC's podcast'Beyond the Valley.' It raises fears that self driving cars and other technologies could be hacked, turning them into makeshift battle weapons. Security expert Guy Caspi claims the technology needed for killer vehicles already exists in self driving cars from firms like Alphabet's Waymo. Caspi says much of the technology needed for killer vehicles already exists.

Last year, eBay launched a visual search capability for its mobile app that makes it possible to find items with pictures instead of words. Now, the auction platform is making it even easier to use -- you don't even have to take screenshots of whatever it is you want to find anymore, because eBay will soon allow you to drag and drop images into its search bar. For instance, if you search for a "Hello Kitty purse" in the app and find one that catches your eye, you can drag that photo into the search bar to find listings featuring identical or similar items. It won't only give you a way to search for purchases quickly, but also to find the best deals on the website. According to eBay, its convolutional neural networks process the photo you use by transforming it into a vector representation.

AI is continually improving climate models. As the planet continues to warm, climate change impacts are worsening. In 2016, there were 772 weather and disaster events, triple the number that occurred in 1980. Twenty percent of species currently face extinction, and that number could rise to 50 percent by 2100. And even if all countries keep their Paris climate pledges, by 2100, it's likely that average global temperatures will be 3 C higher than in pre-industrial times.

The challenge of assigning importance to individual neurons in a network is of interest when interpreting deep learning models. In recent work, Dhamdhere et al. proposed Total Conductance, a "natural refinement of Integrated Gradients" for attributing importance to internal neurons. Unfortunately, the authors found that calculating conductance in tensorflow required the addition of several custom gradient operators and did not scale well. In this work, we show that the formula for Total Conductance is mathematically equivalent to Path Integrated Gradients computed on a hidden layer in the network. We provide a scalable implementation of Total Conductance using standard tensorflow gradient operators that we call Neuron Integrated Gradients. We compare Neuron Integrated Gradients to DeepLIFT, a pre-existing computationally efficient approach that is applicable to calculating internal neuron importance. We find that DeepLIFT produces strong empirical results and is faster to compute, but because it lacks the theoretical properties of Neuron Integrated Gradients, it may not always be preferred in practice. Colab notebook reproducing results: http://bit.ly/neuronintegratedgradients

Neural networks are very powerful learning systems, but they do not readily generalize from one task to the other. This is partly due to the fact that they do not learn in a compositional way, that is, by discovering skills that are shared by different tasks, and recombining them to solve new problems. In this paper, we explore the compositional generalization capabilities of recurrent neural networks (RNNs). We first propose the lookup table composition domain as a simple setup to test compositional behaviour and show that it is theoretically possible for a standard RNN to learn to behave compositionally in this domain when trained with standard gradient descent and provided with additional supervision. We then remove this additional supervision and perform a search over a large number of model initializations to investigate the proportion of RNNs that can still converge to a compositional solution. We discover that a small but non-negligible proportion of RNNs do reach partial compositional solutions even without special architectural constraints. This suggests that a combination of gradient descent and evolutionary strategies directly favouring the minority models that developed more compositional approaches might suffice to lead standard RNNs towards compositional solutions.

Darwiche and Pearl's seminal 1997 article outlined a number of baseline principles for a logic of iterated belief revision. These principles, the DP postulates, have been supplemented in a number of alternative ways. Most of the suggestions made have resulted in a form of `reductionism' that identifies belief states with orderings of worlds. However, this position has recently been criticised as being unacceptably strong. Other proposals, such as the popular principle (P), aka `Independence', characteristic of `admissible' revision operators, remain commendably more modest. In this paper, we supplement both the DP postulates and (P) with a number of novel conditions. While the DP postulates constrain the relation between a prior and a posterior conditional belief set, our new principles notably govern the relation between two posterior conditional belief sets obtained from a common prior by different revisions. We show that operators from the resulting family, which subsumes both lexicographic and restrained revision, can be represented as relating belief states that are associated with a `proper ordinal interval' (POI) assignment, a structure more fine-grained than a simple ordering of worlds. We close the paper by noting that these operators satisfy iterated versions of a large number of AGM era postulates, including Superexpansion, that are not sound for admissible operators in general.

This paper provides a framework for evaluating creativity in co-creative systems: those that involve computer programs collaborating with human users on creative tasks. We situate co-creative systems within a broader context of computational creativity and explain the unique qualities of these systems. We present four main questions that can guide evaluation in co-creative systems: Who is evaluating the creativity, what is being evaluated, when does evaluation occur and how the evaluation is performed. These questions provide a framework for comparing how existing co-creative systems evaluate creativity, and we apply them to examples of co-creative systems in art, humor, games and robotics. We conclude that existing co-creative systems tend to focus on evaluating the user experience. Adopting evaluation methods from autonomous creative systems may lead to co-creative systems that are self-aware and intentional.

Probabilistic representations, such as Bayesian and Markov networks, are fundamental to much of statistical machine learning. Thus, learning probabilistic representations directly from data is a deep challenge, the main computational bottleneck being inference that is intractable. Tractable learning is a powerful new paradigm that attempts to learn distributions that support efficient probabilistic querying. By leveraging local structure, representations such as sum-product networks (SPNs) can capture high tree-width models with many hidden layers, essentially a deep architecture, while still admitting a range of probabilistic queries to be computable in time polynomial in the network size. The leaf nodes in SPNs, from which more intricate mixtures are formed, are tractable univariate distributions, and so the literature has focused on Bernoulli and Gaussian random variables. This is clearly a restriction for handling mixed discrete-continuous data, especially if the continuous features are generated from non-parametric and non-Gaussian distribution families. In this work, we present a framework that systematically integrates SPN structure learning with weighted model integration, a recently introduced computational abstraction for performing inference in hybrid domains, by means of piecewise polynomial approximations of density functions of arbitrary shape. Our framework is instantiated by exploiting the notion of propositional abstractions, thus minimally interfering with the SPN structure learning module, and supports a powerful query interface for conditioning on interval constraints. Our empirical results show that our approach is effective, and allows a study of the trade off between the granularity of the learned model and its predictive power.

KIM LANDERS: From self-driving cars to facial recognition new technology powered by artificial intelligence is changing the way we live and work and how we make decisions. But what is this rapid rise of new technology doing to our Human Rights? That's the question a new project from the Human Rights Commission is going to tackle. It's trying to identify the issues at stake before coming up with a final report by late next year. Edward Santow is the Human Rights Commissioner.