Reliable forecasting of the future behavior of road agents is a critical component to safe planning in autonomous vehicles. Here, we represent continuous trajectories as sequences of discrete motion tokens and cast multi-agent motion prediction as a language modeling task over this domain. Our model, MotionLM, provides several advantages: First, it does not require anchors or explicit latent variable optimization to learn multimodal distributions. Instead, we leverage a single standard language modeling objective, maximizing the average log probability over sequence tokens. Second, our approach bypasses post-hoc interaction heuristics where individual agent trajectory generation is conducted prior to interactive scoring. Instead, MotionLM produces joint distributions over interactive agent futures in a single autoregressive decoding process. In addition, the model’s sequential factorization enables temporally causal conditional rollouts. The proposed approach establishes new state-of-the-art performance for multi-agent motion prediction on the Waymo Open Motion Dataset, ranking 1ston the interactive challenge leaderboard.

The success of generative modeling in continuous domains has led to a surge of interest in generating discrete data such as molecules, source code, and graphs. However, construction histories for these discrete objects are typically not unique and so generative models must reason about intractably large spaces in order to learn. Additionally, structured discrete domains are often characterized by strict constraints on what constitutes a valid object and generative models must respect these requirements in order to produce useful novel samples. Here, we present a generative model for discrete objects employing a Markov chain where transitions are restricted to a set of local operations that preserve validity. Building off of generative interpretations of denoising autoencoders, the Markov chain alternates between producing 1) a sequence of corrupted objects that are valid but not from the data distribution, and 2) a learned reconstruction distribution that attempts to fix the corruptions while also preserving validity. This approach constrains the generative model to only produce valid objects, requires the learner to only discover local modifications to the objects, and avoids marginalization over an unknown and potentially large space of construction histories. We evaluate the proposed approach on two highly structured discrete domains, molecules and Laman graphs, and find that it compares favorably to alternative methods at capturing distributional statistics for a host of semantically relevant metrics.

Despite tremendous progress in computer vision, there has not been an attempt to apply machine learning on very large-scale medical image databases. We present an interleaved text/image deep learning system to extract and mine the semantic interactions of radiology images and reports from a national research hospital's Picture Archiving and Communication System. With natural language processing, we mine a collection of $\sim$216K representative two-dimensional images selected by clinicians for diagnostic reference and match the images with their descriptions in an automated manner. We then employ a weakly supervised approach using all of our available data to build models for generating approximate interpretations of patient images. Finally, we demonstrate a more strictly supervised approach to detect the presence and absence of a number of frequent disease types, providing more specific interpretations of patient scans. A relatively small amount of data is used for this part, due to the challenge in gathering quality labels from large raw text data. Our work shows the feasibility of large-scale learning and prediction in electronic patient records available in most modern clinical institutions. It also demonstrates the trade-offs to consider in designing machine learning systems for analyzing large medical data.

Automated computer-aided detection (CADe) has been an important tool in clinical practice and research. State-of-the-art methods often show high sensitivities at the cost of high false-positives (FP) per patient rates. We design a two-tiered coarse-to-fine cascade framework that first operates a candidate generation system at sensitivities ∼ 100% of but at high FP levels. By leveraging existing CADe systems, coordinates of regions or volumes of interest (ROI or VOI) are generated and function as input for a second tier, which is our focus in this study. In this second stage, we generate 2D (two-dimensional) or 2.5D views via sampling through scale transformations, random translations and rotations. These random views are used to train deep convolutional neural network (ConvNet) classifiers. In testing, the ConvNets assign class (e.g., lesion, pathology) probabilities for a new set of random views that are then averaged to compute a final per-candidate classification probability. This second tier behaves as a highly selective process to reject difficult false positives while preserving high sensitivities. The methods are evaluated on three data sets: 59 patients for sclerotic metastasis detection, 176 patients for lymph node detection, and 1,186 patients for colonic polyp detection. Experimental results show the ability of ConvNets to generalize well to different medical imaging CADe applications and scale elegantly to various data sets. Our proposed methods improve performance markedly in all cases. Sensitivities improved from 57% to 70%, 43% to 77%, and 58% to 75% at 3 FPs per patient for sclerotic metastases, lymph nodes and colonic polyps, respectively.

Despite tremendous progress in computer vision, effective learning on very large-scale (> 100K patients) medical image databases has been vastly hindered. We present an interleaved text/image deep learning system to extract and mine the semantic interactions of radiology images and reports from a national research hospital's picture archiving and communication system. Instead of using full 3D medical volumes, we focus on a collection of representative ~216K 2D key images/slices (selected by clinicians for diagnostic reference) with text-driven scalar and vector labels. Our system interleaves between unsupervised learning (e.g., latent Dirichlet allocation, recurrent neural net language models) on document- and sentence-level texts to generate semantic labels and supervised learning via deep convolutional neural networks (CNNs) to map from images to label spaces. Disease-related key words can be predicted for radiology images in a retrieval manner. We have demonstrated promising quantitative and qualitative results. The large-scale datasets of extracted key images and their categorization, embedded vector labels and sentence descriptions can be harnessed to alleviate the deep learning “data-hungry” obstacle in the medical domain.

Today, there are two major paradigms for vision-based autonomous driving systems: mediated perception approaches that parse an entire scene to make a driving decision, and behavior reflex approaches that directly map an input image to a driving action by a regressor. In this paper, we propose a third paradigm: a direct perception approach to estimate the affordance for driving. We propose to map an input image to a small number of key perception indicators that directly relate to the affordance of a road/traffic state for driving. Our representation provides a set of compact yet complete descriptions of the scene to enable a simple controller to drive autonomously. Falling in between the two extremes of mediated perception and behavior reflex, we argue that our direct perception representation provides the right level of abstraction. To demonstrate this, we train a deep Convolutional Neural Network using recording from 12 hours of human driving in a video game and show that our model can work well to drive a car in a very diverse set of virtual environments. We also train a model for car distance estimation on the KITTI dataset. Results show that our direct perception approach can generalize well to real driving images. Source code and data are available on our project website.