ImageNet pre-training has been regarded as essential for training accurate object detectors for a long time. Recently, it has been shown that object detectors trained from randomly initialized weights can be on par with those fine-tuned from ImageNet pre-trained models. However, the effects of pre-training and the differences caused by pre-training are still not fully understood. In this paper, we analyze the eigenspectrum dynamics of the covariance matrix of each feature map in object detectors. Based on our analysis on ResNet-50, Faster R-CNN with FPN, and Mask R-CNN, we show that object detectors trained from ImageNet pre-trained models and those trained from scratch behave differently from each other even if both object detectors have similar accuracy. Furthermore, we propose a method for automatically determining the widths (the numbers of channels) of object detectors based on the eigenspectrum. We train Faster R-CNN with FPN from randomly initialized weights, and show that our method can reduce ~27% of the parameters of ResNet-50 without increasing Multiply-Accumulate operations and losing accuracy. Our results indicate that we should develop more appropriate methods for transferring knowledge from image classification to object detection (or other tasks).

Hyperparameter optimization is both a practical issue and an interesting theoretical problem in training of deep architectures. Despite many recent advances the most commonly used methods almost universally involve training multiple and decoupled copies of the model, in effect sampling the hyperparameter space. We show that at a negligible additional computational cost, results can be improved by sampling nonlocal paths instead of points in hyperparameter space. To this end we interpret hyperparameters as controlling the level of correlated noise in training, which can be mapped to an effective temperature. The usually independent instances of the model are coupled and allowed to exchange their hyperparameters throughout the training using the well established parallel tempering technique of statistical physics. Each simulation corresponds then to a unique path, or history, in the joint hyperparameter/model-parameter space. We provide empirical tests of our method, in particular for dropout and learning rate optimization. We observed faster training and improved resistance to overfitting and showed a systematic decrease in the absolute validation error, improving over benchmark results.

Context-aware recommender systems (CARSs) apply sensing and analysis of user context in order to provide personalized services. Adding context to a recommendation model is challenging, since the addition of context may increases both the dimensionality and sparsity of the model. Recent research has shown that modeling contextual information as a latent vector may address the sparsity and dimensionality challenges. We suggest a new latent modeling of sequential context by generating sequences of contextual information and reducing their contextual space to a compressed latent space.We train a long short-term memory (LSTM) encoder-decoder network on sequences of contextual information and extract sequential latent context from the hidden layer of the network in order to represent a compressed representation of sequential data. We propose new context-aware recommendation models that extend the neural collaborative filtering approach and learn nonlinear interactions between latent features of users, items, and contexts which take into account the sequential latent context representation as part of the recommendation process. We deployed our approach using two context-aware datasets with different context dimensions. Empirical analysis of our results validates that our proposed sequential latent context-aware model (SLCM), surpasses state of the art CARS models.

We propose a new approach to combine Restricted Boltzmann Machines (RBMs) that can be used to solve combinatorial optimization problems. This allows synthesis of larger models from smaller RBMs that have been pretrained, thus effectively bypassing the problem of learning in large RBMs, and creating a system able to model a large, complex multi-modal space. We validate this approach by using learned representations to create "invertible boolean logic", where we can use Markov chain Monte Carlo (MCMC) approaches to find the solution to large scale boolean satisfiability problems and show viability towards other combinatorial optimization problems. Using this method, we are able to solve 64 bit addition based problems, as well as factorize 16 bit numbers. We find that these combined representations can provide a more accurate result for the same sample size as compared to a fully trained model. The Ising Problem has long been known to be in the class of NP-Hard problems, with no exact polynomial solution existing. Because of this, a large class of combinatorial optimization problems can be reformulated as Ising problems and solved by finding the ground state of that system (Barahona, 1982; Kirkpatrick et al., 1983; Lucas, 2014). The Boltzmann Machine (Ackley et al., 1987) was originally introduced as a constraint satisfaction network based on the Ising model problem, where the weights would encode some global constraints, and stochastic units were used to escape local minima. The original Boltzmann Machine found favor as a method to solve various combinatorial optimization problems (Korst & Aarts, 1989). However, learning was very slow with this model due to the difficulties with sampling and convergence, as well as the inability to exactly calculate the partition function.

Driver identification has emerged as a vital research field, where both practitioners and researchers investigate the potential of driver identification to enable a personalized driving experience. Within recent years, a selection of studies have reported that individuals could be perfectly identified based on their driving behavior under controlled conditions. However, research investigating the potential of driver identification under naturalistic conditions claim accuracies only marginally higher than random guess. The paper at hand provides a comprehensive summary of the recent work, highlighting the main discrepancies in the design of the machine learning approaches, primarily the window length parameter that was considered. Key findings further indicate that the longitudinal vehicle control information is particularly useful for driver identification, leaving the research gap on the extent to which the lateral vehicle control can be used for reliable identification. Building upon existing work, we provide a novel approach for the design of the window length parameter that provides evidence that reliable driver identification can be achieved with data limited to the steering wheel only. The results and insights in this paper are based on data collected from the largest naturalistic driving study conducted in this field. Overall, a neural network based on GRUs was found to provide better identification performance than traditional methods, increasing the prediction accuracy from under 15\% to over 65\% for 15 drivers. When leveraging the full field study dataset, comprising 72 drivers, the accuracy of identification prediction of the approach improved a random guess approach by a factor of 25.

Real Time Trajectory Prediction Using Deep Conditional Generative Models Sebastian Gomez-Gonzalez 1, 2, Sergey Prokudin 1, Bernhard Sch olkopf 1 and Jan Peters 2 Abstract -- Data driven methods for time series forecasting that quantify uncertainty open new important possibilities for robot tasks with hard real time constraints, allowing the robot system to make decisions that trade off between reaction time and accuracy in the predictions. Despite the recent advances in deep learning, it is still challenging to make long term accurate predictions with the low latency required by real time robotic systems. In this paper, we propose a deep conditional generative model for trajectory prediction that is learned from a data set of collected trajectories. Our method uses an encoder and decoder deep networks that maps complete or partial trajectories to a Gaussian distributed latent space and back, allowing for fast inference of the future values of a trajectory given previous observations. The encoder and decoder networks are trained using stochastic gradient variational Bayes. In the experiments, we show that our model provides more accurate long term predictions with a lower latency that popular models for trajectory forecasting like recurrent neural networks or physical models based on differential equations. Finally, we test our proposed approach in a robot table tennis scenario to evaluate the performance of the proposed method in a robotic task with hard real time constraints. I. INTRODUCTION Dynamic high speed robotics tasks often require accurate methods to forecast the future value of a physical quantity based on previous measurements while respecting the real time constraints of the particular application.

We introduce a wide and deep neural network for prediction of progression from patients with mild cognitive impairment to Alzheimer's disease. Information from anatomical shape and tabular clinical data (demographics, biomarkers) are fused in a single neural network. The network is invariant to shape transformations and avoids the need to identify point correspondences between shapes. To account for right censored time-to-event data, i.e., when it is only known that a patient did not develop Alzheimer's disease up to a particular time point, we employ a loss commonly used in survival analysis. Our network is trained end-to-end to combine information from a patient's hippocampus shape and clinical biomarkers. Our experiments on data from the Alzheimer's Disease Neuroimaging Initiative demonstrate that our proposed model is able to learn a shape descriptor that augments clinical biomarkers and outperforms a deep neural network on shape alone and a linear model on common clinical biomarkers.

In this work we systematically analyze general properties of differential equations used as machine learning models. We demonstrate that the gradient of the loss function with respect to to the hidden state can be considered as a generalized momentum conjugate to the hidden state, allowing application of the tools of classical mechanics. In addition, we show that not only residual networks, but also feedforward neural networks with small nonlinearities and the weights matrices deviating only slightly from identity matrices can be related to the differential equations. We propose a differential equation describing such networks and investigate its properties. 1 Introduction Deep learning is a form of machine learning that uses neural networks with many hidden layers [1, 2]. Deep learning models have dramatically improved speech recognition, visual object recognition, object detection and many other domains [2]. Since the number of layers in deep neural networks become large, it is possible to consider the layer number as a continuous variable [3] and represent the neural network by an differential equation. The connection between the neural networks and differential equations first appeared with an additive model for continuous time recurrent neural networks, described by the differential equations [4] τ i dx i dt x i n null j 1w j,iσ ( x j θ j) I i(t) . Hopfield's work [6] pioneered the analog computation of continuous time recurrent neural networks instead of digital computation using complex numerical algorithms on a digital computer. A Hopfield network has a quadratic form as an Lyapunov function for the activity dynamics. As a consequence, the state of the network evolves to a final state that is a minimum of the Lyapunov function when started in any initial state [7].

Continual learning of deep neural networks is a key requirement for scaling them up to more complex applicative scenarios and for achieving real lifelong learning of these architectures. Previous approaches to the problem have considered either the progressive increase in the size of the networks, or have tried to regularize the network behavior to equalize it with respect to previously observed tasks. In the latter case, it is essential to understand what type of information best represents this past behavior. Common techniques include regularizing the past outputs, gradients, or individual weights. In this work, we propose a new, relatively simple and efficient method to perform continual learning by regularizing instead the network internal embeddings. To make the approach scalable, we also propose a dynamic sampling strategy to reduce the memory footprint of the required external storage. We show that our method performs favorably with respect to state-of-the-art approaches in the literature, while requiring significantly less space in memory and computational time. In addition, inspired inspired by to recent works, we evaluate the impact of selecting a more flexible model for the activation functions inside the network, evaluating the impact of catastrophic forgetting on the activation functions themselves.

Granger causality analysis, as one of the most popular time series causality methods, has been widely used in the economics, neuroscience. However, unobserved confounders is a fundamental problem in the observational studies, which is still not solved for the non-linear Granger causality. The application works often deal with this problem in virtue of the proxy variables, who can be treated as a measure of the confounder with noise. But the proxy variables has been proved to be unreliable, because of the bias it may induce. In this paper, we try to "recover" the unobserved confounders for the Granger causality. We use a generative model with latent variable to build the relationship between the unobserved confounders and the observed variables(tested variable and the proxy variables). The posterior distribution of the latent variable is adopted to represent the confounders distribution, which can be sampled to get the estimated confounders. We adopt the variational autoencoder to estimate the intractable posterior distribution. The recurrent neural network is applied to build the temporal relationship in the data. We evaluate our method in the synthetic and semi-synthetic dataset. The result shows our estimated confounders has a better performance than the proxy variables in the non-linear Granger causality with multiple proxies in the semi-synthetic dataset. But the performances of the synthetic dataset and the different noise level of proxy seem terrible. Any advice can really help.