Seminar

This seminar reviewed AI Agents for Inventory Control: Human-LLM-OR Complementarity, which examines how operations research methods, large language models, and human judgment can be combined for sequential inventory decisions under uncertain demand and lead times. The presentation compared alternative decision pipelines in which OR and LLM components play different roles, highlighting how LLMs can incorporate contextual information and adapt when the assumptions underlying traditional inventory models become unreliable. Experiments on synthetic and real-world inventory settings showed that hybrid OR–LLM approaches can outperform either component alone, while a human-in-the-loop study found the strongest performance when humans retained final decision authority while receiving LLM reasoning. The seminar highlighted the broader importance of assigning OR, LLMs, and humans to complementary roles rather than treating them as interchangeable decision makers.

This seminar reviewed Meta-learning with Differentiable Closed-form Solvers, a few-shot learning approach that enables rapid adaptation to new tasks. The presentation focused on integrating differentiable solvers such as ridge regression and logistic regression into deep neural networks, allowing the feature representation to be meta-learned for efficient adaptation. It also discussed the use of the Woodbury identity to reduce the computational cost of closed-form solutions. Experimental results across three few-shot learning benchmarks demonstrated competitive or superior performance compared with existing methods.

This seminar presented federated learning strategies for building personalized models in environments where data cannot be centrally collected. Based on three studies conducted as part of a doctoral dissertation and research at SIA Lab, the presentation introduced global federated learning approaches designed to improve the efficiency and effectiveness of personalized model development while preserving decentralized data. The seminar focused on the problem settings, key methodological ideas, and main experimental findings of each study, providing a high-level overview of how federated learning can support personalization under heterogeneous and distributed data environments.

This seminar reviewed SoftPatch: Unsupervised Anomaly Detection with Noisy Data, a method designed to address the degradation of anomaly detection performance when training data are contaminated with anomalous samples. The presentation focused on how SoftPatch identifies and mitigates patch-level noise by assigning outlier scores to image patches before coreset construction and incorporating these scores into the memory bank. It further discussed how this approach softens the anomaly decision boundary and improves robustness to noisy training data while preserving strong performance under clean settings. Experimental results on the MVTecAD and BTAD benchmarks were also reviewed, highlighting SoftPatch’s effectiveness compared with existing unsupervised anomaly detection methods across various noise conditions.

This seminar presented an analysis of gradient-based meta-learning, specifically comparing feature reuse and representation change in algorithms such as MAML, ANIL, and BOIL. The presentation explored whether rapid learning in MAML relies on feature reuse or representation change, utilizing layer freezing and CCA/CKA analyses. It further introduced ANIL (Almost No Inner Loop) for computational efficiency via feature reuse and BOIL (Body Only update In inner Loop), which enforces representation change in the body while keeping the head fixed to improve performance particularly under domain shifts.

This seminar presented FR-LLM, a unified large language model that jointly performs fault diagnosis and remaining useful life prediction for rotating machinery. The framework converts vibration signals into structured textual inputs using frequency-domain, statistical, and empirical-mode-decomposition features. It further employs an adaptive Convergence Balancer to mitigate interference between classification and regression tasks and uses low-rank adaptation to reduce computational costs, demonstrating improved accuracy, generalization, and training efficiency on bearing datasets.

This seminar introduced PatchCore, an industrial anomaly-detection method that learns exclusively from normal training images. PatchCore constructs a representative memory bank of local patch-level features extracted from pretrained image models and detects defects by comparing test features with the stored normal patterns. The method achieves strong performance in both anomaly detection and localization while maintaining practical inference efficiency, including an image-level AUROC of up to 99.6% on the MVTec AD benchmark.

This seminar introduced the Transformer, a sequence-modeling architecture constructed entirely from attention mechanisms without recurrent or convolutional layers. Multi-head self-attention allows the model to capture relationships between tokens while supporting substantially more parallel computation than recurrent networks. Experiments on machine-translation tasks demonstrate that the Transformer improves predictive quality while reducing training time, establishing the foundation of many subsequent large-scale language and multimodal models.

This seminar presented a domain-adversarial approach that learns from labeled source-domain data and unlabeled target-domain data. A gradient reversal layer encourages the network to learn features that are discriminative for the primary task but invariant to differences between the source and target domains. The entire model can be optimized using standard backpropagation and demonstrates strong adaptation performance even when substantial distribution shifts exist between domains.

This seminar introduced Prototypical Networks, a metric-based approach for classifying previously unseen classes from only a few labeled examples. The method represents each class using a prototype computed from its support examples and assigns a query sample according to its distance from these prototypes in a learned embedding space. Despite its simple inductive structure, the approach achieves strong few-shot classification performance and can also be extended to zero-shot learning.

This seminar presented a general meta-learning framework for learning transferable forecasting strategies from diverse time-series datasets. The study interprets residual connections as an implicit adaptation mechanism that generates task-specific behavior from each input series, connecting neural forecasting architectures with gradient-based meta-learning. Experiments demonstrate that a model trained on source datasets can forecast previously unseen target time series without additional training while remaining competitive with established univariate forecasting methods.

This seminar introduced Meta Test-Time Training, which combines meta-learning, self-supervision, and test-time adaptation to address distribution shifts encountered after model deployment. The model is trained so that minimizing a self-supervised objective on an unlabeled test sample produces parameters that improve the primary prediction task. MT3 can therefore adapt to each test instance without labels and achieves improved robustness on corrupted image-classification benchmarks.

This seminar explored the mechanism through which Transformers perform in-context learning without explicitly updating their parameters. The study establishes a connection between the computation of a linear self-attention layer and a gradient-descent update on a regression objective. Experimental analysis suggests that trained Transformers can operate as implicit optimizers during their forward passes and can improve upon standard gradient descent by learning mechanisms such as iterative curvature correction.

This seminar presented global federated learning strategies for constructing models that can be efficiently personalized without directly sharing private client data. FedFN addresses unstable feature representations caused by heterogeneous client data, while FedDr+ combines Dot Regression and Feature Distillation to improve local adaptation without losing global knowledge. The seminar also discussed personalized reward-model learning under preference heterogeneity, showing that a strong single global model can provide an effective initialization for subsequent client-level fine-tuning.

This seminar presented Moirai, a universal time-series forecasting Transformer trained on the Large-scale Open Time Series Archive, which contains more than 27 billion observations from nine application domains. The model addresses key challenges in large-scale time-series pretraining, including different sampling frequencies, arbitrary numbers of variables, and heterogeneous data distributions. Moirai demonstrates competitive or superior zero-shot forecasting performance compared with models trained separately on individual downstream datasets.

This seminar introduced BatLiNet, a deep learning framework for predicting battery lifetime under diverse cycling protocols, temperatures, and battery chemistries. In addition to learning the degradation pattern of an individual cell, BatLiNet compares target and reference cells to model their lifetime differences through inter-cell learning. Combining intra-cell and inter-cell information improves prediction accuracy and robustness while enabling knowledge transfer across different ageing conditions and battery chemistries.

This seminar investigated whether multimodal Transformers remain reliable when one or more input modalities are unavailable. The study demonstrates that Transformers are sensitive to missing modalities and that their robustness depends strongly on the selected modality-fusion strategy. Because the optimal strategy varies across datasets, the authors propose automatically identifying an appropriate fusion method for the given data, resulting in improved performance on multiple multimodal benchmarks.

This seminar provided an overview of active learning, in which a model selectively requests labels for the most informative unlabeled samples. Major query strategies based on uncertainty, diversity, representativeness, and expected model improvement were introduced, along with pool-based and stream-based learning settings. Selected studies were reviewed to discuss how active learning can reduce labeling costs and the practical challenges associated with sampling bias, noisy annotations, and model calibration.

This seminar examined long-tailed recognition by separating representation learning from classifier training. The study shows that naturally imbalanced data can still produce strong feature representations and that performance on minority classes can be improved by adjusting only the classifier after representation learning. This simple decoupled approach achieves strong results on major long-tailed benchmarks without relying on complicated loss functions or model architectures.

This seminar introduced Model-Agnostic Meta-Learning, or MAML, which learns model parameters that can be rapidly adapted to a new task using only a small number of examples and gradient updates. Unlike approaches designed for a specific architecture, MAML can be applied to any model trained through gradient descent. Its effectiveness was demonstrated across few-shot classification, regression, and reinforcement learning tasks.

This seminar presented a deep latent variable framework for estimating carried hand loads from gait data collected using wearable sensors. The proposed model incorporates unloaded baseline gait and carrying-style information through temporal convolutional networks and bidirectional cross-attention, while removing the need for carrying-style labels during inference. Experiments on real-world motion data demonstrate that explicitly integrating auxiliary information substantially improves load estimation compared with direct prediction and simple feature concatenation.

This seminar introduced Gaussian Processes as probabilistic models that provide both predictions and uncertainty estimates for unknown functions. It also covered Bayesian Optimization, which uses a surrogate model and an acquisition function to efficiently identify promising solutions when evaluating the objective function is expensive. Key concepts, including kernels, posterior inference, and the exploration–exploitation trade-off, were discussed through illustrative examples.