Transfer of pre-trained representations can improve sample efficiency and reduce computational requirements for new tasks. However, representations used for transfer are usually generic, and are not tailored to a particular distribution of downstream tasks. We explore the use of expert representations for transfer with a simple, yet effective, strategy. We train a diverse set of experts by exploiting existing label structures, and use cheap-to-compute performance proxies to select the relevant expert for each target task. This strategy scales the process of transferring to new tasks, since it does not revisit the pre-training data during transfer. Accordingly, it requires little extra compute per target task, and results in a speed-up of 2-3 orders of magnitude compared to competing approaches. Further, we provide an adapter-based architecture able to compress many experts into a single model. We evaluate our approach on two different data sources and demonstrate that it outperforms baselines on over 20 diverse vision tasks in both cases.

Recent studies have demonstrated that noise in stochastic gradient descent (SGD) is closely related to generalization: A larger SGD noise, if not too large, results in better generalization. Since the covariance of the SGD noise is proportional to $\eta^2/B$, where $\eta$ is the learning rate and $B$ is the minibatch size of SGD, the SGD noise has so far been controlled by changing $\eta$ and/or $B$. However, too large $\eta$ results in instability in the training dynamics and a small $B$ prevents scalable parallel computation. It is thus desirable to develop a method of controlling the SGD noise without changing $\eta$ and $B$. In this paper, we propose a method that achieves this goal using ``noise enhancement'', which is easily implemented in practice. We expound the underlying theoretical idea and demonstrate that the noise enhancement actually improves generalization for real datasets. It turns out that large-batch training with the noise enhancement even shows better generalization compared with small-batch training.

We consider training a binary classifier under delayed feedback (DF Learning). In DF Learning, we first receive negative samples; subsequently, some samples turn positive. This problem is conceivable in various real-world applications such as online advertisements, where the user action takes place long after the first click. Owing to the delayed feedback, simply separating the positive and negative data causes a sample selection bias. One solution is to assume that a long time window after first observing a sample reduces the sample selection bias. However, existing studies report that only using a portion of all samples based on the time window assumption yields suboptimal performance, and the use of all samples along with the time window assumption improves empirical performance. Extending these existing studies, we propose a method with an unbiased and convex empirical risk constructed from the whole samples under the time window assumption. We provide experimental results to demonstrate the effectiveness of the proposed method using a real traffic log dataset.

Decision forest algorithms model data by learning a binary tree structure recursively where every node splits the feature space into two regions, sending examples into the left or right branches. This "decision" is the result of the evaluation of a condition. For example, a node may split input data by applying a threshold to a numerical feature value. Such decisions are learned using (often greedy) algorithms that attempt to optimize a local loss function. Crucially, whether an algorithm exists to find and evaluate splits for a feature type (e.g., text) determines whether a decision forest algorithm can model that feature type at all. In this work, we set out to devise such an algorithm for textual features, thereby equipping decision forests with the ability to directly model text without the need for feature transformation. Our algorithm is efficient during training and the resulting splits are fast to evaluate with our extension of the QuickScorer inference algorithm. Experiments on benchmark text classification datasets demonstrate the utility and effectiveness of our proposal.

Diffusion Maps is a nonlinear dimensionality reduction technique used to embed high-dimensional data in a low-dimensional Euclidean space, where the notion of distance is due to the transition probability of a random walk over the dataset. However, the conventional approach is not capable to reveal the dataset underlying subspace structure, a useful information for machine learning applications such as object classification and face recognition. To circumvent this limitation, a novel nonlinear dimensionality reduction technique, referred to as Grassmannian Diffusion Maps, is developed herein relying on the affinity between subspaces represented by points on the Grassmann manifold. To this aim, a kernel matrix is used to construct the transition matrix of a random walk on a graph connecting points on the Grassmann manifold for posterior determination of the diffusion coordinates embedding the data in a low-dimensional Euclidean space. In this paper, three examples are considered to evaluate the performance of both conventional and Grassmannian Diffusion Maps. First, a "toy" example shows that the Grassmannian Diffusion Maps can identify a well-defined parametrization of points on the unit sphere, representing a Grassmann manifold. The second example shows that the Grassmannian Diffusion Maps outperforms the conventional Diffusion Maps in classifying elements, later recovered by a conventional clustering, of a dataset by their intrinsic characteristics. In the last example, a novel data classification/recognition technique is developed based on the construction of an overcomplete dictionary of reduced dimension whose atoms are given by the diffusion coordinates. A face recognition problem is solved and high recognition rates (i.e., 95% in the best-case scenario) are obtained using a fraction of the data required by conventional methods.

Source code representations are key in applying machine learning techniques for processing and analyzing programs. A popular approach in representing source code is neural source code embeddings that represents programs with high-dimensional vectors computed by training deep neural networks on a large volume of programs. Although successful, there is little known about the contents of these vectors and their characteristics. In this paper, we present our preliminary results towards better understanding the contents of code2vec neural source code embeddings. In particular, in a small case study, we use the code2vec embeddings to create binary SVM classifiers and compare their performance with the handcrafted features. Our results suggest that the handcrafted features can perform very close to the highly-dimensional code2vec embeddings, and the information gains are more evenly distributed in the code2vec embeddings compared to the handcrafted features. We also find that the code2vec embeddings are more resilient to the removal of dimensions with low information gains than the handcrafted features. We hope our results serve a stepping stone toward principled analysis and evaluation of these code representations.

The VAT method is a visual technique for determining the potential cluster structure and the possible number of clusters in numerical data. Its improved version, iVAT, uses a path-based distance transform to improve the effectiveness of VAT for "tough" cases. Both VAT and iVAT have also been used in conjunction with a single-linkage(SL) hierarchical clustering algorithm. However, they are sensitive to noise and bridge points between clusters in the dataset, and consequently, the corresponding VAT/iVAT images are often in-conclusive for such cases. In this paper, we propose a constraint-based version of iVAT, which we call ConiVAT, that makes use of background knowledge in the form of constraints, to improve VAT/iVAT for challenging and complex datasets. ConiVAT uses the input constraints to learn the underlying similarity metric and builds a minimum transitive dissimilarity matrix, before applying VAT to it. We demonstrate ConiVAT approach to visual assessment and single linkage clustering on nine datasets to show that, it improves the quality of iVAT images for complex datasets, and it also overcomes the limitation of SL clustering with VAT/iVAT due to "noisy" bridges between clusters. Extensive experiment results on nine datasets suggest that ConiVAT outperforms the other three semi-supervised clustering algorithms in terms of improved clustering accuracy.

In this paper, we consider mixtures of multinomial logistic models (MNL), which are known to $\epsilon$-approximate any random utility model. Despite its long history and broad use, rigorous results are only available for learning a uniform mixture of two MNLs. Continuing this line of research, we study the problem of learning an arbitrary mixture of two MNLs. We show that the identifiability of the mixture models may only fail on an algebraic variety of a negligible measure. This is done by reducing the problem of learning a mixture of two MNLs to the problem of solving a system of univariate quartic equations. We also devise an algorithm to learn any mixture of two MNLs using a polynomial number of samples and a linear number of queries, provided that a mixture of two MNLs over some finite universe is identifiable. Several numerical experiments and conjectures are also presented.

Similarity analysis is one of the crucial steps in most fMRI studies. Representational Similarity Analysis (RSA) can measure similarities of neural signatures generated by different cognitive states. This paper develops Deep Representational Similarity Learning (DRSL), a deep extension of RSA that is appropriate for analyzing similarities between various cognitive tasks in fMRI datasets with a large number of subjects, and high-dimensionality -- such as whole-brain images. Unlike the previous methods, DRSL is not limited by a linear transformation or a restricted fixed nonlinear kernel function -- such as Gaussian kernel. DRSL utilizes a multi-layer neural network for mapping neural responses to linear space, where this network can implement a customized nonlinear transformation for each subject separately. Furthermore, utilizing a gradient-based optimization in DRSL can significantly reduce runtime of analysis on large datasets because it uses a batch of samples in each iteration rather than all neural responses to find an optimal solution. Empirical studies on multi-subject fMRI datasets with various tasks -- including visual stimuli, decision making, flavor, and working memory -- confirm that the proposed method achieves superior performance to other state-of-the-art RSA algorithms.

Undertaking causal inference with observational data is extremely useful across a wide range of domains including the development of medical treatments, advertisements and marketing, and policy making. There are two main challenges associated with undertaking causal inference using observational data: treatment assignment heterogeneity (i.e., differences between the treated and untreated groups), and an absence of counterfactual data (i.e. not knowing what would have happened if an individual who did get treatment, were instead to have not been treated). We address these two challenges by combining structured inference and targeted learning. To our knowledge, Targeted Variational AutoEncoder (TVAE) is the first method to incorporate targeted learning into deep latent variable models. Results demonstrate competitive and state of the art performance.