Recently, Ainsworth et al. showed that using weight matching (WM) to minimize the $L^2$ distance in a permutation search of model parameters effectively identifies permutations that satisfy linear mode connectivity (LMC), where the loss along a linear path between two independently trained models with different seeds remains nearly constant. This paper analyzes LMC using WM, which is useful for understanding stochastic gradient descent's effectiveness and its application in areas like model merging. We first empirically show that permutations found by WM do not significantly reduce the $L^2$ distance between two models, and the occurrence of LMC is not merely due to distance reduction by WM itself. We then demonstrate that permutations can change the directions of the singular vectors, but not the singular values, of the weight matrices in each layer. This finding shows that permutations found by WM primarily align the directions of singular vectors associated with large singular values across models. This alignment brings the singular vectors with large singular values, which determine the model's functionality, closer between the original and merged models, allowing the merged model to retain functionality similar to the original models, thereby satisfying LMC. This paper also analyzes activation matching (AM) in terms of singular vectors and finds that the principle of AM is likely the same as that of WM. Finally, we analyze the difference between WM and the straight-through estimator (STE), a dataset-dependent permutation search method, and show that WM can be more advantageous than STE in achieving LMC among three or more models.

Building multisensory AI systems that learn from multiple sensory inputs such as text, speech, video, real-world sensors, wearable devices, and medical data holds great promise for impact in many scientific areas with practical benefits, such as in supporting human health and well-being, enabling multimedia content processing, and enhancing real-world autonomous agents. By synthesizing a range of theoretical frameworks and application domains, this thesis aims to advance the machine learning foundations of multisensory AI. In the first part, we present a theoretical framework formalizing how modalities interact with each other to give rise to new information for a task. These interactions are the basic building blocks in all multimodal problems, and their quantification enables users to understand their multimodal datasets, design principled approaches to learn these interactions, and analyze whether their model has succeeded in learning. In the second part, we study the design of practical multimodal foundation models that generalize over many modalities and tasks, which presents a step toward grounding large language models to real-world sensory modalities. We introduce MultiBench, a unified large-scale benchmark across a wide range of modalities, tasks, and research areas, followed by the cross-modal attention and multimodal transformer architectures that now underpin many of today's multimodal foundation models. Scaling these architectures on MultiBench enables the creation of general-purpose multisensory AI systems, and we discuss our collaborative efforts in applying these models for real-world impact in affective computing, mental health, cancer prognosis, and robotics. Finally, we conclude this thesis by discussing how future work can leverage these ideas toward more general, interactive, and safe multisensory AI.

This paper proposes a non-autoregressive extension of our previously proposed sequence-to-sequence (S2S) model-based voice conversion (VC) methods. S2S model-based VC methods have attracted particular attention in recent years for their flexibility in converting not only the voice identity but also the pitch contour and local duration of input speech, thanks to the ability of the encoder-decoder architecture with the attention mechanism. However, one of the obstacles to making these methods work in real-time is the autoregressive (AR) structure. To overcome this obstacle, we develop a method to obtain a model that is free from an AR structure and behaves similarly to the original S2S models, based on a teacher-student learning framework. In our method, called "FastS2S-VC", the student model consists of encoder, decoder, and attention predictor. The attention predictor learns to predict attention distributions solely from source speech along with a target class index with the guidance of those predicted by the teacher model from both source and target speech. Thanks to this structure, the model is freed from an AR structure and allows for parallelization. Furthermore, we show that FastS2S-VC is suitable for real-time implementation based on a sliding-window approach, and describe how to make it run in real-time. Through speaker-identity and emotional-expression conversion experiments, we confirmed that FastS2S-VC was able to speed up the conversion process by 70 to 100 times compared to the original AR-type S2S-VC methods, without significantly degrading the audio quality and similarity to target speech. We also confirmed that the real-time version of FastS2S-VC can be run with a latency of 32 ms when run on a GPU.

Pre-trained models are essential as feature extractors in modern machine learning systems in various domains. In this study, we hypothesize that representations effective for general audio tasks should provide multiple aspects of robust features of the input sound. For recognizing sounds regardless of perturbations such as varying pitch or timbre, features should be robust to these perturbations. For serving the diverse needs of tasks such as recognition of emotions or music genres, representations should provide multiple aspects of information, such as local and global features. To implement our principle, we propose a self-supervised learning method: Bootstrap Your Own Latent (BYOL) for Audio (BYOL-A, pronounced "viola"). BYOL-A pre-trains representations of the input sound invariant to audio data augmentations, which makes the learned representations robust to the perturbations of sounds. Whereas the BYOL-A encoder combines local and global features and calculates their statistics to make the representation provide multi-aspect information. As a result, the learned representations should provide robust and multi-aspect information to serve various needs of diverse tasks. We evaluated the general audio task performance of BYOL-A compared to previous state-of-the-art methods, and BYOL-A demonstrated generalizability with the best average result of 72.4% and the best VoxCeleb1 result of 57.6%. Extensive ablation experiments revealed that the BYOL-A encoder architecture contributes to most performance, and the final critical portion resorts to the BYOL framework and BYOL-A augmentations. Our code is available online at this https URL for future studies.

To reduce the need for skilled clinicians in heart sound interpretation, recent studies on automating cardiac auscultation have explored deep learning approaches. However, despite the demands for large data for deep learning, the size of the heart sound datasets is limited, and no pre-trained model is available. On the contrary, many pre-trained models for general audio tasks are available as general-purpose audio representations. This study explores the potential of general-purpose audio representations pre-trained on large-scale datasets for transfer learning in heart murmur detection. Experiments on the CirCor DigiScope heart sound dataset show that the recent self-supervised learning Masked Modeling Duo (M2D) outperforms previous methods with the results of a weighted accuracy of 0.832 and an unweighted average recall of 0.713. Experiments further confirm improved performance by ensembling M2D with other models. These results demonstrate the effectiveness of general-purpose audio representation in processing heart sounds and open the way for further applications. Our code is available online which runs on a 24 GB consumer GPU at this https URL

Subgraph matching is a compute-intensive problem that asks to enumerate all the isomorphic embeddings of a query graph within a data graph. This problem is generally solved with backtracking, which recursively evolves every possible partial embedding until it becomes an isomorphic embedding or is found unable to become it. While existing methods reduce the search space by analyzing graph structures before starting the backtracking, it is often ineffective for complex graphs. In this paper, we propose an efficient algorithm for subgraph matching that performs on-the-fly pruning during the backtracking. Our main idea is to `learn from failure'. That is, our algorithm generates failure patterns when a partial embedding is found unable to become an isomorphic embedding. Then, in the subsequent process of the backtracking, our algorithm prunes partial embeddings matched with a failure pattern. This pruning does not change the result because failure patterns are designed to represent the conditions that never yield an isomorphic embedding. Additionally, we introduce an efficient representation of failure patterns for constant-time pattern matching. The experimental results show that our method improves the performance by up to 10000 times than existing methods.

In this paper, we propose a non-parallel any-to-many voice conversion (VC) method termed VoiceGrad. Inspired by WaveGrad, a recently introduced novel waveform generation method, VoiceGrad is based upon the concepts of score matching and Langevin dynamics. It uses weighted denoising score matching to train a score approximator, a fully convolutional network with a U-Net structure designed to predict the gradient of the log density of the speech feature sequences of multiple speakers, and performs VC by using annealed Langevin dynamics to iteratively update an input feature sequence towards the nearest stationary point of the target distribution based on the trained score approximator network. Thanks to the nature of this concept, VoiceGrad enables any-to-many VC, a VC scenario in which the speaker of input speech can be arbitrary, and allows for non-parallel training, which requires no parallel utterances or transcriptions.

We propose a non-learning depth completion method for a sparse depth map captured using a light detection and ranging (LiDAR) sensor guided by a pair of stereo images. Generally, conventional stereo-aided depth completion methods have two limiations. (i) They assume the given sparse depth map is accurately aligned to the input image, whereas the alignment is difficult to achieve in practice. (ii) They have limited accuracy in the long range because the depth is estimated by pixel disparity. To solve the abovementioned limitations, we propose selective stereo matching (SSM) that searches the most appropriate depth value for each image pixel from its neighborly projected LiDAR points based on an energy minimization framework. This depth selection approach can handle any type of mis-projection. Moreover, SSM has an advantage in terms of long-range depth accuracy because it directly uses the LiDAR measurement rather than the depth acquired from the stereo. SSM is a discrete process; thus, we apply variational smoothing with binary anisotropic diffusion tensor (B-ADT) to generate a continuous depth map while preserving depth discontinuity across object boundaries. Experimentally, compared with the previous state-of-the-art stereo-aided depth completion, the proposed method reduced the mean absolute error (MAE) of the depth estimation to 0.65 times and demonstrated approximately twice more accurate estimation in the long range. Moreover, under various LiDAR-camera calibration errors, the proposed method reduced the depth estimation MAE to 0.34-0.93 times from previous depth completion methods.

Recent work [4] analyses the local convergence of Adam in a neighbourhood of an optimal solution for a twice-differentiable function. It is found that the learning rate has to be sufficiently small to ensure local stability of the optimal solution. The above convergence results also hold for AdamW. In this work, we propose a new adaptive optimisation method by extending AdamW in two aspects with the purpose to relax the requirement on small learning rate for local stability, which we refer to as Aida. Firstly, we consider tracking the 2nd moment r_t of the pth power of the gradient-magnitudes. r_t reduces to v_t of AdamW when p=2. Suppose {m_t} is the first moment of AdamW. It is known that the update direction m_{t+1}/(v_{t+1}+epsilon)^0.5 (or m_{t+1}/(v_{t+1}^0.5+epsilon) of AdamW (or Adam) can be decomposed as the sign vector sign(m_{t+1}) multiplied elementwise by a vector of magnitudes |m_{t+1}|/(v_{t+1}+epsilon)^0.5 (or |m_{t+1}|/(v_{t+1}^0.5+epsilon)). Aida is designed to compute the qth power of the magnitude in the form of |m_{t+1}|^q/(r_{t+1}+epsilon)^(q/p) (or |m_{t+1}|^q/((r_{t+1})^(q/p)+epsilon)), which reduces to that of AdamW when (p,q)=(2,1). Suppose the origin 0 is a local optimal solution of a twice-differentiable function. It is found theoretically that when q>1 and p>1 in Aida, the origin 0 is locally stable only when the weight-decay is non-zero. Experiments are conducted for solving ten toy optimisation problems and training Transformer and Swin-Transformer for two deep learning (DL) tasks. The empirical study demonstrates that in a number of scenarios (including the two DL tasks), Aida with particular setups of (p,q) not equal to (2,1) outperforms the setup (p,q)=(2,1) of AdamW.

Artificial objects often subjectively look eerie when their appearance to some extent resembles a human, which is known as the uncanny valley phenomenon. From a cognitive psychology perspective, several explanations of the phenomenon have been put forth, two of which are object categorization and realism inconsistency. Recently, MacDorman and Chattopadhyay (2016) reported experimental data as evidence in support of the latter. In our estimation, however, their results are still consistent with categorization-based stranger avoidance. In this Discussions paper, we try to describe why categorization-based stranger avoidance remains a viable explanation, despite the evidence of MacDorman and Chattopadhyay, and how it offers a more inclusive explanation of the impression of eeriness in the uncanny valley phenomenon.

This paper explores the problem of 3D human pose estimation from only low-level acoustic signals. The existing active acoustic sensing-based approach for 3D human pose estimation implicitly assumes that the target user is positioned along a line between loudspeakers and a microphone. Because reflection and diffraction of sound by the human body cause subtle acoustic signal changes compared to sound obstruction, the existing model degrades its accuracy significantly when subjects deviate from this line, limiting its practicality in real-world scenarios. To overcome this limitation, we propose a novel method composed of a position discriminator and reverberation-resistant model. The former predicts the standing positions of subjects and applies adversarial learning to extract subject position-invariant features. The latter utilizes acoustic signals before the estimation target time as references to enhance robustness against the variations in sound arrival times due to diffraction and reflection. We construct an acoustic pose estimation dataset that covers diverse human locations and demonstrate through experiments that our proposed method outperforms existing approaches.

We make contributions towards improving adaptive-optimizer performance. Our improvements are based on suppression of the range of adaptive stepsizes in the AdaBelief optimizer. Firstly, we show that the particular placement of the parameter epsilon within the update expressions of AdaBelief reduces the range of the adaptive stepsizes, making AdaBelief closer to SGD with momentum. Secondly, we extend AdaBelief by further suppressing the range of the adaptive stepsizes. To achieve the above goal, we perform mutual layerwise vector projections between the gradient g_t and its first momentum m_t before using them to estimate the second momentum. The new optimization method is referred to as Aida. Thirdly, extensive experimental results show that Aida outperforms nine optimizers when training transformers and LSTMs for NLP, and VGG and ResNet for image classification over CIAF10 and CIFAR100 while matching the best performance of the nine methods when training WGAN-GP models for image generation tasks. Furthermore, Aida produces higher validation accuracies than AdaBelief for training ResNet18 over ImageNet. Code is available at this URL

We experimentally demonstrate the decomposition of entropy production during free-energy generation in a nanometer-scale dot transitioning to a non-equilibrium steady state via single-electron counting statistics. An alternating-current signal driving a reservoir that injects multiple electrons into the dot makes it non-equilibrium, leading to free-energy generation, heat dissipation, and Shannon-entropy production. By analyzing the time-domain probability distributions of multiple electron states of the dot, we quantitatively decompose the heat dissipation into housekeeping and excess heats, revealing the correlation between the free energy and the decomposed components of heat dissipation. This correlation suggests that the ratio of the generated free energy to the work applied to the dot, can potentially reach 0.5 under far-from-equilibrium conditions induced by a large signal, while an efficiency of 0.25 was experimentally achieved. Our results, providing the theoretical and experimental efficiencies from the relation between decomposed heat dissipation and free-energy generation, promise to connect non-equilibrium thermodynamics perspectives to electronic devices.

Data center networks (DCNs) require a low-cost, low-power optical transceiver to handle increased traffic from generative artificial intelligence, video streaming services, and more. Improving the required signal-to-noise ratio (RSNR) by digital signal processing such as forward error correction (FEC) mitigates the requirements for electrical and optical components. The optical transceivers in DCNs exploit a low-complexity soft-decision (SD) FEC, consisting of short block-length linear error-correcting codes and a low-complexity SD decoder (SDD), such as a Chase decoder and ordered statistical decoding. The low complexity SDD efficiently approaches a maximum likelihood decoding (MLD). However, the decoding performance of MLD is limited by its finite block length. In this paper, we describe the detail of our proposed channel-polarized multilevel coding with iterative decoding (CP-MLC-ID), which improves the decoding performance. The 19.5$\%$-OH CP-MLC-ID 128-bit extended Bose-Chaudhuri-Hocquenghem (eBCH) and KP4 codes outperform the concatenated eBCH and KP4 codes with a net coding gain of 0.25 and 0.40 dB for the same and double the number of SDDs, respectively. We also investigate the dependency of the decoding performance on the size of a bit interleaver. The performance degradation of CP-MLC-ID using an 8-bit interleaver is about 0.1 dB compared to using the large-bit interleaver. Our results indicate that even a weak connection by exclusive-OR between codewords improves the decoding performance, compared to simple concatenated codes in the DCNs.

The non-commutative harmonic oscillator (NCHO) was introduced as a specific Hamiltonian operator on $L^2(\mathbb{R})\otimes\mathbb{C}^2$ by Parmeggiani and Wakayama. Then it was proved by Ochiai and Wakayama that the eigenvalue problem for NCHO is reduced to a Heun differential equation. In this article, we consider some generalization of NCHO for $L^2(\mathbb{R}^n)\otimes\mathbb{C}^p$ as a rotation-invariant differential equation. Then by applying a representation theory of $\mathfrak{sl}(2,\mathbb{R})\simeq \mathfrak{su}(1,1)$, we check that its restriction to the space of products of radial functions and homogeneous harmonic polynomials is reduced to a holomorphic differential equation on the unit disk, which is generically Fuchsian.