Twin-to-twin transfusion syndrome treatment requires fetoscopic laser photocoagulation of placental vascular anastomoses to regulate blood flow to both fetuses. Limited field-of-view (FoV) and low visual quality during fetoscopy make it challenging to identify all vascular connections. Mosaicking can align multiple overlapping images to generate an image with increased FoV, however, existing techniques apply poorly to fetoscopy due to the low visual quality, texture paucity, and hence fail in longer sequences due to the drift accumulated over time. Deep learning techniques can facilitate in overcoming these challenges. Therefore, we present a new generalized Deep Sequential Mosaicking (DSM) framework for fetoscopic videos captured from different settings such as simulation, phantom, and real environments. DSM extends an existing deep image-based homography model to sequential data by proposing controlled data augmentation and outlier rejection methods. Unlike existing methods, DSM can handle visual variations due to specular highlights and reflection across adjacent frames, hence reducing the accumulated drift. We perform experimental validation and comparison using 5 diverse fetoscopic videos to demonstrate the robustness of our framework.

We can define a neural network that can learn to recognize objects in less than 100 lines of code. However, after training, it is characterized by millions of weights that contain the knowledge about many object types across visual scenes. Such networks are thus dramatically easier to understand in terms of the code that makes them than the resulting properties, such as tuning or connections. In analogy, we conjecture that rules for development and learning in brains may be far easier to understand than their resulting properties. The analogy suggests that neuroscience would benefit from a focus on learning and development.

Parameterized state space models in the form of recurrent networks are often used in machine learning to learn from data streams exhibiting temporal dependencies. To break the black box nature of such models it is important to understand the dynamical features of the input driving time series that are formed in the state space. We propose a framework for rigorous analysis of such state representations in vanishing memory state space models such as echo state networks (ESN). In particular, we consider the state space a temporal feature space and the readout mapping from the state space a kernel machine operating in that feature space. We show that: (1) The usual ESN strategy of randomly generating input-to-state, as well as state coupling leads to shallow memory time series representations, corresponding to cross-correlation operator with fast exponentially decaying coefficients; (2) Imposing symmetry on dynamic coupling yields a constrained dynamic kernel matching the input time series with straightforward exponentially decaying motifs or exponentially decaying motifs of the highest frequency; (3) Simple cycle high-dimensional reservoir topology specified only through two free parameters can implement deep memory dynamic kernels with a rich variety of matching motifs. We quantify richness of feature representations imposed by dynamic kernels and demonstrate that for dynamic kernel associated with cycle reservoir topology, the kernel richness undergoes a phase transition close to the edge of stability.

Training data-driven approaches for complex industrial system health monitoring is challenging. When data on faulty conditions are rare or not available, the training has to be performed in a unsupervised manner. In addition, when the observation period, used for training, is kept short, to be able to monitor the system in its early life, the training data might not be representative of all the system normal operating conditions. In this paper, we propose five approaches to perform fault detection in such context. Two approaches rely on the data from the unit to be monitored only: the baseline is trained on the early life of the unit. An incremental learning procedure tries to learn new operating conditions as they arise. Three other approaches take advantage of data from other similar units within a fleet. In two cases, units are directly compared to each other with similarity measures, and the data from similar units are combined in the training set. We propose, in the third case, a new deep-learning methodology to perform, first, a feature alignment of different units with an Unsupervised Feature Alignment Network (UFAN). Then, features of both units are combined in the training set of the fault detection neural network. The approaches are tested on a fleet comprising 112 units, observed over one year of data. All approaches proposed here are an improvement to the baseline, trained with two months of data only. As units in the fleet are found to be very dissimilar, the new architecture UFAN, that aligns units in the feature space, is outperforming others.

The aim of multi-task reinforcement learning is two-fold: (1) efficiently learn by training against multiple tasks and (2) quickly adapt, using limited samples, to a variety of new tasks. In this work, the tasks correspond to reward functions for environments with the same (or similar) dynamical models. We propose to learn a dynamical model during the training process and use this model to perform sample-efficient adaptation to new tasks at test time. We use significantly fewer samples by performing policy optimization only in a "virtual" environment whose transitions are given by our learned dynamical model. Our algorithm sequentially trains against several tasks. Upon encountering a new task, we first warm-up a policy on our learned dynamical model, which requires no new samples from the environment. We then adapt the dynamical model with samples from this policy in the real environment. We evaluate our approach on several continuous control benchmarks and demonstrate its efficacy over MAML, a state-of-the-art meta-learning algorithm, on these tasks.

The goal of this paper is to develop a machine learning based approach that utilizes phase space alone to separate the Gaia DR2 stars into two categories: those accreted onto the Milky Way from in situ stars that were born within the Galaxy. Traditional selection methods that have been used to identify accreted stars typically rely on full 3D velocity and/or metallicity information, which significantly reduces the number of classifiable stars. The approach advocated here is applicable to a much larger fraction of Gaia DR2. A method known as transfer learning is shown to be effective through extensive testing on a set of mock Gaia catalogs that are based on the FIRE cosmological zoom-in hydrodynamic simulations of Milky Way-mass galaxies. The machine is first trained on simulated data using only 5D kinematics as inputs, and is then further trained on a cross-matched Gaia/RAVE data set, which improves sensitivity to properties of the real Milky Way. The result is a catalog that identifies ~650,000 accreted stars within Gaia DR2. This catalog can yield empirical insights into the merger history of the Milky Way, and could be used to infer properties of the dark matter distribution.

In data mining, machine learning, and other areas of AI, we are often faced with datasets that contain many more attributes than needed, or that can even be helpful for tasks such as clustering or classification. Problems associated with such high dimensional data are for example the concentration effect of distances [13, 20] or irrelevant features [25, 49]. For clustering [31] and outlier detection [49], researchers have made use of various techniques to identify relevant subspaces, as defined by subsets of features that are informative for a particular task. Examples of how relevant subspaces can be determined for individual clusters or outliers include local density estimation in a systematic search through candidate subspaces (often following the Apriori principle [7] in various adaptations to the subspace search problem [48]), or the adaptation of distance measures based on the distribution within local neighborhoods (using some analysis of variance or even covariance -- typically based on PCA -- to allow also for an adaptation to correlated features). For sufficiently tight local neighborhoods, the underlying local data manifold can be regarded as approaching a linear form [40], an assumption that further justifies the determination of locally relevant features for subspace determination.

Recently deep learning-based methods have been applied in image compression and achieved many promising results. In this paper, we propose an improved hybrid layered image compression framework by combining deep learning and the traditional image codecs. At the encoder, we first use a convolutional neural network (CNN) to obtain a compact representation of the input image, which is losslessly encoded by the FLIF codec as the base layer of the bit stream. A coarse reconstruction of the input is obtained by another CNN from the reconstructed compact representation. The residual between the input and the coarse reconstruction is then obtained and encoded by the H.265/HEVC-based BPG codec as the enhancement layer of the bit stream. Experimental results using the Kodak and Tecnick datasets show that the proposed scheme outperforms the state-of-the-art deep learning-based layered coding scheme and traditional codecs including BPG in both PSNR and MS-SSIM metrics across a wide range of bit rates, when the images are coded in the RGB444 domain.

One of the challenges in display advertising is that the distribution of features and click through rate (CTR) can exhibit large shifts over time due to seasonality, changes to ad campaigns and other factors. The predominant strategy to keep up with these shifts is to train predictive models continuously, on fresh data, in order to prevent them from becoming stale. However, in many ad systems positive labels are only observed after a possibly long and random delay. These delayed labels pose a challenge to data freshness in continuous training: fresh data may not have complete label information at the time they are ingested by the training algorithm. Naive strategies which consider any data point a negative example until a positive label becomes available tend to underestimate CTR, resulting in inferior user experience and suboptimal performance for advertisers. The focus of this paper is to identify the best combination of loss functions and models that enable large-scale learning from a continuous stream of data in the presence of delayed labels. In this work, we compare 5 different loss functions, 3 of them applied to this problem for the first time. We benchmark their performance in offline settings on both public and proprietary datasets in conjunction with shallow and deep model architectures. We also discuss the engineering cost associated with implementing each loss function in a production environment. Finally, we carried out online experiments with the top performing methods, in order to validate their performance in a continuous training scheme. While training on 668 million in-house data points offline, our proposed methods outperform previous state-of-the-art by 3% relative cross entropy (RCE). During online experiments, we observed 55% gain in revenue per thousand requests (RPMq) against naive log loss.

One of the main, long-term objectives of artificial intelligence is the creation of thinking machines. To that end, substantial effort has been placed into designing cognitive systems; i.e. systems that can manipulate semantic-level information. A substantial part of that effort is oriented towards designing the mathematical machinery underlying cognition in a way that is very efficiently implementable in hardware. In this work we propose a 'semi-holographic' representation system that can be implemented in hardware using only multiplexing and addition operations, thus avoiding the need for expensive multiplication. The resulting architecture can be readily constructed by recycling standard microprocessor elements and is capable of performing two key mathematical operations frequently used in cognition, superposition and binding, within a budget of below 6 pJ for 64- bit operands. Our proposed 'cognitive processing unit' (CoPU) is intended as just one (albeit crucial) part of much larger cognitive systems where artificial neural networks of all kinds and associative memories work in concord to give rise to intelligence.