Ideal supervised classification assumes known correct labels, but various truthing issues can arise in practice: noisy labels; multiple, conflicting labels for a sample; missing labels; and different labeler combinations for different samples. Previous work introduced a noisy-label model, which views the observed noisy labels as random variables conditioned on the unobserved correct labels. It has mainly focused on estimating the conditional distribution of the noisy labels and the class prior, as well as estimating the correct labels or training with noisy labels. In a complementary manner, given the conditional distribution and class prior, we apply estimation theory to classifier testing, training, and comparison of different combinations of labelers. First, for binary classification, we construct a testing model and derive approximate marginal posteriors for accuracy, precision, recall, probability of false alarm, and F-score, and joint posteriors for ROC and precision-recall analysis. We propose minimum mean-square error (MMSE) testing, which employs empirical Bayes algorithms to estimate the testing-model parameters and then computes optimal point estimates and credible regions for the metrics. We extend the approach to multi-class classification to obtain optimal estimates of accuracy and individual confusion-matrix elements. Second, we present a unified view of training that covers probabilistic (i.e., discriminative or generative) and non-probabilistic models. For the former, we adjust maximum-likelihood or maximum a posteriori training for truthing issues; for the latter, we propose MMSE training, which minimizes the MMSE estimate of the empirical risk. We also describe suboptimal training that is compatible with existing infrastructure. Third, we observe that mutual information lets one express any labeler combination as an equivalent single labeler, implying that multiple mediocre labelers can be as informative as, or more informative than, a single expert labeler. Experiments demonstrate the effectiveness of the methods and confirm the implication.

Virtual humans are frequently used to help medical students practice communication skills. Here, we show that communication skills features drawn from the literature on best practices for doctor-patient communication can be used to predict student interviewers' success in a given domain skill. We also demonstrate the viability of Bayesian Rule Lists, an interpretable machine learning model, for this use case. Bayesian Rule Lists' predictive performance is comparable to that of other other commonly used algorithms, including decision trees. This suggests that Bayesian Rule Lists, which produce simple, human-readable trained binary classifiers, may be suitable for providing feedback for educational purposes.

The paper presents a model for watermarking and some attacks on watermarks. Given the watermarked signal, the so-called Wiener attack performs minimum mean-squared error (MMSE) estimation of the watermark and subtracts the weighted MMSE estimate from the watermarked signal. Under the assumption of a fixed correlation detector, the attack is shown to minimize the expected correlation statistic for the same attack distortion among linear, shift-invariant filtering attacks. It also leads to the idea of energy-efficient watermarking: watermarking that resists MMSE estimation as much as possible, and provides a meaningful way to evaluate robustness. The paper shows that energy-efficient watermarks must satisfy a power-spectrum condition (PSC), which states that the watermark's power spectrum should be directly proportional to the original signal's. PSC-compliant watermarks are proven to be most robust. Experiments with signal models and natural images demonstrate that watermarks that do not closely fulfill the PSC are vulnerable to the Wiener attack, while PSC-compliant watermarks are highly resistant to it. These theoretical and experimental results justify prior heuristic arguments that, for maximum robustness, a watermark should be closely matched to the spectral content of the original signal. The results also discourage the use of watermarks that do not approximately satisfy the PSC.

It was shown previously that, in some situations, blind watermarking can perform as well as watermarking schemes with the host signal available to the decoder. In this paper, blind watermarking of colored Gaussian host signals in the presence of filtering and additive Gaussian noise attacks is discussed. Three suboptimal but practical schemes are compared with a scheme where the host signal is available at the decoder. The performance is analyzed theoretically and experimentally for image watermarking.

An extension of conventional block motion compensation (BMC), overlapped block motion compensation (OBMC) has been shown to reduce residual errors and blocking effects in motion-compensated video. However, the overlap creates a noncausal spatial dependence between blocks and complicates motion estimation (ME) for OBMC. Iterative methods have traditionally been employed for overlapped block motion estimation (OBME). For compression, the rate for the motion vector field (MVF) may also be constrained. This work considers several rate-constrained OBME algorithms, both iterative and noniterative. Experiments demonstrate that a simple raster-scan algorithm is effective as a suboptimal, noniterative solution, with comparable or better rate-distortion performance and computational complexity than iterative OBME algorithms. Depending on the application, either this method or a simple block-matching algorithm plus iteration are the most attractive of the tested OBME schemes.

A simple but powerful theoretical model for watermarking and attacks is presented. The attack consists of Wiener estimation of the watermark, followed by weighted feedback of the estimate to reduce the detector's correlation value. The Wiener attack is provably optimal, and it leads to the notion of energy-efficient watermarking, the generation of watermarks that resist Wiener estimation. It also leads to a meaningful way to measure watermark robustness. Designing watermarks to resist the attack produces a power-spectrum condition (PSC): the watermark's power spectrum should be a scaled version of the original document's. Experiments with signal models demonstrate the importance of the PSC and show that watermarks that fail to meet the PSC may be vulnerable to attack. These analytic predictions are confirmed with experiments on natural images. The paper provides a theoretical basis for the popular heuristic argument that a watermark should be embedded in the "perceptually significant frequency components" of the original document and discourages the use of watermarks whose power spectra are not closely matched to the original document's power spectrum.

For many video compression systems, overlapped block motion compensation (OBMC) provides an effective extension of traditional block MC (BMC). However, because OBMC causes noncausal spatial interaction between blocks, iterative methods have typically been used for motion estimation (ME) for OBMC. For compression purposes, the rate for the motion vector field may also be constrained. This paper considers several rate-constrained ME methods for OBMC. Experiments demonstrate that a simple raster-scan algorithm is effective as a suboptimal non-iterative solution, with better rate-distortion performance and equal or lower computational complexity than several iterative methods.

In transform image coding, the histograms of transform coefficients can be approximately modeled by generalized Gaussian (GG) random variables. However, the GG models may not fit the DC distribution. One approach uses DPCM for the DC data, which greatly complicates bit allocation; another assumes a single Gaussian (SG) model, which may be a poor model. As an alternative, this paper proposes a finite Gaussian mixture (GM) model for the DC data. The GM approach does not require tweaking of the DPCM quantizer stepsize and can allocate bits optimally between the DC and AC data; it is also more flexible than the SG model. Experimentally, the GM method matched DPCM at medium rates and gave 1-5 dB higher PSNR at low and high rates. The GM method also matched the performance of the SG model and gave 0.5-2 dB higher PSNR when the SG assumption failed.

An iterative post-processing algorithm for reducing blocking and ringing in block transform-coded images is presented. It smooths the decompressed image and uses convex projections theory to preserve the image's edges and bit-stream. The algorithm exploits the coding scheme's block structure and ringing behavior. In experiments it outperforms an existing method and gives significant subjective and objective improvement over the decompressed image. Sufficient conditions for algorithm convergence are given.