Noise can speed backpropagation learning and deep bidirectional pretraining.

We show that the backpropagation algorithm is a special case of the generalized Expectation-Maximization (EM) algorithm for iterative maximum likelihood estimation. We then apply the recent result that carefully chosen noise can speed the average convergence of the EM algorithm as it climbs a hill of probability or log-likelihood. Then injecting such noise can speed the average convergence of the backpropagation algorithm for both the training and pretraining of multilayer neural networks. The beneficial noise adds to the hidden and visible neurons and related parameters. The noise also applies to regularized regression networks. This beneficial noise is just that noise that makes the current signal more probable. We show that such noise also tends to improve classification accuracy. The geometry of the noise-benefit region depends on the probability structure of the neurons in a given layer. The noise-benefit region in noise space lies above the noisy-EM (NEM) hyperplane for classification and involves a hypersphere for regression. Simulations demonstrate these noise benefits using MNIST digit classification. The NEM noise benefits substantially exceed those of simply adding blind noise to the neural network. We further prove that the noise speed-up applies to the deep bidirectional pretraining of neural-network bidirectional associative memories (BAMs) or their functionally equivalent restricted Boltzmann machines. We then show that learning with basic contrastive divergence also reduces to generalized EM for an energy-based network probability. The optimal noise adds to the input visible neurons of a BAM in stacked layers of trained BAMs. Global stability of generalized BAMs guarantees rapid convergence in pretraining where neural signals feed back between contiguous layers. Bipolar coding of inputs further improves pretraining performance.

Modeling multiple time series annotations as noisy distortions of the ground truth: An Expectation-Maximization approach.

Studies of time-continuous human behavioral phenomena often rely on ratings from multiple annotators. Since the ground truth of the target construct is often latent, the standard practice is to use ad-hoc metrics (such as averaging annotator ratings). Despite being easy to compute, such metrics may not provide accurate representations of the underlying construct. In this paper, we present a novel method for modeling multiple time series annotations over a continuous variable that computes the ground truth by modeling annotator specific distortions. We condition the ground truth on a set of features extracted from the data and further assume that the annotators provide their ratings as modification of the ground truth, with each annotator having specific distortion tendencies. We train the model using an Expectation-Maximization based algorithm and evaluate it on a study involving natural interaction between a child and a psychologist, to predict confidence ratings of the children's smiles. We compare and analyze the model against two baselines where: (i) the ground truth in considered to be framewise mean of ratings from various annotators and, (ii) each annotator is assumed to bear a distinct time delay in annotation and their annotations are aligned before computing the framewise mean.

Noise-enhanced convolutional neural networks.

Injecting carefully chosen noise can speed convergence in the backpropagation training of a convolutional neural network (CNN). The Noisy CNN algorithm speeds training on average because the backpropagation algorithm is a special case of the generalized expectation-maximization (EM) algorithm and because such carefully chosen noise always speeds up the EM algorithm on average. The CNN framework gives a practical way to learn and recognize images because backpropagation scales with training data. It has only linear time complexity in the number of training samples. The Noisy CNN algorithm finds a special separating hyperplane in the network's noise space. The hyperplane arises from the likelihood-based positivity condition that noise-boosts the EM algorithm. The hyperplane cuts through a uniform-noise hypercube or Gaussian ball in the noise space depending on the type of noise used. Noise chosen from above the hyperplane speeds training on average. Noise chosen from below slows it on average. The algorithm can inject noise anywhere in the multilayered network. Adding noise to the output neurons reduced the average per-iteration training-set cross entropy by 39% on a standard MNIST image test set of handwritten digits. It also reduced the average per-iteration training-set classification error by 47%. Adding noise to the hidden layers can also reduce these performance measures. The noise benefit is most pronounced for smaller data sets because the largest EM hill-climbing gains tend to occur in the first few iterations. This noise effect can assist random sampling from large data sets because it allows a smaller random sample to give the same or better performance than a noiseless sample gives.

Detecting paralinguistic events in audio stream using context in features and probabilistic decisions.

Non-verbal communication involves encoding, transmission and decoding of non-lexical cues and is realized using vocal (e.g. prosody) or visual (e.g. gaze, body language) channels during conversation. These cues perform the function of maintaining conversational flow, expressing emotions, and marking personality and interpersonal attitude. In particular, non-verbal cues in speech such as paralanguage and non-verbal vocal events (e.g. laughters, sighs, cries) are used to nuance meaning and convey emotions, mood and attitude. For instance, laughters are associated with affective expressions while fillers (e.g. um, ah, um) are used to hold floor during a conversation. In this paper we present an automatic non-verbal vocal events detection system focusing on the detect of laughter and fillers. We extend our system presented during Interspeech 2013 Social Signals Sub-challenge (that was the winning entry in the challenge) for frame-wise event detection and test several schemes for incorporating local context during detection. Specifically, we incorporate context at two separate levels in our system: (i) the raw frame-wise features and, (ii) the output decisions. Furthermore, our system processes the output probabilities based on a few heuristic rules in order to reduce erroneous frame-based predictions. Our overall system achieves an Area Under the Receiver Operating Characteristics curve of 95.3% for detecting laughters and 90.4% for fillers on the test set drawn from the data specifications of the Interspeech 2013 Social Signals Sub-challenge. We perform further analysis to understand the interrelation between the features and obtained results. Specifically, we conduct a feature sensitivity analysis and correlate it with each feature's stand alone performance. The observations suggest that the trained system is more sensitive to a feature carrying higher discriminability with implications towards a better system design.

Applying machine learning to facilitate autism diagnostics: pitfalls and promises.

Machine learning has immense potential to enhance diagnostic and intervention research in the behavioral sciences, and may be especially useful in investigations involving the highly prevalent and heterogeneous syndrome of autism spectrum disorder. However, use of machine learning in the absence of clinical domain expertise can be tenuous and lead to misinformed conclusions. To illustrate this concern, the current paper critically evaluates and attempts to reproduce results from two studies (Wall et al. in Transl Psychiatry 2(4):e100, 2012a; PloS One 7(8), 2012b) that claim to drastically reduce time to diagnose autism using machine learning. Our failure to generate comparable findings to those reported by Wall and colleagues using larger and more balanced data underscores several conceptual and methodological problems associated with these studies. We conclude with proposed best-practices when using machine learning in autism research, and highlight some especially promising areas for collaborative work at the intersection of computational and behavioral science.

A globally-variant locally-constant model for fusion of labels from multiple diverse experts without using reference labels.

Researchers have shown that fusion of categorical labels from multiple experts—humans or machine classifiers—improves the accuracy and generalizability of the overall classification system. Simple plurality is a popular technique for performing this fusion, but it gives equal importance to labels from all experts, who may not be equally reliable or consistent across the dataset. Estimation of expert reliability without knowing the reference labels is, however, a challenging problem. Most previous works deal with these challenges by modeling expert reliability as constant over the entire data (feature) space. This paper presents a model based on the consideration that in dealing with real-world data, expert reliability is variable over the complete feature space but constant over local clusters of homogeneous instances. This model jointly learns a classifier and expert reliability parameters without assuming knowledge of the reference labels using the Expectation-Maximization (EM) algorithm. Classification experiments on simulated data, data from the UCI Machine Learning Repository, and two emotional speech classification datasets show the benefits of the proposed model. Using a metric based on the Jensen-Shannon divergence, we empirically show that the proposed model gives greater benefit for datasets where expert reliability is highly variable over the feature space.