Hidden representations in deep neural networks: Part 2. Regression problems

We explore how DNNs can be used to develop a computational understanding of individual differences in high-level visual cognition given their ability to generate rich meaningful object representations informed by their architecture, experience, and training protocols. As a first step to quantifying individual differences in DNN representations, we systematically explored the robustness of a variety of representational similarity measures: Representational Similarity Analysis (RSA), Centered Kernel Alignment (CKA), and Projection-Weighted Canonical Correlation Analysis (PWCCA), with an eye to how these measures are used in cognitive science, cognitive neuroscience, and vision science. To manipulate object representations, we next created a large set of models varying in random initial weights and random training image order, training image frequencies, training category frequencies, and model size and architecture and measured the representational variation caused by each manipulation. We examined both small (All-CNN-C) and commonly-used large (VGG and ResNet) DNN architectures. To provide a comparison for the magnitude of representational differences, we established a baseline based on the representational variation caused by image-augmentation techniques used to train those DNNs. We found that variation in model randomization and model size never exceeded baseline. By contrast, differences in training image frequency and training category frequencies caused representational variation that exceeded baseline, with training category frequency manipulations exceeding baseline earlier in the networks. These findings provide insights into the magnitude of representational variations that can be expected with a range of manipulations and provide a springboard for further exploration of systematic model variations aimed at modeling individual differences in high-level visual cognition.

Transformability, generalizability, but limited diffusibility: Comparing global vs. task-specific language representations in deep neural networks

In the human brain, the incoming light to the retina is transformed into meaningful representations that allow us to interact with the world. In a similar vein, the RGB pixel values are transformed by a deep neural network (DNN) into meaningful representations relevant to solving a computer vision task it was trained for. Therefore, in my research, I aim to reveal insights into the visual representations in the human visual cortex and DNNs solving vision tasks. In the previous decade, DNNs have emerged as the state-of-the-art models for predicting neural responses in the human and monkey visual cortex. Research has shown that training on a task related to a brain region’s function leads to better predictivity than a randomly initialized network. Based on this observation, we proposed that we can use DNNs trained on different computer vision tasks to identify functional mapping of the human visual cortex. To validate our proposed idea, we first investigate a brain region occipital place area (OPA) using DNNs trained on scene parsing task and scene classification task. From the previous investigations about OPA’s functions, we knew that it encodes navigational affordances that require spatial information about the scene. Therefore, we hypothesized that OPA’s representation should be closer to a scene parsing model than a scene classification model as the scene parsing task explicitly requires spatial information about the scene. Our results showed that scene parsing models had representation closer to OPA than scene classification models thus validating our approach. We then selected multiple DNNs performing a wide range of computer vision tasks ranging from low-level tasks such as edge detection, 3D tasks such as surface normals, and semantic tasks such as semantic segmentation. We compared the representations of these DNNs with all the regions in the visual cortex, thus revealing the functional representations of different regions of the visual cortex. Our results highly converged with previous investigations of these brain regions validating the feasibility of the proposed approach in finding functional representations of the human brain. Our results also provided new insights into underinvestigated brain regions that can serve as starting hypotheses and promote further investigation into those brain regions. We applied the same approach to find representational insights about the DNNs. A DNN usually consists of multiple layers with each layer performing a computation leading to the final layer that performs prediction for a given task. Training on different tasks could lead to very different representations. Therefore, we first investigate at which stage does the representation in DNNs trained on different tasks starts to differ. We further investigate if the DNNs trained on similar tasks lead to similar representations and on dissimilar tasks lead to more dissimilar representations. We selected the same set of DNNs used in the previous work that were trained on the Taskonomy dataset on a diverse range of 2D, 3D and semantic tasks. Then, given a DNN trained on a particular task, we compared the representation of multiple layers to corresponding layers in other DNNs. From this analysis, we aimed to reveal where in the network architecture task-specific representation is prominent. We found that task specificity increases as we go deeper into the DNN architecture and similar tasks start to cluster in groups. We found that the grouping we found using representational similarity was highly correlated with grouping based on transfer learning thus creating an interesting application of the approach to model selection in transfer learning. During previous works, several new measures were introduced to compare DNN representations. So, we identified the commonalities in different measures and unified different measures into a single framework referred to as duality diagram similarity. This work opens up new possibilities for similarity measures to understand DNN representations. While demonstrating a much higher correlation with transfer learning than previous state-of-the-art measures we extend it to understanding layer-wise representations of models trained on the Imagenet and Places dataset using different tasks and demonstrate its applicability to layer selection for transfer learning. In all the previous works, we used the task-specific DNN representations to understand the representations in the human visual cortex and other DNNs. We were able to interpret our findings in terms of computer vision tasks such as edge detection, semantic segmentation, depth estimation, etc. however we were not able to map the representations to human interpretable concepts. Therefore in our most recent work, we developed a new method that associates individual artificial neurons with human interpretable concepts. Overall, the works in this thesis revealed new insights into the representation of the visual cortex and DNNs...

At the core of any inference procedure, deep neural networks are dot product operations, which are the component that requires the highest computational resources. For instance, deep neural networks, such as VGG-16, require up to 15-G operations in order to perform the dot products present in a single forward pass, which results in significant energy consumption and thus limits their use in resource-limited environments, e.g., on embedded devices or smartphones. One common approach to reduce the complexity of the inference is to prune and quantize the weight matrices of the neural network. Usually, this results in matrices whose entropy values are low, as measured relative to the empirical probability mass distribution of its elements. In order to efficiently exploit such matrices, one usually relies on, inter alia, sparse matrix representations. However, most of these common matrix storage formats make strong statistical assumptions about the distribution of the elements; therefore, cannot efficiently represent the entire set of matrices that exhibit low-entropy statistics (thus, the entire set of compressed neural network weight matrices). In this paper, we address this issue and present new efficient representations for matrices with low-entropy statistics. Alike sparse matrix data structures, these formats exploit the statistical properties of the data in order to reduce the size and execution complexity. Moreover, we show that the proposed data structures can not only be regarded as a generalization of sparse formats but are also more energy and time efficient under practically relevant assumptions. Finally, we test the storage requirements and execution performance of the proposed formats on compressed neural networks and compare them to dense and sparse representations. We experimentally show that we are able to attain up to ×42 compression ratios, ×5 speed ups, and ×90 energy savings when we lossless convert the state-of-the-art networks, such as AlexNet, VGG-16, ResNet152, and DenseNet, into the new data structures and benchmark their respective dot product.

There is an important challenge in systematically interpreting the internal representations of deep neural networks (DNNs). Existing techniques are often less effective for non-tabular tasks, or they primarily focus on qualitative, ad-hoc interpretations of models. In response, this study introduces a cognitive science-inspired, multi-dimensional quantification and visualization approach that captures two temporal dimensions of model learning: the "information-processing trajectory" and the "developmental trajectory." The former represents the influence of incoming signals on an agent's decision-making, while the latter conceptualizes the gradual improvement in an agent's performance throughout its lifespan. Tracking the learning curves of DNNs enables researchers to explicitly identify the model appropriateness of a given task, examine the properties of the underlying input signals, and assess the model's alignment (or lack thereof) with human learning experiences. To illustrate this method, we conducted 750 runs of simulations on two temporal tasks: gesture detection and sentence classification, showcasing its applicability across different types of deep learning tasks. Using four descriptive metrics to quantify the mapped learning curves-start, end - start, max, tmax-, we identified significant differences in learning patterns based on data sources and class distinctions (all p's < .0001), the prominent role of spatial semantics in gesture learning, and larger information gains in language learning. We highlight three key insights gained from mapping learning curves: non-monotonic progress, pairwise comparisons, and domain distinctions. We reflect on the theoretical implications of this method for cognitive processing, language models and representations from multiple modalities.

Product Quantization (PQ) is one of the most popular Approximate Nearest Neighbor (ANN) methods for large-scale image retrieval, bringing better performance than hashing based methods. In recent years, several works extend the hard quantization to soft quantization with specially designed deep neural architectures. We propose a simple but effective deep Product Quantization Module (PQM) to jointly learn discriminative codebook and precise hard assignment in an end-to-end manner. In this work, we use the straight-through estimator to make it feasible to directly optimize the discrete binary representations in deep neural networks with stochastic gradient descent. Different from previous deep vector quantization methods, PQM is a plug-and-play module which can be adaptive to various base networks in the scenarios of image search or compression. Besides, we propose a reconstruction loss to minimize the domain gap between the original embedding features and codebook. Experimental results show that PQM outperforms state-of-the-art deep supervised hashing and quantization methods on several image retrieval benchmarks.

Automated image classification is an essential task of the computer vision field. The tagging of images into a set of predefined groups is referred to as image classification. The implementation of computer vision to automate image classification would be beneficial because manual image evaluation and identification can be time-consuming, particularly when there are many images of different classes. Deep learning approaches are proven to overperform existing machine learning techniques in many fields in recent years, and computer vision is one of the most notable examples. The very deep neural network (VDCNN) is a powerful deep learning model for image classification, and this paper examines it briefly using MNIST handwritten digit dataset. This dataset is used to prove the efficacy of very deep neural networks over other deep learning models. The proposed study aims to comprehend the very deep neural network architecture used to accomplish a handwritten digit recognition task. The feasibility of the proposed model is evaluated using mean accuracy, validation accuracy, and standard deviation. The study results of the very deep neural network model are compared to a convolutional neural network and convolutional neural network with batch normalization. According to the results of the comparison study, very deep neural networks achieve high accuracy of 99.1% for a handwritten dataset. The outcome of the proposed work is used to interpret how well a very deep neural network performs when compared to the other two models of deep neural networks. This proposed architecture may be used to automate the classification of handwritten digits dataset.

Real-time electricity price forecasting of wind farms with deep neural network transfer learning and hybrid datasets

Visual features versus categories: Explaining object representations in primate IT and deep neural networks with weighted representational modeling

Analysis and classification of sequences is one of the key research areas in bioinformatics. The basic tool for sequence analysis is alignment, but there are also other techniques that can be used. Frequency Chaos Game Representation is a technique that builds an image characteristic of the sequence The paper describes the first experiment in the use of a deep neural network for classification of DNA sequences represented as images by using the Frequency Chaos Game Representation.

Deep Learning algorithms are designed in such a way that they mimic the function of the human cerebral cortex. These algorithms are representations of deep neural networks i.e. neural networks with many hidden layers. Convolutional neural networks are deep learning algorithms that can train large datasets with millions of parameters, in form of 2D images as input and convolve it with filters to produce the desired outputs. In this article, CNN models are built to evaluate its performance on image recognition and detection datasets. The algorithm is implemented on MNIST and CIFAR-10 dataset and its performance are evaluated. The accuracy of models on MNIST is 99.6 %, CIFAR-10 is using real-time data augmentation and dropout on CPU unit.

Multi-model ensemble is an important fundamental technique of practical value for many artificial intelligence applications. However, the usage for multi-model ensemble has been limited when it is combined with deep neural networks to construct ensemble of deep neural networks. Due to the big time and computing resources required to train and to integrate multiple deep neural networks for the achievement of multi-model ensemble, the engineering application field where developing time and computing resources are usually restricted, has not yet widespreadly benefited from ensemble of deep neural networks. To alleviate this situation, we present a new multi-model ensemble approach entitled feature transferring based multi-model ensemble (FTBME), for ensemble of deep neural networks. Primarily, we propose a feature transferring based multi-model training strategy to more affordably find multiple extra models based on a given previously optimized deep neural network model. Sequentially, to develop better ensemble solutions, we design a more effective random greedy based ensemble selection strategy to filter out models non-positive to ensemble generalization. Finally, inspired by the idea of averaging parameter points, we propose to fuse the obtained models in weight space which eventually reduces the expense of ensemble at the testing stage to a single deep neural network model while retaining the generalization. These three advances constitute the resulting technique FTBME. We conducted extensive experiments using deep neural networks, from light weight to complex, on ImageNet, CIFAR-10 and CIFAR-100. Results show that, given a deep neural network model which has been well-optimized and reaching its limit, FTBME can obtain better generalization with minor extra training expense while maintaining the expense to a single model at ensemble testing. This promising property of FTBME make us believe that it could be leveraged to broaden the usage for ensemble of deep neural networks, alleviating the situation that the engineering application field has not yet widespreadly benefited from ensemble of deep neural networks.

Deep learning is being incorporated in many modern software systems. Deep learning approaches train a deep neural network (DNN) model using training examples, and then use the DNN model for prediction. While the structure of a DNN model as layers is observable, the model is treated in its entirety as a monolithic component. To change the logic implemented by the model, e.g. to add/remove logic that recognizes inputs belonging to a certain class, or to replace the logic with an alternative, the training examples need to be changed and the DNN needs to be retrained using the new set of examples. We argue that decomposing a DNN into DNN modules— akin to decomposing a monolithic software code into modules—can bring the benefits of modularity to deep learning. In this work, we develop a methodology for decomposing DNNs for multi-class problems into DNN modules. For four canonical problems, namely MNIST, EMNIST, FMNIST, and KMNIST, we demonstrate that such decomposition enables reuse of DNN modules to create different DNNs, enables replacement of one DNN module in a DNN with another without needing to retrain. The DNN models formed by composing DNN modules are at least as good as traditional monolithic DNNs in terms of test accuracy for our problems.

Computer-aided diagnosis of ocular disorders like glaucoma can be effectively performed with retinal fundus images. Since the advent of machine learning and later deep learning techniques in medical image analysis, research focusing on automated and early detection of glaucoma has gained prominence. In this paper, we show how a deep Convolutional Neural Network model can be used in visualising the features related to Glaucoma that are consistent with (a subset of) those used by ophthalmologists for diagnosis. However, since the doctors do not necessarily depend entirely on these features but back them up with other important measurements such as visual field testing (HVF), any deep neural network such as CNN may not necessarily be expected to yield 100% accuracy. Further, other important parameters such as Intraocular Pressure (IOP), which may be built into a deep learning model, may not necessarily correlate well with the presence of Glaucoma. In this work, a deep CNN model has been trained and tested on a large number of high-quality fundus images, both normal and glaucomatous. We compare the results with transfer learning models such as VGG16, ResNet50, and MobileNetV2. On an average, we obtained an accuracy of 93.75% in identifying glaucoma, focusing only on the features of the fundus image.

Land cover and its dynamic information is the basis for characterizing surface conditions, supporting land resource management and optimization, and assessing the impacts of climate change and human activities. In land cover information extraction, the traditional convolutional neural network (CNN) method has several problems, such as the inability to be applied to multispectral and hyperspectral satellite imagery, the weak generalization ability of the model and the difficulty of automating the construction of a training database. To solve these problems, this study proposes a new type of deep convolutional neural network based on Landsat-8 Operational Land Imager (OLI) imagery. The network integrates cascaded cross-channel parametric pooling and average pooling layer, applies a hierarchical sampling strategy to realize the automatic construction of the training dataset, determines the technical scheme of model-related parameters, and finally performs the automatic classification of remote sensing images. This study used the new type of deep convolutional neural network to extract land cover information from Qinhuangdao City, Hebei Province, and compared the experimental results with those obtained by traditional methods. The results show that: (1) The proposed deep convolutional neural network (DCNN) model can automatically construct the training dataset and classify images. This model performs the classification of multispectral and hyperspectral satellite images using deep neural networks, which improves the generalization ability of the model and simplifies the application of the model. (2) The proposed DCNN model provides the best classification results in the Qinhuangdao area. The overall accuracy of the land cover data obtained is 82.0%, and the kappa coefficient is 0.76. The overall accuracy is improved by 5% and 14% compared to the support vector machine method and the maximum likelihood classification method, respectively.