The area of automatic image caption evaluation is still undergoing intensive research to address the needs of generating captions which can meet adequacy and fluency requirements. Based on our past attempts at developing highly sophisticated learning-based metrics, we have discovered that a simple cosine similarity measure using the Mean of Word Embeddings(MOWE) of captions can actually achieve a surprisingly high performance on unsupervised caption evaluation. This inspires our proposed work on an effective metric WEmbSim, which beats complex measures such as SPICE, CIDEr and WMD at system-level correlation with human judgments. Moreover, it also achieves the best accuracy at matching human consensus scores for caption pairs, against commonly used unsupervised methods. Therefore, we believe that WEmbSim sets a new baseline for any complex metric to be justified.

Automatic evaluation metrics hold a fundamental importance in the development and fine-grained analysis of captioning systems. While current evaluation metrics tend to achieve an acceptable correlation with human judgements at the system level, they fail to do so at the caption level. In this work, we propose a neural network-based learned metric to improve the caption-level caption evaluation. To get a deeper insight into the parameters which impact a learned metrics performance, this paper investigates the relationship between different linguistic features and the caption-level correlation of the learned metrics. We also compare metrics trained with different training examples to measure the variations in their evaluation. Moreover, we perform a robustness analysis, which highlights the sensitivity of learned and handcrafted metrics to various sentence perturbations. Our empirical analysis shows that our proposed metric not only outperforms the existing metrics in terms of caption-level correlation but it also shows a strong system-level correlation against human assessments.

Existing Image Captioning (IC) systems model words as atomic units in captions and are unable to exploit the structural information in the words. This makes representation of rare words very difficult and out-of-vocabulary words impossible. Moreover, to avoid computational complexity, existing IC models operate over a modest sized vocabulary of frequent words, such that the identity of rare words is lost. In this work we address this common limitation of IC systems in dealing with rare words in the corpora. We decompose words into smaller constituent units 'subwords' and represent captions as a sequence of subwords instead of words. This helps represent all words in the corpora using a significantly lower subword vocabulary, leading to better parameter learning. Using subword language modeling, our captioning system improves various metric scores, with a training vocabulary size approximately 90% less than the baseline and various state-of-the-art word-level models. Our quantitative and qualitative results and analysis signify the efficacy of our proposed approach.

Training deep neural networks (DNNs) with noisy labels is a challenging problem due to over-parameterization. DNNs tend to essentially fit on clean samples at a higher rate in the initial stages, and later fit on the noisy samples at a relatively lower rate. Thus, with a noisy dataset, the test accuracy increases initially and drops in the later stages. To find an early stopping point at the maximum obtainable test accuracy (MOTA), recent studies assume either that i) a clean validation set is available or ii) the noise ratio is known, or, both. However, often a clean validation set is unavailable, and the noise estimation can be inaccurate. To overcome these issues, we provide a novel training solution, free of these conditions. We analyze the rate of change of the training accuracy for different noise ratios under different conditions to identify a training stop region. We further develop a heuristic algorithm based on a small-learning assumption to find a training stop point (TSP) at or close to MOTA. To the best of our knowledge, our method is the first to rely solely on the \textit{training behavior}, while utilizing the entire training set, to automatically find a TSP. We validated the robustness of our algorithm (AutoTSP) through several experiments on CIFAR-10, CIFAR-100, and a real-world noisy dataset for different noise ratios, noise types and architectures.

We describe the convex semi-infinite dual of the two-layer vector-output ReLU neural network training problem. This semi-infinite dual admits a finite dimensional representation, but its support is over a convex set which is difficult to characterize. In particular, we demonstrate that the non-convex neural network training problem is equivalent to a finite-dimensional convex copositive program. Our work is the first to identify this strong connection between the global optima of neural networks and those of copositive programs. We thus demonstrate how neural networks implicitly attempt to solve copositive programs via semi-nonnegative matrix factorization, and draw key insights from this formulation. We describe the first algorithms for provably finding the global minimum of the vector output neural network training problem, which are polynomial in the number of samples for a fixed data rank, yet exponential in the dimension. However, in the case of convolutional architectures, the computational complexity is exponential in only the filter size and polynomial in all other parameters. We describe the circumstances in which we can find the global optimum of this neural network training problem exactly with soft-thresholded SVD, and provide a copositive relaxation which is guaranteed to be exact for certain classes of problems, and which corresponds with the solution of Stochastic Gradient Descent in practice.

Background: Drug-drug interactions (DDIs) refer to processes triggered by the administration of two or more drugs leading to side effects beyond those observed when drugs are administered by themselves. Due to the massive number of possible drug pairs, it is nearly impossible to experimentally test all combinations and discover previously unobserved side effects. Therefore, machine learning based methods are being used to address this issue. Methods: We propose a Siamese self-attention multi-modal neural network for DDI prediction that integrates multiple drug similarity measures that have been derived from a comparison of drug characteristics including drug targets, pathways and gene expression profiles. Results: Our proposed DDI prediction model provides multiple advantages: 1) It is trained end-to-end, overcoming limitations of models composed of multiple separate steps, 2) it offers model explainability via an Attention mechanism for identifying salient input features and 3) it achieves similar or better prediction performance (AUPR scores ranging from 0.77 to 0.92) compared to state-of-the-art DDI models when tested on various benchmark datasets. Novel DDI predictions are further validated using independent data resources. Conclusions: We find that a Siamese multi-modal neural network is able to accurately predict DDIs and that an Attention mechanism, typically used in the Natural Language Processing domain, can be beneficially applied to aid in DDI model explainability.

Efficient sampling of complex data distributions can be achieved using trained invertible flows (IF), where the model distribution is generated by pushing a simple base distribution through multiple non-linear bijective transformations. However, the iterative nature of the transformations in IFs can limit the approximation to the target distribution. In this paper we seek to mitigate this by implementing RBM-Flow, an IF model whose base distribution is a Restricted Boltzmann Machine (RBM) with a continuous smoothing applied. We show that by using RBM-Flow we are able to improve the quality of samples generated, quantified by the Inception Scores (IS) and Frechet Inception Distance (FID), over baseline models with the same IF transformations, but with less expressive base distributions. Furthermore, we also obtain D-Flow, an IF model with uncorrelated discrete latent variables. We show that D-Flow achieves similar likelihoods and FID/IS scores to those of a typical IF with Gaussian base variables, but with the additional benefit that global features are meaningfully encoded as discrete labels in the latent space.

In 2016, Alphabet's DeepMind came out with AlphaGo, an AI which consistently beat the best human Go players. One year later, the subsidiary went on to refine its work, creating AlphaGo Zero. Where its predecessor learned to play Go by observing amateur and professional matches, AlphaGo Zero mastered the ancient game by simply playing against itself. DeepMind then created AlphaZero, which could play Go, chess and shogi with a single algorithm. What tied all those AIs together is that they knew the rules of the games they had to master going into their training.

The successor to AlphaGo is being used to create a more efficient type of video compression.

What magical trick makes us intelligent? The trick is that there is no trick. The power of intelligence stems from our vast diversity, not from any single, perfect principle. Artificial intelligence has recently beaten world champions in Go and poker and made extraordinary progress in domains such as machine translation, object classification, and speech recognition. However, most AI systems are extremely narrowly focused. AlphaGo, the champion Go player, does not know that the game is played by putting stones onto a board; it has no idea what a "stone" or a "board" is, and would need to be retrained from scratch if you presented it with a rectangular board rather than a square grid.