Refine
Document Type
- Article (1)
- Part of a Book (1)
- Doctoral Thesis (1)
Language
- English (3)
Has Fulltext
- yes (3)
Is part of the Bibliography
- no (3)
Keywords
- neural networks (3) (remove)
Institute
- Informatik (2)
- Informatik und Mathematik (1)
AttendAffectNet-emotion prediction of movie viewers using multimodal fusion with self-attention
(2021)
In this paper, we tackle the problem of predicting the affective responses of movie viewers, based on the content of the movies. Current studies on this topic focus on video representation learning and fusion techniques to combine the extracted features for predicting affect. Yet, these typically, while ignoring the correlation between multiple modality inputs, ignore the correlation between temporal inputs (i.e., sequential features). To explore these correlations, a neural network architecture—namely AttendAffectNet (AAN)—uses the self-attention mechanism for predicting the emotions of movie viewers from different input modalities. Particularly, visual, audio, and text features are considered for predicting emotions (and expressed in terms of valence and arousal). We analyze three variants of our proposed AAN: Feature AAN, Temporal AAN, and Mixed AAN. The Feature AAN applies the self-attention mechanism in an innovative way on the features extracted from the different modalities (including video, audio, and movie subtitles) of a whole movie to, thereby, capture the relationships between them. The Temporal AAN takes the time domain of the movies and the sequential dependency of affective responses into account. In the Temporal AAN, self-attention is applied on the concatenated (multimodal) feature vectors representing different subsequent movie segments. In the Mixed AAN, we combine the strong points of the Feature AAN and the Temporal AAN, by applying self-attention first on vectors of features obtained from different modalities in each movie segment and then on the feature representations of all subsequent (temporal) movie segments. We extensively trained and validated our proposed AAN on both the MediaEval 2016 dataset for the Emotional Impact of Movies Task and the extended COGNIMUSE dataset. Our experiments demonstrate that audio features play a more influential role than those extracted from video and movie subtitles when predicting the emotions of movie viewers on these datasets. The models that use all visual, audio, and text features simultaneously as their inputs performed better than those using features extracted from each modality separately. In addition, the Feature AAN outperformed other AAN variants on the above-mentioned datasets, highlighting the importance of taking different features as context to one another when fusing them. The Feature AAN also performed better than the baseline models when predicting the valence dimension.
The brain is a highly dynamic and variable system: when the same stimulus is presented to the same animal on the same day multiple times, the neural responses show high trial-to-trial variability. In addition, even in the absence of sensory stimulation neural recordings spontaneously show seemingly random activity patterns. Evoked and spontaneous neural variability is not restricted to activity but is also found in structure: most synapses do not survive for longer than two weeks and even those that do show high fluctuations in their efficacy.
Both forms of variability are further affected by stochastic components of neural processing such as frequent transmission failure. At present it is unclear how these observations relate to each other and how they arise in cortical circuits.
Here, we will investigate how the self-organizational processes of neural circuits affect the high variability in two different directions: First, we will show that recurrent dynamics of self-organizing neural networks can account for key features of neural variability. This is achieved in the absence of any intrinsic noise sources by the neural network models learning a predictive model of their environment with sampling-like dynamics. Second, we will show that the same self-organizational processes can compensate for intrinsic noise sources. For this, an analytical model and more biologically plausible models are established to explain the alignment of parallel synapses in the presence of synaptic failure.
Both modeling studies predict properties of neural variability, of which two are subsequently tested on a synapse database from a dense electron microscopy reconstruction from mouse somatosensory cortex and on multi-unit recordings from the visual cortex of macaque monkeys during a passive viewing task. While both analyses yield interesting results, the predicted properties were not confirmed, guiding the next iteration of experiments and modeling studies.
One of the most interesting domains of feedforward networks is the processing of sensor signals. There do exist some networks which extract most of the information by implementing the maximum entropy principle for Gaussian sources. This is done by transforming input patterns to the base of eigenvectors of the input autocorrelation matrix with the biggest eigenvalues. The basic building block of these networks is the linear neuron, learning with the Oja learning rule. Nevertheless, some researchers in pattern recognition theory claim that for pattern recognition and classification clustering transformations are needed which reduce the intra-class entropy. This leads to stable, reliable features and is implemented for Gaussian sources by a linear transformation using the eigenvectors with the smallest eigenvalues. In another paper (Brause 1992) it is shown that the basic building block for such a transformation can be implemented by a linear neuron using an Anti-Hebb rule and restricted weights. This paper shows the analog VLSI design for such a building block, using standard modules of multiplication and addition. The most tedious problem in this VLSI-application is the design of an analog vector normalization circuitry. It can be shown that the standard approaches of weight summation will not give the convergence to the eigenvectors for a proper feature transformation. To avoid this problem, our design differs significantly from the standard approaches by computing the real Euclidean norm. Keywords: minimum entropy, principal component analysis, VLSI, neural networks, surface approximation, cluster transformation, weight normalization circuit.