Submodular functions are a broad class of set functions, which naturally arise in diverse areas. Many algorithms have been suggested for the maximization of these functions. Unfortunately, once the function deviates from submodularity, the known algorithms may perform arbitrarily poorly. Amending this issue, by obtaining approximation results for set functions generalizing submodular functions, has been the focus of recent works. One such class, known as weakly submodular functions, has received a lot of attention. A key result proved by Das and Kempe (2011) showed that the approximation ratio of the greedy algorithm for weakly submodular maximization subject to a cardinality constraint degrades smoothly with the distance from submodularity. However, no results have been obtained for maximization subject to constraints beyond cardinality. In particular, it is not known whether the greedy algorithm achieves any non-trivial approximation ratio for such constraints. In this paper, we prove that a randomized version of the greedy algorithm (previously used by Buchbinder et al. (2014) for a different problem) achieves an approximation ratio of $(1 + 1/\gamma)^{-2}$ for the maximization of a weakly submodular function subject to a general matroid constraint, where $\gamma$ is a parameter measuring the distance of the function from submodularity. Moreover, we also experimentally compare the performance of this version of the greedy algorithm on real world problems against natural benchmarks, and show that the algorithm we study performs well also in practice. To the best of our knowledge, this is the first algorithm with a non-trivial approximation guarantee for maximizing a weakly submodular function subject to a constraint other than the simple cardinality constraint. In particular, it is the first algorithm with such a guarantee for the important and broad class of matroid constraints.

The recent success of neural networks in applications such as speech recognition, vision and autonomous navigation has led to great excitement by members of the artificial intelligence (AI) community and the general public at large. Over a relatively short period, by the science clock, we managed to automate some tasks that have defied us for decades and using one of the more classical techniques coming out of artificial intelligence research. The triumph over these achievements has led some to describe the automation of these tasks as having reached human level intelligence. This perception, originally hinted at in academic circles, has been gaining momentum more broadly and is leading to some implications. For example, a trend is emerging in which machine learning research is being streamlined into neural network research, under its newly acquired label "deep learning." This perception has also caused some to question the wisdom of continuing to invest in other machine learning approaches, or even mainstream areas of artificial intelligence, such as knowledge representation, symbolic reasoning and planning. Some coverage of AI in public arenas, particularly comments made by some visible figures, has led to mixing this excitement with fear of what AI may be bringing us in the future (i.e., doomsday scenarios).

We consider the problem of deep neural net compression by quantization: given a large, reference net, we want to quantize its real-valued weights using a codebook with $K$ entries so that the training loss of the quantized net is minimal. The codebook can be optimally learned jointly with the net, or fixed, as for binarization or ternarization approaches. Previous work has quantized the weights of the reference net, or incorporated rounding operations in the backpropagation algorithm, but this has no guarantee of converging to a loss-optimal, quantized net. We describe a new approach based on the recently proposed framework of model compression as constrained optimization \citep{Carreir17a}. This results in a simple iterative "learning-compression" algorithm, which alternates a step that learns a net of continuous weights with a step that quantizes (or binarizes/ternarizes) the weights, and is guaranteed to converge to local optimum of the loss for quantized nets. We develop algorithms for an adaptive codebook or a (partially) fixed codebook. The latter includes binarization, ternarization, powers-of-two and other important particular cases. We show experimentally that we can achieve much higher compression rates than previous quantization work (even using just 1 bit per weight) with negligible loss degradation.

We present the first general purpose framework for marginal maximum a posteriori estimation of probabilistic program variables. By using a series of code transformations, the evidence of any probabilistic program, and therefore of any graphical model, can be optimized with respect to an arbitrary subset of its sampled variables. To carry out this optimization, we develop the first Bayesian optimization package to directly exploit the source code of its target, leading to innovations in problem-independent hyperpriors, unbounded optimization, and implicit constraint satisfaction; delivering significant performance improvements over prominent existing packages.

Kernel adaptive filters, a class of adaptive nonlinear time-series models, are known by their ability to learn expressive autoregressive patterns from sequential data. However, for trivial monotonic signals, they struggle to perform accurate predictions and at the same time keep computational complexity within desired boundaries. This is because new observations are incorporated to the dictionary when they are far from what the algorithm has seen in the past. We propose a novel approach to kernel adaptive filtering that compares new observations against dictionary samples in terms of their unit-norm (normalised) versions, meaning that new observations that look like previous samples but have a different magnitude are not added to the dictionary. We achieve this by proposing the unit-norm Gaussian kernel and define a sparsification criterion for this novel kernel. This new methodology is validated on two real-world datasets against standard KAF in terms of the normalised mean square error and the dictionary size.

Representing a word by its co-occurrences with other words in context is an effective way to capture the meaning of the word. However, the theory behind remains a challenge. In this work, taking the example of a word classification task, we give a theoretical analysis of the approaches that represent a word X by a function f(P(C X)), where C is a context feature, P(C X) is the conditional probability estimated from a text corpus, and the function f maps the co-occurrence measure to a prediction score. We investigate the impact of context feature C and the function f. We also explain the reasons why using the co-occurrences with multiple context features may be better than just using a single one. In addition, some of the results shed light on the theory of feature learning and machine learning in general.

Most deep reinforcement learning algorithms are data inefficient in complex and rich environments, limiting their applicability to many scenarios. One direction for improving data efficiency is multitask learning with shared neural network parameters, where efficiency may be improved through transfer across related tasks. In practice, however, this is not usually observed, because gradients from different tasks can interfere negatively, making learning unstable and sometimes even less data efficient. Another issue is the different reward schemes between tasks, which can easily lead to one task dominating the learning of a shared model. We propose a new approach for joint training of multiple tasks, which we refer to as Distral (Distill & transfer learning). Instead of sharing parameters between the different workers, we propose to share a "distilled" policy that captures common behaviour across tasks. Each worker is trained to solve its own task while constrained to stay close to the shared policy, while the shared policy is trained by distillation to be the centroid of all task policies. Both aspects of the learning process are derived by optimizing a joint objective function. We show that our approach supports efficient transfer on complex 3D environments, outperforming several related methods. Moreover, the proposed learning process is more robust and more stable---attributes that are critical in deep reinforcement learning.

This paper presents an approach for automation of interpretable feature selection for Internet Of Things Analytics (IoTA) using machine learning (ML) techniques. Authors have conducted a survey over different people involved in different IoTA based application development tasks. The survey reveals that feature selection is the most time consuming and niche skill demanding part of the entire workflow. This paper shows how feature selection is successfully automated without sacrificing the decision making accuracy and thereby reducing the project completion time and cost of hiring expensive resources. Several pattern recognition principles and state of art (SoA) ML techniques are followed to design the overall approach for the proposed automation. Three data sets are considered to establish the proof-of-concept. Experimental results show that the proposed automation is able to reduce the time for feature selection to $2$ days instead of $4-6$ months which would have been required in absence of the automation. This reduction in time is achieved without any sacrifice in the accuracy of the decision making process. Proposed method is also compared against Multi Layer Perceptron (MLP) model as most of the state of the art works on IoTA uses MLP based Deep Learning. Moreover the feature selection method is compared against SoA feature reduction technique namely Principal Component Analysis (PCA) and its variants. The results obtained show that the proposed method is effective.

The aim of semi-supervised learning is to improve supervised learners by exploiting potentially large amounts of, typically easier to obtain, unlabeled data [1]. Up to now, however, semi-supervised learners have reported mixed results when it comes to such improvements: it is not always the case that semi-supervision results in lower expected error rates. On the contrary, severely deteriorated performances have been observed in empirical studies and theory shows that improvement guarantees can often only be provided under rather stringent conditions [2-5]. Now, the principal suggestion put forward in this chapter is that, when dealing with semi-supervised learning, one may not only want to study the (expected) error rates these classifiers produce, but also to measure the classifiers' performances by means of the intrinsic loss they may be optimizing in the first place. That is, for classification routines that optimize a so-called surrogate loss at training time--which is what many machine learning and Bayesian decision theoretic approaches do [6, 7], we propose to also investigate how this loss behaves on the test set as this can provide us with an alternative view on the classifier's behavior that a mere error rate cannot capture. In fact, though the main example is concerned with semi-supervision, we would like to argue that in other learning scenarios, similar considerations might be beneficial. For instance in active learning [8], where rather than sampling randomly from ones input data to provide these instances with labels, one aims to do the sampling in a systematic way, trying to keep labeling cost as low as one can or, similarly, to learn from as few labeled examples as possible. Also here it may (or, we believe, it should) be of interest to not only compare the error rates that different approaches (e.g.

This paper proposes a novel approach for constructing effective personalized policies when the observed data lacks counter-factual information, is biased and possesses many features. The approach is applicable in a wide variety of settings from healthcare to advertising to education to finance. These settings have in common that the decision maker can observe, for each previous instance, an array of features of the instance, the action taken in that instance, and the reward realized -- but not the rewards of actions that were not taken: the counterfactual information. Learning in such settings is made even more difficult because the observed data is typically biased by the existing policy (that generated the data) and because the array of features that might affect the reward in a particular instance -- and hence should be taken into account in deciding on an action in each particular instance -- is often vast. The approach presented here estimates propensity scores for the observed data, infers counterfactuals, identifies a (relatively small) number of features that are (most) relevant for each possible action and instance, and prescribes a policy to be followed. Comparison of the proposed algorithm against the state-of-art algorithm on actual datasets demonstrates that the proposed algorithm achieves a significant improvement in performance.