Learning goal-directed behavior in environments with sparse feedback is a major challenge for reinforcement learning algorithms. The primary difficulty arises due to insufficient exploration, resulting an agent being unable learn robust value functions. Intrinsically motivated agents can explore new its own sake rather than directly solve problems. Such intrinsic behaviors could eventually help the tasks posed by environment. We present hierarchical-DQN (h-DQN), framework integrate...

The predictive performance of supervised learning algorithms depends on the quality labels. In a typical label collection process, multiple annotators provide subjective noisy estimates ``truth" under influence their varying skill-levels and biases. Blindly treating these labels as ground truth limits accuracy in presence strong disagreement. This problem is critical for applications domains such medical imaging where both annotation cost inter-observer variability are high. this work, we...

Traditional topic models do not account for semantic regularities in language.Recent distributional representations of words exhibit consistency over directional metrics such as cosine similarity.However, neither categorical nor Gaussian observational distributions used existing are appropriate to leverage correlations.In this paper, we propose use the von Mises-Fisher distribution model density a unit sphere.Such representation is well-suited data.We Hierarchical Dirichlet Process our base...

Learning robust value functions given raw observations and rewards is now possible with model-free model-based deep reinforcement learning algorithms. There a third alternative, called Successor Representations (SR), which decomposes the function into two components -- reward predictor successor map. The map represents expected future state occupancy from any maps states to scalar rewards. of can be computed as inner product between weights. In this paper, we present DSR, generalizes SR...

The predictive performance of supervised learning algorithms depends on the quality labels. In a typical label collection process, multiple annotators provide subjective noisy estimates "truth" under influence their varying skill-levels and biases. Blindly treating these labels as ground truth limits accuracy in presence strong disagreement. This problem is critical for applications domains such medical imaging where both annotation cost inter-observer variability are high. this work, we...

Markov jump processes (MJPs) are used to model a wide range of phenomena from disease progression RNA path folding. However, maximum likelihood estimation parametric models leads degenerate trajectories and inferential performance is poor in nonparametric models. We take small-variance asymptotics (SVA) approach overcome these limitations. derive the for MJPs both directly observed hidden state In case we obtain novel objective function which non-degenerate trajectories. To version introduce...

Traditional topic models do not account for semantic regularities in language. Recent distributional representations of words exhibit consistency over directional metrics such as cosine similarity. However, neither categorical nor Gaussian observational distributions used existing are appropriate to leverage correlations. In this paper, we propose use the von Mises-Fisher distribution model density a unit sphere. Such representation is well-suited data. We Hierarchical Dirichlet Process our...

We propose the segmented iHMM (siHMM), a hierarchical infinite hidden Markov model (iHMM) that supports simple, efficient inference scheme. The siHMM is well suited to segmentation problems, where goal identify points at which time series transitions from one relatively stable regime new regime. Conventional iHMMs often struggle with such since they have no mechanism for distinguishing between high- and low-level dynamics. Hierarchical HMMs (HHMMs) can do better, but require much more...

Professional-grade software applications are powerful but complicated$-$expert users can achieve impressive results, novices often struggle to complete even basic tasks. Photo editing is a prime example: after loading photo, the user confronted with an array of cryptic sliders like "clarity", "temp", and "highlights". An automatically generated suggestion could help, there no single "correct" edit for given image$-$different experts may make very different aesthetic decisions when faced same...

Clinical Question Answering (QA) systems enable doctors to quickly access patient information from electronic health records (EHRs). However, training these requires significant annotated data, which is limited due the expertise needed and privacy concerns associated with clinical data. This paper explores generating QA data using large language models (LLMs) in a zero-shot setting. We find that naive prompting often results easy questions do not reflect complexity of scenarios. To address...

Knowledge distillation has been used to capture the knowledge of a teacher model and distill it into student with some desirable characteristics such as being smaller, more efficient, or generalizable. In this paper, we propose framework for distilling powerful discriminative neural network commonly graphical models known be interpretable (e.g., topic models, autoregressive Hidden Markov Models). Posterior latent variables in these proportions models) is often feature representation...

In this note we provide detailed derivations of two versions small-variance asymptotics for hierarchical Dirichlet process (HDP) mixture models and the HDP hidden Markov model (HDP-HMM, a.k.a. infinite HMM). We include probabilities certain CRP CRF partitions, which are more general interest.