Task-specific Systems Research Articles

On the self-adjustment of privacy safeguards for query log streams

Internet-based services process and store numerous search queries around the globe. The use of web search engines, such as Bing and Google, as well as personal assistants (e.g., Alexa and Cortana) and task specific systems (e.g., YouTube, Netflix, Amazon) are relevant examples. The queries associated to such services may be stored and sold out for profit. Before doing so, personal and sensitive information must be sanitized, as requested by current regulations. This can be cumbersome for some organizations. We present an automated solution for anonymizing unstructured data, like the one used within query logs. Our solution uses a light-weight probabilistic k-anonymity approach, which allows verifiable real-time privacy protection. It addresses previous limitations and improves performance. We validate the feasibility of the approach, under some evaluation metrics including data utility, privacy and speed.

Intricate Twitternlp Modelling

Abstract: In this research, we introduce TweetNLP, a platform for social media Natural Language Processing (NLP). An extensive range of NLP tasks are supported by TweetNLP, including standard focus areas like sentiment analysis and named entity recognition as well as social media-specific tasks like emoticon prediction and offensive language detection. Task-specific systems run on moderately small Transformer-based language models that are focused on social media text, particularly Twitter, and don't require specialized hardware or cloud services to operate. TweetNLP's major contributions are: (1) an integrated Python library for a contemporary toolkit supporting social media analysis using various task-specific models tailored to the social domain; (2) an interactive online demo for codeless experimentation using our models; and (3) a tutorial covering a wide range of typical social media applications.

MT-clinical BERT: scaling clinical information extraction with multitask learning.

Clinical notes contain an abundance of important, but not-readily accessible, information about patients. Systems that automatically extract this information rely on large amounts of training data of which there exists limited resources to create. Furthermore, they are developed disjointly, meaning that no information can be shared among task-specific systems. This bottleneck unnecessarily complicates practical application, reduces the performance capabilities of each individual solution, and associates the engineering debt of managing multiple information extraction systems. We address these challenges by developing Multitask-Clinical BERT: a single deep learning model that simultaneously performs 8 clinical tasks spanning entity extraction, personal health information identification, language entailment, and similarity by sharing representations among tasks. We compare the performance of our multitasking information extraction system to state-of-the-art BERT sequential fine-tuning baselines. We observe a slight but consistent performance degradation in MT-Clinical BERT relative to sequential fine-tuning. These results intuitively suggest that learning a general clinical text representation capable of supporting multiple tasks has the downside of losing the ability to exploit dataset or clinical note-specific properties when compared to a single, task-specific model. We find our single system performs competitively with all state-the-art task-specific systems while also benefiting from massive computational benefits at inference.

Visualization of the Mechanical Wave Effect on Liquid Microphases and Its Application for the Tuning of Dissipative Soft Microreactors.

The development of approaches for creation of adaptive and stimuli-responsive chemical systems is particularly important for chemistry, materials science, and biotechnology. The understanding of response mechanisms for various external forces is highly demanded for the rational design of task-specific systems. Here, we report direct liquid-phase scanning electron microscopy (SEM) observations of the high frequency sound-wave-driven restructuring of liquid media on the microlevel, leading to switching of its chemical behavior. We show that under the action of ultrasound, the microstructured ionic liquid/water mixture undergoes rearrangement resulting in formation of separated phases with specific compositions and reactivities. The observed effect was successfully utilized for creation of dissipative soft microreactors formed in ionic liquid/water media during the sonication-driven water transfer. The performance of the microreactors was demonstrated using the example of controlled synthesis of small and uniform gold and palladium nanoparticles. The microsonication stage, designed and used in the present study, opened unique opportunities for direct sonochemical studies with the use of electron microscopy.

DARPA's Role in Machine Learning

Machine learning methods provide a way for artificial intelligence systems to learn from experience. This article describes four threads of machine learning research supported and guided by the Defense Advanced Research Projects Agency — probabilistic modeling for speech recognition, probabilistic relational models, the integration of multiple machine learning approaches into a task‐specific system, and neural network technology. These threads illustrate the Defense Advanced Research Projects Agency way of creating timely advances in a field.

Individual differences in verbal short-term memory and reading aloud: Semantic compensation for weak phonological processing across tasks.

According to contemporary accounts, linguistic behavior reflects the interaction of distinct representations supporting word meaning and phonology. However, there is controversy about the extent to which this interaction occurs within task-specific systems, specialized for reading and short-term memory (STM), as opposed to between components that support the full range of linguistic tasks. We examined whether individual differences in the efficiency of phonological processing would relate to the application of lexical-semantic knowledge to support verbal STM, single word reading and repetition. In a sample of 83 participants, we related nonword performance in each task (as a marker of phonological capacity in the absence of meaning) to the effects of word imageability (a lexical-semantic variable). We found stronger reliance on lexical-semantic knowledge in participants with weaker phonological processing. This relationship held across tasks, suggesting that lexical-semantic processing can compensate for phonological weakness which would otherwise give rise to poor performance. Our results are consistent with separable yet interacting primary systems for phonology and semantics, with lexical-semantic knowledge supporting pattern completion within the phonological system in a similar way across STM and reading tasks. (PsycINFO Database Record (c) 2019 APA, all rights reserved).

Similar but separate systems underlie perceptual bistability in vision and audition

The dynamics of perceptual bistability, the phenomenon in which perception switches between different interpretations of an unchanging stimulus, are characterised by very similar properties across a wide range of qualitatively different paradigms. This suggests that perceptual switching may be triggered by some common source. However, it is also possible that perceptual switching may arise from a distributed system, whose components vary according to the specifics of the perceptual experiences involved. Here we used a visual and an auditory task to determine whether individuals show cross-modal commonalities in perceptual switching. We found that individual perceptual switching rates were significantly correlated across modalities. We then asked whether perceptual switching arises from some central (modality-) task-independent process or from a more distributed task-specific system. We found that a log-normal distribution best explained the distribution of perceptual phases in both modalities, suggestive of a combined set of independent processes causing perceptual switching. Modality- and/or task-dependent differences in these distributions, and lack of correlation with the modality-independent central factors tested (ego-resiliency, creativity, and executive function), also point towards perceptual switching arising from a distributed system of similar but independent processes.

A Bayesian Model of Diachronic Meaning Change

Word meanings change over time and an automated procedure for extracting this information from text would be useful for historical exploratory studies, information retrieval or question answering. We present a dynamic Bayesian model of diachronic meaning change, which infers temporal word representations as a set of senses and their prevalence. Unlike previous work, we explicitly model language change as a smooth, gradual process. We experimentally show that this modeling decision is beneficial: our model performs competitively on meaning change detection tasks whilst inducing discernible word senses and their development over time. Application of our model to the SemEval-2015 temporal classification benchmark datasets further reveals that it performs on par with highly optimized task-specific systems.

Pseudo-Haptic Feedback in Teleoperation.

In this paper, we develop possible realizations of pseudo-haptic feedback in teleoperation systems based on existing works for pseudo-haptic feedback in virtual reality and the intended applications. We derive four potential factors affecting the performance of haptic feedback (calculation operator, maximum displacement, offset force, and scaling factor), which are analyzed in three compliance identification experiments. First, we analyze the principle usability of pseudo-haptic feedback by comparing information transfer measures for teleoperation and direct interaction. Pseudo-haptic interaction yields well above-chance performance, while direct interaction performs almost perfectly. In order to optimize pseudo-haptic feedback, in the second study we perform a full-factorial experimental design with 36subjects performing 6,480trials with 36different treatments. Information transfer ranges from 0.68bit to 1.72bit in a task with a theoretical maximum of 2.6bit, with a predominant effect of the calculation operator and a minor effect of the maximum displacement. In a third study, short- and long-term learning effects are analyzed. Learning effects regarding the performance of pseudo-haptic feedback cannot be observed for single-day experiments. Tests over 10 days show a maximum increase in information transfer of 0.8bit. The results show the feasibility of pseudo-haptic feedback for teleoperation and can be used as design basis for task-specific systems.

Neural Mechanisms Underlying Motivation of Mental Versus Physical Effort

Mental and physical efforts, such as paying attention and lifting weights, have been shown to involve different brain systems. These cognitive and motor systems, respectively, include cortical networks (prefronto-parietal and precentral regions) as well as subregions of the dorsal basal ganglia (caudate and putamen). Both systems appeared sensitive to incentive motivation: their activity increases when we work for higher rewards. Another brain system, including the ventral prefrontal cortex and the ventral basal ganglia, has been implicated in encoding expected rewards. How this motivational system drives the cognitive and motor systems remains poorly understood. More specifically, it is unclear whether cognitive and motor systems can be driven by a common motivational center or if they are driven by distinct, dedicated motivational modules. To address this issue, we used functional MRI to scan healthy participants while performing a task in which incentive motivation, cognitive, and motor demands were varied independently. We reasoned that a common motivational node should (1) represent the reward expected from effort exertion, (2) correlate with the performance attained, and (3) switch effective connectivity between cognitive and motor regions depending on task demand. The ventral striatum fulfilled all three criteria and therefore qualified as a common motivational node capable of driving both cognitive and motor regions of the dorsal striatum. Thus, we suggest that the interaction between a common motivational system and the different task-specific systems underpinning behavioral performance might occur within the basal ganglia.

Encoding of emotional facial expressions in direct and incidental tasks: Two event-related potential studies*

Emotional face encoding process was explored through electroencephalographic measures (event-related potentials [ERPs]). Previous studies have demonstrated an emotion-specific cognitive process in face comprehension. However, the effect of emotional significance of the stimuli (type of emotion) and task (direct or indirect task) on the ERP is uncertain. In Experiment 1 (indirect task) ERP correlates of 21 subjects were recorded when they viewed emotional (anger, sadness and happiness) or neutral facial stimuli. An emotion-specific cortical variation was found, a negative deflection at approximately 200 ms after simulus (N2 effect). This effect was sensitive to the emotional valence of faces, because it differentiated high arousal emotions (i.e., anger) from low arousal emotions (i.e., sadness). Moreover, a specific cortical site (posterior) was activated by emotional faces but not by neutral faces. In Experiment 2 (direct task), the authors investigated whether encoding for emotional faces relies on a single neural system irrespective of the task, or whether it is supported by multiple, task-specific systems. Differences in the cortical distribution (posterior for incidental task; central and posterior for direct task) and lateralisation (right-distribution for the negative emotions in direct task) of N2 on the scalp were observed in the different tasks. This indicates that distinct task-specific cortical responses to emotional focus can be detected with ERP methodology.

Genericity and portability for task-independent speech recognition

As core speech recognition technology improves, opening up a wider range of applications, genericity and portability are becoming important issues. Most of todays recognition systems are still tuned to a particular task and porting the system to a new task (or language) requires a substantial investment of time and money, as well as human expertise. This paper addresses issues in speech recognizer portability and in the development of generic core speech recognition technology. First, the genericity of wide domain models is assessed by evaluating their performance on several tasks of varied complexity. Then, techniques aimed at enhancing the genericity of these wide domain models are investigated. Multi-source acoustic training is shown to reduce the performance gap between task-independent and task-dependent acoustic models, and for some tasks to out-perform task-dependent acoustic models. Transparent methods for porting generic models to a specific task are also explored. Transparent unsupervised acoustic model adaptation is contrasted with supervised adaptation, and incremental unsupervised adaptation of both the acoustic and linguistic models is investigated. Experimental results on a dialog task show that with the proposed scheme, a transparently adapted generic system can perform nearly as well (about a 1% absolute gap in word error rate) as a task-specific system trained on several tens of hours of manually transcribed data.

Does metacognition necessarily involve metarepresentation?

Against the view that metacognition is a capacity that parallels theory of mind, it is argued that metacognition need involve neither metarepresentation nor semantic forms of reflexivity, but only process-reflexivity, through which a task-specific system monitors its own internal feedback by using quantitative cues. Metacognitive activities, however, may be redescribed in metarepresentational, mentalistic terms in species endowed with a theory of mind.