Decoding Weights Research Articles

Contrastive deep convolutional transform k-means clustering

Deep clustering has gained the immense attention of researchers in recent years. Most of the deep clustering approaches are based on auto-encoders which consist of an encoder-decoder framework. In these approaches, the clustering module is embedded in the latent space of auto-encoders. The auto-encoder based deep clustering approaches require learning of encoder weights as well as decoder weights. Moreover, due to the unsupervised learning strategy, these approaches lack in learning the discriminative features that can help in generating better clusters. This work introduces a novel clustering approach based on Contrastive Deep Convolutional Transform Learning (DCTL) framework. The proposed approach mitigates the problem of lack of supervision in DCTL based K-means clustering approach by embedding the contrastive learning into it. To embed the contrastive learning, the positive pairs and negative pairs of data samples are generated by reconstructing the data samples from the DCTL learnt representation itself and thus eliminates the requirement of data augmentation for embedding contrastive learning. The experimental results on several benchmark facial images datasets demonstrate that the proposed framework gives better clustering performance as compared to the current state-of-the-art deep clustering approaches especially in data constrained scenarios.

Efficient Decoding of Large-Scale Neural Population Responses With Gaussian-Process Multiclass Regression.

Neural decoding methods provide a powerful tool for quantifying the information content of neural population codes and the limits imposed by correlations in neural activity. However, standard decoding methods are prone to overfitting and scale poorly to high-dimensional settings. Here, we introduce a novel decoding method to overcome these limitations. Our approach, the gaussian process multiclass decoder (GPMD), is well suited to decoding a continuous low-dimensional variable from high-dimensional population activity and provides a platform for assessing the importance of correlations in neural population codes. The GPMD is a multinomial logistic regression model with a gaussian process prior over the decoding weights. The prior includes hyperparameters that govern the smoothness of each neuron's decoding weights, allowing automatic pruning of uninformative neurons during inference. We provide a variational inference method for fitting the GPMD to data, which scales to hundreds or thousands of neurons and performs well even in data sets with more neurons than trials. We apply the GPMD to recordings from primary visual cortex in three species: monkey, ferret, and mouse. Our decoder achieves state-of-the-art accuracy on all three data sets and substantially outperforms independent Bayesian decoding, showing that knowledge of the correlation structure is essential for optimal decoding in all three species.

RADAM: Texture recognition through randomized aggregated encoding of deep activation maps

Texture analysis is a classical yet challenging task in computer vision for which deep neural networks are actively being applied. Most approaches are based on building feature aggregation modules around a pre-trained backbone and then fine-tuning the new architecture on specific texture recognition tasks. Here we propose a new method named Random encoding of Aggregated Deep Activation Maps (RADAM) which extracts rich texture representations without ever changing the backbone. The technique consists of encoding the output at different depths of a pre-trained deep convolutional network using a Randomized Autoencoder (RAE). The RAE is trained locally to each image using a closed-form solution, and its decoder weights are used to compose a 1-dimensional texture representation that is fed into a linear SVM. This means that no fine-tuning or backpropagation is needed for the backbone. We explore RADAM on several texture benchmarks and achieve state-of-the-art results with different computational budgets. Our results suggest that pre-trained backbones may not require additional fine-tuning for texture recognition if their learned representations are better encoded.

Unraveling Functional Diversity of Cortical Synaptic Architecture Through the Lens of Population Coding.

The synaptic inputs to single cortical neurons exhibit substantial diversity in their sensory-driven activity. What this diversity reflects is unclear, and appears counter-productive in generating selective somatic responses to specific stimuli. One possibility is that this diversity reflects the propagation of information from one neural population to another. To test this possibility, we bridge population coding theory with measurements of synaptic inputs recorded in vivo with two-photon calcium imaging. We construct a probabilistic decoder to estimate the stimulus orientation from the responses of a realistic, hypothetical input population of neurons to compare with synaptic inputs onto individual neurons of ferret primary visual cortex (V1) recorded with two-photon calcium imaging in vivo. We find that optimal decoding requires diverse input weights and provides a straightforward mapping from the decoder weights to excitatory synapses. Analytically derived weights for biologically realistic input populations closely matched the functional heterogeneity of dendritic spines imaged in vivo with two-photon calcium imaging. Our results indicate that synaptic diversity is a necessary component of information transmission and reframes studies of connectivity through the lens of probabilistic population codes. These results suggest that the mapping from synaptic inputs to somatic selectivity may not be directly interpretable without considering input covariance and highlights the importance of population codes in pursuit of the cortical connectome.

A Manifold Learning Perspective on Representation Learning: Learning Decoder and Representations without an Encoder

Autoencoders are commonly used in representation learning. They consist of an encoder and a decoder, which provide a straightforward method to map n-dimensional data in input space to a lower m-dimensional representation space and back. The decoder itself defines an m-dimensional manifold in input space. Inspired by manifold learning, we showed that the decoder can be trained on its own by learning the representations of the training samples along with the decoder weights using gradient descent. A sum-of-squares loss then corresponds to optimizing the manifold to have the smallest Euclidean distance to the training samples, and similarly for other loss functions. We derived expressions for the number of samples needed to specify the encoder and decoder and showed that the decoder generally requires much fewer training samples to be well-specified compared to the encoder. We discuss the training of autoencoders in this perspective and relate it to previous work in the field that uses noisy training examples and other types of regularization. On the natural image data sets MNIST and CIFAR10, we demonstrated that the decoder is much better suited to learn a low-dimensional representation, especially when trained on small data sets. Using simulated gene regulatory data, we further showed that the decoder alone leads to better generalization and meaningful representations. Our approach of training the decoder alone facilitates representation learning even on small data sets and can lead to improved training of autoencoders. We hope that the simple analyses presented will also contribute to an improved conceptual understanding of representation learning.

Transfer Learning for LiDAR-Based Lane Marking Detection and Intensity Profile Generation

Recently, light detection and ranging (LiDAR)-based mobile mapping systems (MMS) have been utilized for extracting lane markings using deep learning frameworks. However, huge datasets are required for training neural networks. Furthermore, with accurate lane markings being detected utilizing LiDAR data, an algorithm for automatically reporting their intensity information is beneficial for identifying worn-out or missing lane markings. In this paper, a transfer learning approach based on fine-tuning of a pretrained U-net model for lane marking extraction and a strategy for generating intensity profiles using the extracted results are presented. Starting from a pretrained model, a new model can be trained better and faster to make predictions on a target domain dataset with only a few training examples. An original U-net model trained on two-lane highways (source domain dataset) was fine-tuned to make accurate predictions on datasets with one-lane highway patterns (target domain dataset). Specifically, encoder- and decoder-trained U-net models are presented wherein, during retraining of the former, only weights in the encoder path of U-net were allowed to change with decoder weights frozen and vice versa for the latter. On the test data (target domain), the encoder-trained model (F1-score: 86.9%) outperformed the decoder-trained (F1-score: 82.1%). Additionally, on an independent dataset, the encoder-trained one (F1-score: 90.1%) performed better than the decoder-trained one (F1-score: 83.2%). Lastly, on the basis of lane marking results obtained from the encoder-trained U-net, intensity profiles were generated. Such profiles can be used to identify lane marking gaps and investigate their cause through RGB imagery visualization.

A novel weight initialization with adaptive hyper-parameters for deep semantic segmentation

The semantic segmentation process divides an image into its constituent objects and background by assigning a corresponding class label to each pixel in the image. Semantic segmentation is an important area in computer vision with wide practical applications. The contemporary semantic segmentation approaches are primarily based on two types of deep neural networks architectures i.e., symmetric and asymmetric networks. Both types of networks consist of several layers of neurons which are arranged in two sections called encoder and decoder. The encoder section receives the input image and the decoder section outputs the segmented image. However, both sections in symmetric networks have the same number of layers and the number of neurons in an encoder layer is the same as that of the corresponding layer in the decoder section but asymmetric networks do not strictly follow such one-one correspondence between encoder and decoder layers. At the moment, SegNet and ESNet are the two leading state-of-the-art symmetric encoder-decoder deep neural network architectures. However, both architectures require extensive training for good generalization and need several hundred epochs for convergence. This paper aims to improve the convergence and enhance network generalization by introducing two novelties into the network training process. The first novelty is a weight initialization method and the second contribution is an adaptive mechanism for dynamic layer learning rate adjustment in training loop. The proposed initialization technique uses transfer learning to initialize the encoder section of the network, but for initialization of decoder section, the weights of the encoder section layers are copied to the corresponding layers of the decoder section. The second contribution of the paper is an adaptive layer learning rate method, wherein the learning rates of the encoder layers are updated based on a metric representing the difference between the probability distributions of the input images and encoder weights. Likewise, the learning rates of the decoder layers are updated based on the difference between the probability distributions of the output labels and decoder weights. Intensive empirical validation of the proposed approach shows significant improvement in terms of faster convergence and generalization.

Pairwise Synchrony and Correlations Depend on the Structure of the Population Code in Visual Cortex

In visual areas of primates, neurons activate in parallel while the animal is engaged in a behavioral task. In this study, we examine the structure of the population code while the animal performs delayed match-to-sample tasks on complex natural images. The macaque monkeys visualized two consecutive stimuli that were either the same or different, while being recorded with laminar arrays across the cortical depth in cortical areas V1 and V4. We decode correct choice behavior from neural populations of simultaneously recorded units. Utilizing decoding weights, we divide neurons into most informative and less informative and show that most informative neurons in V4, but not in V1, are more strongly synchronized, coupled, and correlated than less informative neurons. Because neurons are divided into two coding pools according to their coding preference, in V4, but not in V1, spiking synchrony, coupling, and correlations within the coding pool are stronger than across coding pools.

Plug-and-play control of a brain-computer interface through neural map stabilization.

Brain-computer interfaces (BCIs) enable control of assistive devices in individuals with severe motor impairments. A limitation of BCIs that has hindered real-world adoption is poor long-term reliability and lengthy daily recalibration times. To develop methods that allow stable performance without recalibration, we used a 128-channel chronic electrocorticography (ECoG) implant in a paralyzed individual, which allowed stable monitoring of signals. We show that long-term closed-loop decoder adaptation, in which decoder weights are carried across sessions over multiple days, results in consolidation of a neural map and 'plug-and-play' control. In contrast, daily reinitialization led to degradation of performance with variable relearning. Consolidation also allowed the addition of control features over days, that is, long-term stacking of dimensions. Our results offer an approach for reliable, stable BCI control by leveraging the stability of ECoG interfaces and neural plasticity.

An Autuencoder-based Data Augmentation Strategy for Generalization Improvement of DCNNs

Inspired by the phenomenon that the decoding weights of a well-trained autoencoder contain the information of the training samples, we proposed a data augmentation method by utilizing the decoding weights. Given a batch of training data, the autoencoder is trained and the decoding weights are activated; the decoding weights are then combined with the raw samples to generate augmented samples. Furthermore, we probe its working mechanism in three ways: (i) we prove that the decoding weights and the raw samples are of linear relationship under the transformation of a certain invertible function; (ii) the proposed method can sample from a larger range in both feature dimensions and label dimension, which can be interpreted as a broader distribution vicinity compared with those by other approaches; (iii) the model trained with our data augmentation approach has better representation capability, which is reflected by the higher Fisher’s criteria value in deep feature space. We conduct extensive experiments on image and tabular dataset with multiple network architectures. The proposed method provides significant generalization performance improvement compared with the baseline and better or comparable performance compared with the other state-of-the-art data augmentation approaches.

An adaptive decoder design based on the receding horizon optimization in BMI system.

In a motor brain-machine interface system, since the electroencephalogram signal is changing through out the process of the arm movement, the offline trained decoder with fixed weights is often unable to convert the electroencephalogram signal accurately, resulting in poor recovery of joint motor function. In this paper, a receding horizon optimization strategy is chosen to online update the decoder weights and design an adaptive Wiener-filter-based decoder. Firstly, a classical Wiener-filter-based decoder with fixed weights is brief reviewed. Secondly, the weights in Wiener-filter-based decoder are updated by minimizing the cost function, which is composed by the sum of squared position errors in the given horizon at each sampling time. The simulation shows that the recovery effect of joint motor function and neuron activity in the BMI system with the adaptive decoder are both better than that in the BMI system with the fixed decoder.

EEG decoding of the target speaker in a cocktail party scenario: considerations regarding dynamic switching of talker location

Objective. It has been shown that attentional selection in a simple dichotic listening paradigm can be decoded offline by reconstructing the stimulus envelope from single-trial neural response data. Here, we test the efficacy of this approach in an environment with non-stationary talkers. We then look beyond the envelope reconstructions themselves and consider whether incorporating the decoder values—which reflect the weightings applied to the multichannel EEG data at different time lags and scalp locations when reconstructing the stimulus envelope—can improve decoding performance. Approach. High-density EEG was recorded as subjects attended to one of two talkers. The two speech streams were filtered using HRTFs, and the talkers were alternated between the left and right locations at varying intervals to simulate a dynamic environment. We trained spatio-temporal decoders mapping from EEG data to the attended and unattended stimulus envelopes. We then decoded auditory attention by (1) using the attended decoder to reconstruct the envelope and (2) exploiting the fact that decoder weightings themselves contain signatures of attention, resulting in consistent patterns across subjects that can be classified. Main results. The previously established decoding approach was found to be effective even with non-stationary talkers. Signatures of attentional selection and attended direction were found in the spatio-temporal structure of the decoders and were consistent across subjects. The inclusion of decoder weights into the decoding algorithm resulted in significantly improved decoding accuracies (from 61.07% to 65.31% for 4 s windows). An attempt was made to include alpha power lateralization as another feature to improve decoding, although this was unsuccessful at the single-trial level. Significance. This work suggests that the spatial-temporal decoder weights can be utilised to improve decoding. More generally, looking beyond envelope reconstruction and incorporating other signatures of attention is an avenue that should be explored to improve selective auditory attention decoding.

Weighting neurons by selectivity produces near-optimal population codes

Perception is produced by "reading out" the representation of a sensory stimulus contained in the activity of a population of neurons. To examine experimentally how populations code information, a common approach is to decode a linearly weighted sum of the neurons' spike counts. This approach is popular because of the biological plausibility of weighted, nonlinear integration. For neurons recorded in vivo, weights are highly variable when derived through optimization methods, but it is unclear how the variability affects decoding performance in practice. To address this, we recorded from neurons in the middle temporal area (MT) of anesthetized marmosets (Callithrix jacchus) viewing stimuli comprising a sheet of dots that moved coherently in 1 of 12 different directions. We found that high peak response and direction selectivity both predicted that a neuron would be weighted more highly in an optimized decoding model. Although learned weights differed markedly from weights chosen according to a priori rules based on a neuron's tuning profile, decoding performance was only marginally better for the learned weights. In the models with a priori rules, selectivity is the best predictor of weighting, and defining weights according to a neuron's preferred direction and selectivity improves decoding performance to very near the maximum level possible, as defined by the learned weights. NEW & NOTEWORTHY We examined which aspects of a neuron's tuning account for its contribution to sensory coding. Strongly direction-selective neurons are weighted most highly by optimal decoders trained to discriminate motion direction. Models with predefined decoding weights demonstrate that this weighting scheme causally improved direction representation by a neuronal population. Optimizing decoders (using a generalized linear model or Fisher's linear discriminant) led to only marginally better performance than decoders based purely on a neuron's preferred direction and selectivity.

Neuronal Correlations in MT and MST Impair Population Decoding of Opposite Directions of Random Dot Motion.

The study of neuronal responses to random-dot motion patterns has provided some of the most valuable insights into how the activity of neurons is related to perception. In the opposite directions of motion paradigm, the motion signal strength is decreased by manipulating the coherence of random dot patterns to examine how well the activity of single neurons represents the direction of motion. To extend this paradigm to populations of neurons, studies have used modelling based on data from pairs of neurons, but several important questions require further investigation with larger neuronal datasets. We recorded neuronal populations in the middle temporal (MT) and medial superior temporal (MST) areas of anaesthetized marmosets with electrode arrays, while varying the coherence of random dot patterns in two opposite directions of motion (left and right). Using the spike rates of simultaneously recorded neurons, we decoded the direction of motion at each level of coherence with linear classifiers. We found that the presence of correlations had a detrimental effect to decoding performance, but that learning the correlation structure produced better decoding performance compared to decoders that ignored the correlation structure. We also found that reducing motion coherence increased neuronal correlations, but decoders did not need to be optimized for each coherence level. Finally, we showed that decoder weights depend of left-right selectivity at 100% coherence, rather than the preferred direction. These results have implications for understanding how the information encoded by populations of neurons is affected by correlations in spiking activity.

Dissociation of Self-Motion and Object Motion by Linear Population Decoding That Approximates Marginalization.

We use visual image motion to judge the movement of objects, as well as our own movements through the environment. Generally, image motion components caused by object motion and self-motion are confounded in the retinal image. Thus, to estimate heading, the brain would ideally marginalize out the effects of object motion (or vice versa), but little is known about how this is accomplished neurally. Behavioral studies suggest that vestibular signals play a role in dissociating object motion and self-motion, and recent computational work suggests that a linear decoder can approximate marginalization by taking advantage of diverse multisensory representations. By measuring responses of MSTd neurons in two male rhesus monkeys and by applying a recently-developed method to approximate marginalization by linear population decoding, we tested the hypothesis that vestibular signals help to dissociate self-motion and object motion. We show that vestibular signals stabilize tuning for heading in neurons with congruent visual and vestibular heading preferences, whereas they stabilize tuning for object motion in neurons with discrepant preferences. Thus, vestibular signals enhance the separability of joint tuning for object motion and self-motion. We further show that a linear decoder, designed to approximate marginalization, allows the population to represent either self-motion or object motion with good accuracy. Decoder weights are broadly consistent with a readout strategy, suggested by recent computational work, in which responses are decoded according to the vestibular preferences of multisensory neurons. These results demonstrate, at both single neuron and population levels, that vestibular signals help to dissociate self-motion and object motion.SIGNIFICANCE STATEMENT The brain often needs to estimate one property of a changing environment while ignoring others. This can be difficult because multiple properties of the environment may be confounded in sensory signals. The brain can solve this problem by marginalizing over irrelevant properties to estimate the property-of-interest. We explore this problem in the context of self-motion and object motion, which are inherently confounded in the retinal image. We examine how diversity in a population of multisensory neurons may be exploited to decode self-motion and object motion from the population activity of neurons in macaque area MSTd.

Spatial co-adaptation of cortical control columns in a micro-ECoG brain–computer interface

Objective. Electrocorticography (ECoG) has been used for a range of applications including electrophysiological mapping, epilepsy monitoring, and more recently as a recording modality for brain–computer interfaces (BCIs). Studies that examine ECoG electrodes designed and implanted chronically solely for BCI applications remain limited. The present study explored how two key factors influence chronic, closed-loop ECoG BCI: (i) the effect of inter-electrode distance on BCI performance and (ii) the differences in neural adaptation and performance when fixed versus adaptive BCI decoding weights are used. Approach. The amplitudes of epidural micro-ECoG signals between 75 and 105 Hz with 300 μm diameter electrodes were used for one-dimensional and two-dimensional BCI tasks. The effect of inter-electrode distance on BCI control was tested between 3 and 15 mm. Additionally, the performance and cortical modulation differences between constant, fixed decoding using a small subset of channels versus adaptive decoding weights using the entire array were explored. Main results. Successful BCI control was possible with two electrodes separated by 9 and 15 mm. Performance decreased and the signals became more correlated when the electrodes were only 3 mm apart. BCI performance in a 2D BCI task improved significantly when using adaptive decoding weights (80%–90%) compared to using constant, fixed weights (50%–60%). Additionally, modulation increased for channels previously unavailable for BCI control under the fixed decoding scheme upon switching to the adaptive, all-channel scheme. Significance. Our results clearly show that neural activity under a BCI recording electrode (which we define as a ‘cortical control column’) readily adapts to generate an appropriate control signal. These results show that the practical minimal spatial resolution of these control columns with micro-ECoG BCI is likely on the order of 3 mm. Additionally, they show that the combination and interaction between neural adaptation and machine learning are critical to optimizing ECoG BCI performance.

Conservativeness of Untied Auto-Encoders

We discuss necessary and sufficient conditions for an auto-encoder to define a conservative vector field, in which case it is associated with anenergy function akin to the unnormalized log-probability of the data.We show that the conditions for conservativeness are more general than for encoder and decoder weights to be the same ("tied weights''), and thatthey also depend on the form of the hidden unit activation functions.Moreover, we show that contractive training criteria, such as denoising, enforces these conditions locally.Based on these observations, we show how we can use auto-encoders to extract the conservative component of a vector field.

Decoding multiscale word and category-specific spatiotemporal representations from intracranial EEG

Decoding multiscale word and category-specific spatiotemporal representations from intracranial EEG

Naïve coadaptive cortical control

The ability to control a prosthetic device directly from the neocortex has been demonstrated in rats, monkeys and humans. Here we investigate whether neural control can be accomplished in situations where (1) subjects have not received prior motor training to control the device (naïve user) and (2) the neural encoding of movement parameters in the cortex is unknown to the prosthetic device (naïve controller). By adopting a decoding strategy that identifies and focuses on units whose firing rate properties are best suited for control, we show that naïve subjects mutually adapt to learn control of a neural prosthetic system. Six untrained Long-Evans rats, implanted with silicon micro-electrodes in the motor cortex, learned cortical control of an auditory device without prior motor characterization of the recorded neural ensemble. Single- and multi-unit activities were decoded using a Kalman filter to represent an audio ‘cursor’ (90 ms tone pips ranging from 250 Hz to 16 kHz) which subjects controlled to match a given target frequency. After each trial, a novel adaptive algorithm trained the decoding filter based on correlations of the firing patterns with expected cursor movement. Each behavioral session consisted of 100 trials and began with randomized decoding weights. Within 7 ± 1.4 (mean ± SD) sessions, all subjects were able to significantly score above chance (P < 0.05, randomization method) in a fixed target paradigm. Training lasted 24 sessions in which both the behavioral performance and signal to noise ratio of the peri-event histograms increased significantly (P < 0.01, ANOVA). Two rats continued training on a more complex task using a bilateral, two-target control paradigm. Both subjects were able to significantly discriminate the target tones (P < 0.05, Z-test), while one subject demonstrated control above chance (P < 0.05, Z-test) after 12 sessions and continued improvement with many sessions achieving over 90% correct targets. Dynamic analysis of binary trial responses indicated that early learning for this subject occurred during session 6. This study demonstrates that subjects can learn to generate neural control signals that are well suited for use with external devices without prior experience or training.