25 April 2017 in IEEE Transactions on Neural Systems and Rehabilitation Engineering
In brain-computer interfaces (BCIs) that use the steady-state visual evoked response (SSVEP), the user selects a control command by directing attention overtly or covertly to one out of several flicker stimuli. The different control channels are encoded in the frequency, phase, or time domain of the flicker signals. Here, we present a new type of SSVEP BCI, which uses only a single flicker stimulus and yet affords controlling multiple channels. The approach rests on the observation that the relative position between the stimulus and the foci of overt attention result in distinct topographies of the SSVEP response on the scalp. By classifying these topographies, the computer can determine at which position the user is gazing. Offline data analysis in a study on 12 healthy volunteers revealed that 9 targets can be recognized with about 95 ± 3% accuracy, corresponding to an information transfer rate (ITR) of 40.8 ± 3.3 b/min on average. We explored how the classification accuracy is affected by the number of control channels, the trial length, and the number of EEG channels. Our findings suggest that the EEG data from five channels over parieto-occipital brain areas are sufficient for reliably classifying the topographies and that there is a large potential to improve the ITR by optimizing the trial length. The robust performance and the simple stimulation setup suggest that this approach is a prime candidate for applications on desktop and tablet computers.
30 September 2020 in Frontiers in Robotics and AI
Given the features of a video, recurrent neural networks can be used to automatically generate a caption for the video. Existing methods for video captioning have at least three limitations. First, semantic information has been widely applied to boost the performance of video captioning models, but existing networks often fail to provide meaningful semantic features. Second, the Teacher Forcing algorithm is often utilized to optimize video captioning models, but during training and inference, different strategies are applied to guide word generation, leading to poor performance. Third, current video captioning models are prone to generate relatively short captions that express video contents inappropriately. Toward resolving these three problems, we suggest three corresponding improvements. First of all, we propose a metric to compare the quality of semantic features, and utilize appropriate features as input for a semantic detection network (SDN) with adequate complexity in order to generate meaningful semantic features for videos. Then, we apply a scheduled sampling strategy that gradually transfers the training phase from a teacher-guided manner toward a more self-teaching manner. Finally, the ordinary logarithm probability loss function is leveraged by sentence length so that the inclination of generating short sentences is alleviated. Our model achieves better results than previous models on the YouTube2Text dataset and is competitive with the previous best model on the MSR-VTT dataset.
1 January 2021 in NeuroImage
The phase of neural oscillatory signals aligns to the predicted onset of upcoming stimulation. Whether such phase alignments represent phase resets of underlying neural oscillations or just rhythmically evoked activity, and whether they can be observed in a rhythm-free visual context, however, remains unclear. Here, we recorded the magnetoencephalogram while participants were engaged in a temporal prediction task, judging the visual or tactile reappearance of a uniformly moving stimulus. The prediction conditions were contrasted with a control condition to dissociate phase adjustments of neural oscillations from stimulus-driven activity. We observed stronger delta band inter-trial phase consistency (ITPC) in a network of sensory, parietal and frontal brain areas, but no power increase reflecting stimulus-driven or prediction-related evoked activity. Delta ITPC further correlated with prediction performance in the cerebellum and visual cortex. Our results provide evidence that phase alignments of low-frequency neural oscillations underlie temporal predictions in a non-rhythmic visual and crossmodal context.
13 May 2021 in IEEE Transactions on Neural Systems and Rehabilitation Engineering
Conventional visual BCIs, in which control channels are tagged with stimulation patterns to elicit distinguishable brain patterns, has made impressive progress in terms of the information transfer rates (ITRs). However, less development has been seen with respect to user experience and complexity of the technical setup. The requirement to tag each of targets by a unique stimulus substantially limits the flexibility of conventional visual BCI systems. A method for decoding the targets in the environment flexibly was therefore proposed in the present study. A BCI speller with thirteen symbols drawn on paper was developed. The symbols were interspersed with four flickers with distinct frequencies, but the user did not have to gaze at flickers. Rather, subjects could spell a sequence by looking at the symbols on the paper. In a cue-guided spelling task, the average offline and online accuracies reached 89.3± 7.3% and 90.3± 6.9% for 13 subjects, corresponding to ITRs of 43.0± 7.4 bit/min and 43.8± 6.8 bit/min. In an additional free-spelling task for seven out of thirteen subjects, an accuracy of 92.3± 3.1% and an ITR of 45.6± 3.3 bit/min were achieved. Analysis of a simulated online system showed the possibility to reach an average ITR of 105.8 bit/min by reducing the epoch duration from 4 to 1 second. Reliable BCI control is possible by gazing at targets in the environment instead of dedicated stimuli which encode control channels. The proposed method can drastically reduce the technical effort for visual BCIs and thereby advance their applications outside the laboratory.