Predicting the neural response to natural images in the visual cortex requires extracting relevant features from the images and relating those feature to the observed responses. In this work, we optimize the feature extraction in order to maximize the information shared between the image features and the neural response across voxels in a given region of interest (ROI) extracted from the BOLD signal measured by fMRI. We adapt contrastive learning (CL) to fine-tune a convolutional neural network, which was pretrained for image classification, such that a mapping of a given image's features are more similar to the corresponding fMRI response than to the responses to other images. We exploit the recently released Natural Scenes Dataset (Allen et al., 2022) as organized for the Algonauts Project (Gifford et al., 2023), which contains the high-resolution fMRI responses of eight subjects to tens of thousands of naturalistic images. We show that CL fine-tuning creates feature extraction models that enable higher encoding accuracy in early visual ROIs as compared to both the pretrained network and a baseline approach that uses a regression loss at the output of the network to tune it for fMRI response encoding. We investigate inter-subject transfer of the CL fine-tuned models, including subjects from another, lower-resolution dataset (Gong et al., 2023). We also pool subjects for fine-tuning to further improve the encoding performance. Finally, we examine the performance of the fine-tuned models on common image classification tasks, explore the landscape of ROI-specific models by applying dimensionality reduction on the Bhattacharya dissimilarity matrix created using the predictions on those tasks (Mao et al., 2024), and investigate lateralization of the processing for early visual ROIs using salience maps of the classifiers built on the CL-tuned models.

Since neural networks can make wrong predictions even with high confidence, monitoring their behavior at runtime is important, especially in safety-critical domains like autonomous driving. In this paper, we combine ideas from previous monitoring approaches based on observing the activation values of hidden neurons. In particular, we combine the Gaussian-based approach, which observes whether the current value of each monitored neuron is similar to typical values observed during training, and the Outside-the-Box monitor, which creates clusters of the acceptable activation values, and, thus, considers the correlations of the neurons' values. Our experiments evaluate the achieved improvement.

Learning low-dimensional numerical representations from symbolic data, e.g., embedding the nodes of a graph into a geometric space, is an important concept in machine learning. While embedding into Euclidean space is common, recent observations indicate that hyperbolic geometry is better suited to represent hierarchical information and heterogeneous data (e.g., graphs with a scale-free degree distribution). Despite their potential for more accurate representations, hyperbolic embeddings also have downsides like being more difficult to compute and harder to use in downstream tasks. We propose embedding into a weighted space, which is closely related to hyperbolic geometry but mathematically simpler. We provide the embedding algorithm WEmbed and demonstrate, based on generated as well as over 2000 real-world graphs, that our weighted embeddings heavily outperform state-of-the-art Euclidean embeddings for heterogeneous graphs while using fewer dimensions. The running time of WEmbed and embedding quality for the remaining instances is on par with state-of-the-art Euclidean embedders.

Many recent developments for robots to represent environments have focused on photorealistic reconstructions. This paper particularly focuses on generating sequences of images from the photorealistic Gaussian Splatting models, that match instructions that are given by user-inputted language. We contribute a novel framework, SplaTraj, which formulates the generation of images within photorealistic environment representations as a continuous-time trajectory optimization problem. Costs are designed so that a camera following the trajectory poses will smoothly traverse through the environment and render the specified spatial information in a photogenic manner. This is achieved by querying a photorealistic representation with language embedding to isolate regions that correspond to the user-specified inputs. These regions are then projected to the camera's view as it moves over time and a cost is constructed. We can then apply gradient-based optimization and differentiate through the rendering to optimize the trajectory for the defined cost. The resulting trajectory moves to photogenically view each of the specified objects. We empirically evaluate our approach on a suite of environments and instructions, and demonstrate the quality of generated image sequences.

The increasing size of deep learning models has created the need for more efficient alternatives to the standard error backpropagation algorithm, that make better use of asynchronous, parallel and distributed computing. One major shortcoming of backpropagation is the interlocking between the forward phase of the algorithm, which computes a global loss, and the backward phase where the loss is backpropagated through all layers to compute the gradients, which are used to update the network parameters. To address this problem, we propose a method that parallelises SGD updates across the layers of a model by asynchronously updating them from multiple threads. Furthermore, since we observe that the forward pass is often much faster than the backward pass, we use separate threads for the forward and backward pass calculations, which allows us to use a higher ratio of forward to backward threads than the usual 1:1 ratio, reducing the overall staleness of the parameters. Thus, our approach performs asynchronous stochastic gradient descent using separate threads for the loss (forward) and gradient (backward) computations and performs layer-wise partial updates to parameters in a distributed way. We show that this approach yields close to state-of-the-art results while running up to 2.97 faster than Hogwild! We theoretically prove the convergence of the algorithm using a novel theoretical framework based on stochastic differential equations and the drift diffusion process, by modeling the asynchronous parameter updates as a stochastic process. Scaling up modern deep learning models requires massive resources and training time.

As training datasets grow larger, we aspire to develop models that generalize well to any diverse test distribution, even if the latter deviates significantly from the training data. Various approaches like domain adaptation, domain generalization, and robust optimization attempt to address the out-of-distribution challenge by posing assumptions about the relation between training and test distribution. Differently, we adopt a more conservative perspective by accounting for the worst-case error across all sufficiently diverse test distributions within a known domain. Our first finding is that training on a uniform distribution over this domain is optimal. We also interrogate practical remedies when uniform samples are unavailable by considering methods for mitigating non-uniformity through finetuning and rebalancing. Our theory provides a mathematical grounding for previous observations on the role of entropy and rebalancing for o.o.d. generalization and foundation model training. We also provide new empirical evidence across tasks involving o.o.d. shifts which illustrate the broad applicability of our perspective.

--Audio-visual speaker tracking aims to determine the location of human targets in a scene using signals captured by a multi-sensor platform, whose accuracy and robustness can be improved by multi-modal fusion methods. Recently, several fusion methods have been proposed to model the correlation in multiple modalities. However, for the speaker tracking problem, the cross-modal interaction between audio and visual signals hasn't been well exploited. T o this end, we present a novel Speaker Tracking Network (STNet) with a deep audio-visual fusion model in this work. We design a visual-guided acoustic measurement method to fuse heterogeneous cues in a unified localization space, which employs visual observations via a camera model to construct the enhanced acoustic map. For feature fusion, a cross-modal attention module is adopted to jointly model multi-modal contexts and interactions. The correlated information between audio and visual features is further interacted in the fusion model. Moreover, the STNet-based tracker is applied to multi-speaker cases by a quality-aware module, which evaluates the reliability of multi-modal observations to achieve robust tracking in complex scenarios. Experiments on the A V16.3 and CA V3D datasets show that the proposed STNet-based tracker outperforms uni-modal methods and state-of-the-art audio-visual speaker trackers. PEAKER tracking is a fundamental task in human-computer interaction that determines the position of the speaker in each time step by analyzing data from sensors such as microphones and cameras [1]. It has wide applications in intelligent surveillance [2], multimedia systems [3], and robot navigation [4]. In general, the basic approaches for solving the tracking problem include computer vision-based face or body tracking methods [5-7] and auditory-based Sound Source Localization (SSL) methods [8, 9]. However, it is difficult for uni-modal methods to adapt to complex dynamic environments. For example, visual trackers are susceptible to object occlusion and changes in illumination and appearance. Besides, acoustic tracking is not subject to visual interference, but the intermittent nature of speech signals, background noise, and room reverberation constrain the performance of SSL-based trackers. This work is supported by National Natural Science Foundation of China (No. 62403345).

We present the first mini-batch kernel $k$-means algorithm, offering an order of magnitude improvement in running time compared to the full batch algorithm. A single iteration of our algorithm takes $\widetilde{O}(kb^2)$ time, significantly faster than the $O(n^2)$ time required by the full batch kernel $k$-means, where $n$ is the dataset size and $b$ is the batch size. Extensive experiments demonstrate that our algorithm consistently achieves a 10-100x speedup with minimal loss in quality, addressing the slow runtime that has limited kernel $k$-means adoption in practice. We further complement these results with a theoretical analysis under an early stopping condition, proving that with a batch size of $\widetilde{\Omega}(\max \{\gamma^{4}, \gamma^{2}\} \cdot \epsilon^{-2})$, the algorithm terminates in $O(\gamma^2/\epsilon)$ iterations with high probability, where $\gamma$ bounds the norm of points in feature space and $\epsilon$ is a termination threshold. Our analysis holds for any reasonable center initialization, and when using $k$-means++ initialization, the algorithm achieves an approximation ratio of $O(\log k)$ in expectation. For normalized kernels, such as Gaussian or Laplacian it holds that $\gamma=1$. Taking $\epsilon = O(1)$ and $b=\Theta(\log n)$, the algorithm terminates in $O(1)$ iterations, with each iteration running in $\widetilde{O}(k)$ time.

Lung radiotherapy treatment systems are subject to a latency that leads to uncertainty in the estimated tumor location and high irradiation of healthy tissue. This work addresses future frame prediction in chest dynamic MRI sequences to compensate for that delay using RNNs trained with online learning algorithms. The latter enable networks to mitigate irregular movements, as they update synaptic weights with each new training example. Experiments were conducted using four publicly available 2D thoracic cine-MRI sequences. PCA decomposes the time-varying deformation vector field (DVF), computed with the Lucas-Kanade optical flow algorithm, into static deformation fields and low-dimensional time-dependent weights. We compare various algorithms to forecast the latter: linear regression, least mean squares (LMS), and RNNs trained with real-time recurrent learning (RTRL), unbiased online recurrent optimization, decoupled neural interfaces and sparse 1-step approximation (SnAp-1). That enables estimating the future DVFs and, in turn, the next frames by warping the initial image. Linear regression led to the lowest mean DVF error at a horizon h = 0.32s (the time interval in advance for which the prediction is made), equal to 1.30mm, followed by SnAp-1 and RTRL, whose error increased from 1.37mm to 1.44mm as h increased from 0.62s to 2.20s. Similarly, the structural similarity index measure (SSIM) of LMS decreased from 0.904 to 0.898 as h increased from 0.31s to 1.57s and was the highest among the algorithms compared for the latter horizons. SnAp-1 attained the highest SSIM for h $\geq$ 1.88s, with values of less than 0.898. The predicted images look similar to the original ones, and the highest errors occurred at challenging areas such as the diaphragm boundary at the end-of-inhale phase, where motion variability is more prominent, and regions where out-of-plane motion was more prevalent.

The underlying mechanism by which simple gradient-based iterative algorithms can effectively handle the non-convex problem of deep model training remains incompletely understood within the traditional convex and non-convex analysis frameworks, which often require the Lipschitz smoothness of the gradient and strong convexity. In this paper, we introduce $\mathcal{H}(\phi)$-convexity and $\mathcal{H}(\Phi)$-smoothness, which broaden the existing concepts of smoothness and convexity, and delineate their fundamental properties. Building on these concepts, we introduce the high-order gradient descent and high-order stochastic gradient descent methods, which serve as extensions to the traditional gradient descent and stochastic gradient descent methods, respectively. Furthermore, we establish descent lemmas for the $\mathcal{H}(\phi)$-convex and $\mathcal{H}(\Phi)$-smooth objective functions when utilizing these four methods. On the basis of these findings, we develop the gradient structure control algorithm to address non-convex optimization objectives, encompassing both the functions represented by machine learning models and common loss functions in deep learning. The effectiveness of the proposed methodology is empirically validated through experiments.