Neural network pruning is a popular technique used to reduce the inference costs of modern, potentially overparameterized, networks. Starting from a pre-trained network, the process is as follows: remove redundant parameters, retrain, and repeat while maintaining the same test accuracy. The result is a model that is a fraction of the size of the original with comparable predictive performance (test accuracy). Here, we reassess and evaluate whether the use of test accuracy alone in the terminating condition is sufficient to ensure that the resulting model performs well across a wide spectrum of "harder" metrics such as generalization to out-of-distribution data and resilience to noise. Across evaluations on varying architectures and data sets, we find that pruned networks effectively approximate the unpruned model, however, the prune ratio at which pruned networks achieve commensurate performance varies significantly across tasks. These results call into question the extent of \emph{genuine} overparameterization in deep learning and raise concerns about the practicability of deploying pruned networks, specifically in the context of safety-critical systems, unless they are widely evaluated beyond test accuracy to reliably predict their performance. Our code is available at https://github.com/lucaslie/torchprune.
We propose a methodology for robust, real-time place recognition using an imaging lidar, which yields image-quality high-resolution 3D point clouds. Utilizing the intensity readings of an imaging lidar, we project the point cloud and obtain an intensity image. ORB feature descriptors are extracted from the image and encoded into a bag-of-words vector. The vector, used to identify the point cloud, is inserted into a database that is maintained by DBoW for fast place recognition queries. The returned candidate is further validated by matching visual feature descriptors. To reject matching outliers, we apply PnP, which minimizes the reprojection error of visual features' positions in Euclidean space with their correspondences in 2D image space, using RANSAC. Combining the advantages from both camera and lidar-based place recognition approaches, our method is truly rotation-invariant and can tackle reverse revisiting and upside-down revisiting. The proposed method is evaluated on datasets gathered from a variety of platforms over different scales and environments. Our implementation is available at https://git.io/imaging-lidar-place-recognition
Learning composable policies for environments with complex rules and tasks is a challenging problem. We introduce a hierarchical reinforcement learning framework called the Logical Options Framework (LOF) that learns policies that are satisfying, optimal, and composable. LOF efficiently learns policies that satisfy tasks by representing the task as an automaton and integrating it into learning and planning. We provide and prove conditions under which LOF will learn satisfying, optimal policies. And lastly, we show how LOF's learned policies can be composed to satisfy unseen tasks with only 10-50 retraining steps. We evaluate LOF on four tasks in discrete and continuous domains, including a 3D pick-and-place environment.
Learning competitive behaviors in multi-agent settings such as racing requires long-term reasoning about potential adversarial interactions. This paper presents Deep Latent Competition (DLC), a novel reinforcement learning algorithm that learns competitive visual control policies through self-play in imagination. The DLC agent imagines multi-agent interaction sequences in the compact latent space of a learned world model that combines a joint transition function with opponent viewpoint prediction. Imagined self-play reduces costly sample generation in the real world, while the latent representation enables planning to scale gracefully with observation dimensionality. We demonstrate the effectiveness of our algorithm in learning competitive behaviors on a novel multi-agent racing benchmark that requires planning from image observations. Code and videos available at https://sites.google.com/view/deep-latent-competition.
We present a novel, fast differentiable simulator for soft-body learning and control applications. Existing differentiable soft-body simulators can be classified into two categories based on their time integration methods. Simulators using explicit time-stepping scheme require tiny time steps to avoid numerical instabilities in gradient computation, and simulators using implicit time integration typically compute gradients by employing the adjoint method to solve the expensive linearized dynamics. Inspired by Projective Dynamics (PD), we present DiffPD, an efficient differentiable soft-body simulator with implicit time integration. The key idea in DiffPD is to speed up backpropagation by exploiting the prefactorized Cholesky decomposition in PD to achieve a super-linear convergence rate. To handle contacts, DiffPD solves contact forces by analyzing a linear complementarity problem (LCP) and its gradients. With the assumption that contacts occur on a small number of nodes, we develop an efficient method for gradient computation by exploring the low-rank structure in the linearized dynamics. We evaluate the performance of DiffPD and observe a speedup of 4-19 times compared to the standard Newton's method in various applications including system identification, inverse design problems, trajectory optimization, and closed-loop control.
Learning complex behaviors through interaction requires coordinated long-term planning. Random exploration and novelty search lack task-centric guidance and waste effort on non-informative interactions. Instead, decision making should target samples with the potential to optimize performance far into the future, while only reducing uncertainty where conducive to this objective. This paper presents latent optimistic value exploration (LOVE), a strategy that enables deep exploration through optimism in the face of uncertain long-term rewards. We combine finite horizon rollouts from a latent model with value function estimates to predict infinite horizon returns and recover associated uncertainty through ensembling. Policy training then proceeds on an upper confidence bound (UCB) objective to identify and select the interactions most promising to improve long-term performance. We apply LOVE to visual control tasks in continuous state-action spaces and demonstrate improved sample complexity on a selection of benchmarking tasks.
The recent coronavirus pandemic has highlighted the many challenges faced by the healthcare, public safety, and economic systems when confronted with a surge in patients that require intensive treatment and a population that must be quarantined or shelter in place. The most obvious and pressing challenge is taking care of acutely ill patients while managing spread of infection within the care facility, but this is just the tip of the iceberg if we consider what could be done to prepare in advance for future pandemics. Beyond the obvious need for strengthening medical knowledge and preparedness, there is a complementary need to anticipate and address the engineering challenges associated with infectious disease emergencies. Robotic technologies are inherently programmable, and robotic systems have been adapted and deployed, to some extent, in the current crisis for such purposes as transport, logistics, and disinfection. As technical capabilities advance and as the installed base of robotic systems increases in the future, they could play a much more significant role in future crises. This report is the outcome of a virtual workshop co-hosted by the National Academy of Engineering (NAE) and the Computing Community Consortium (CCC) held on July 9-10, 2020. The workshop consisted of over forty participants including representatives from the engineering/robotics community, clinicians, critical care workers, public health and safety experts, and emergency responders. It identifies key challenges faced by healthcare responders and the general population and then identifies robotic/technological responses to these challenges. Then it identifies the key research/knowledge barriers that need to be addressed in developing effective, scalable solutions. Finally, the report ends with the following recommendations on how to implement this strategy.
A common approach for compressing NLP networks is to encode the embedding layer as a matrix $A\in\mathbb{R}^{n\times d}$, compute its rank-$j$ approximation $A_j$ via SVD, and then factor $A_j$ into a pair of matrices that correspond to smaller fully-connected layers to replace the original embedding layer. Geometrically, the rows of $A$ represent points in $\mathbb{R}^d$, and the rows of $A_j$ represent their projections onto the $j$-dimensional subspace that minimizes the sum of squared distances ("errors") to the points. In practice, these rows of $A$ may be spread around $k>1$ subspaces, so factoring $A$ based on a single subspace may lead to large errors that turn into large drops in accuracy. Inspired by \emph{projective clustering} from computational geometry, we suggest replacing this subspace by a set of $k$ subspaces, each of dimension $j$, that minimizes the sum of squared distances over every point (row in $A$) to its \emph{closest} subspace. Based on this approach, we provide a novel architecture that replaces the original embedding layer by a set of $k$ small layers that operate in parallel and are then recombined with a single fully-connected layer. Extensive experimental results on the GLUE benchmark yield networks that are both more accurate and smaller compared to the standard matrix factorization (SVD). For example, we further compress DistilBERT by reducing the size of the embedding layer by $40\%$ while incurring only a $0.5\%$ average drop in accuracy over all nine GLUE tasks, compared to a $2.8\%$ drop using the existing SVD approach. On RoBERTa we achieve $43\%$ compression of the embedding layer with less than a $0.8\%$ average drop in accuracy as compared to a $3\%$ drop previously. Open code for reproducing and extending our results is provided.
Analyzing surgical workflow is crucial for computers to understand surgeries. Deep learning techniques have recently been widely applied to recognize surgical workflows. Many of the existing temporal neural network models are limited in their capability to handle long-term dependencies in the data, instead of relying upon strong performance of the underlying per-frame visual models. We propose a new temporal network structure that leverages task-specific network representation to collect long-term sufficient statistics that are propagated by a sufficient statistics model (SSM). We leverage our approach within an LSTM back-bone for the task of surgical phase recognition and explore several choices for propagated statistics. We demonstrate superior results over existing state-of-the-art segmentation and novel segmentation techniques, on two laparoscopic cholecystectomy datasets: the already published Cholec80dataset and MGH100, a novel dataset with more challenging, yet clinically meaningful, segment labels.