The aim of the notes is to demonstrate the potential for ideas in machine learning to impact on the fields of inverse problems and data assimilation. The perspective is one that is primarily aimed at researchers from inverse problems and/or data assimilation who wish to see a mathematical presentation of machine learning as it pertains to their fields. As a by-product of the presentation we present a succinct mathematical treatment of various topics in machine learning. The material on machine learning, along with some other related topics, is summarized in Part III, Appendix. Part I of the notes is concerned with inverse problems, employing material from Part III; Part II of the notes is concerned with data assimilation, employing material from Parts I and III.

We tackle the general differentiable meta learning problem that is ubiquitous in modern deep learning, including hyperparameter optimization, loss function learning, few-shot learning, invariance learning and more. These problems are often formalized as Bi-Level optimizations (BLO). We introduce a novel perspective by turning a given BLO problem into a stochastic optimization, where the inner loss function becomes a smooth probability distribution, and the outer loss becomes an expected loss over the inner distribution. To solve this stochastic optimization, we adopt Stochastic Gradient Langevin Dynamics (SGLD) MCMC to sample inner distribution, and propose a recurrent algorithm to compute the MC-estimated hypergradient. Our derivation is similar to forward-mode differentiation, but we introduce a new first-order approximation that makes it feasible for large models without needing to store huge Jacobian matrices. The main benefits are two-fold: i) Our stochastic formulation takes into account uncertainty, which makes the method robust to suboptimal inner optimization or non-unique multiple inner minima due to overparametrization; ii) Compared to existing methods that often exhibit unstable behavior and hyperparameter sensitivity in practice, our method leads to considerably more reliable solutions. We demonstrate that the new approach achieves promising results on diverse meta learning problems and easily scales to learning 87M hyperparameters in the case of Vision Transformers.

In binary ($0/1$) online classification with apple tasting feedback, the learner receives feedback only when predicting $1$. Besides some degenerate learning tasks, all previously known learning algorithms for this model are randomized. Consequently, prior to this work it was unknown whether deterministic apple tasting is generally feasible. In this work, we provide the first widely-applicable deterministic apple tasting learner, and show that in the realizable case, a hypothesis class is learnable if and only if it is deterministically learnable, confirming a conjecture of [Raman, Subedi, Raman, Tewari-24]. Quantitatively, we show that every class $\mathcal{H}$ is learnable with mistake bound $O \left(\sqrt{\mathtt{L}(\mathcal{H}) T \log T} \right)$ (where $\mathtt{L}(\mathcal{H})$ is the Littlestone dimension of $\mathcal{H}$), and that this is tight for some classes. We further study the agnostic case, in which the best hypothesis makes at most $k$ many mistakes, and prove a trichotomy stating that every class $\mathcal{H}$ must be either easy, hard, or unlearnable. Easy classes have (both randomized and deterministic) mistake bound $\Theta_{\mathcal{H}}(k)$. Hard classes have randomized mistake bound $\tilde{\Theta}_{\mathcal{H}} \left(k + \sqrt{T} \right)$, and deterministic mistake bound $\tilde{\Theta}_{\mathcal{H}} \left(\sqrt{k \cdot T} \right)$, where $T$ is the time horizon. Unlearnable classes have (both randomized and deterministic) mistake bound $\Theta(T)$. Our upper bound is based on a deterministic algorithm for learning from expert advice with apple tasting feedback, a problem interesting in its own right. For this problem, we show that the optimal deterministic mistake bound is $\Theta \left(\sqrt{T (k + \log n)} \right)$ for all $k$ and $T \leq n \leq 2^T$, where $n$ is the number of experts.

The utilization of sketching techniques has progressively emerged as a pivotal method for enhancing the efficiency of online learning. In linear bandit settings, current sketch-based approaches leverage matrix sketching to reduce the per-round time complexity from \(\Omega\left(d^2\right)\) to \(O(d)\), where \(d\) is the input dimension. Despite this improved efficiency, these approaches encounter critical pitfalls: if the spectral tail of the covariance matrix does not decrease rapidly, it can lead to linear regret. In this paper, we revisit the regret analysis and algorithm design concerning approximating the covariance matrix using matrix sketching in linear bandits. We illustrate how inappropriate sketch sizes can result in unbounded spectral loss, thereby causing linear regret. To prevent this issue, we propose Dyadic Block Sketching, an innovative streaming matrix sketching approach that adaptively manages sketch size to constrain global spectral loss. This approach effectively tracks the best rank-\( k \) approximation in an online manner, ensuring efficiency when the geometry of the covariance matrix is favorable. Then, we apply the proposed Dyadic Block Sketching to linear bandits and demonstrate that the resulting bandit algorithm can achieve sublinear regret without prior knowledge of the covariance matrix, even under the worst case. Our method is a general framework for efficient sketch-based linear bandits, applicable to all existing sketch-based approaches, and offers improved regret bounds accordingly. Additionally, we conduct comprehensive empirical studies using both synthetic and real-world data to validate the accuracy of our theoretical findings and to highlight the effectiveness of our algorithm.

Recurrent stochastic configuration networks (RSCNs) are a class of randomized learner models that have shown promise in modelling nonlinear dynamics. In many fields, however, the data generated by industry systems often exhibits nonstationary characteristics, leading to the built model performing well on the training data but struggling with the newly arriving data. This paper aims at developing a self-organizing version of RSCNs, termed as SORSCNs, to enhance the continuous learning ability of the network for modelling nonstationary data. SORSCNs can autonomously adjust the network parameters and reservoir structure according to the data streams acquired in real-time. The output weights are updated online using the projection algorithm, while the network structure is dynamically adjusted in the light of the recurrent stochastic configuration algorithm and an improved sensitivity analysis. Comprehensive comparisons among the echo state network (ESN), online self-learning stochastic configuration network (OSL-SCN), self-organizing modular ESN (SOMESN), RSCN, and SORSCN are carried out. Experimental results clearly demonstrate that the proposed SORSCNs outperform other models with sound generalization, indicating great potential in modelling nonlinear systems with nonstationary dynamics.

We show how the quality of decisions based on the aggregated opinions of the crowd can be conveniently studied using a sample of individual responses to a standard IQ questionnaire. We aggregated the responses to the IQ questionnaire using simple majority voting and a machine learning approach based on a probabilistic graphical model. The score for the aggregated questionnaire, Crowd IQ, serves as a quality measure of decisions based on aggregating opinions, which also allows quantifying individual and crowd performance on the same scale. We show that Crowd IQ grows quickly with the size of the crowd but saturates, and that for small homogeneous crowds the Crowd IQ significantly exceeds the IQ of even their most intelligent member. We investigate alternative ways of aggregating the responses and the impact of the aggregation method on the resulting Crowd IQ. We also discuss Contextual IQ, a method of quantifying the individual participant's contribution to the Crowd IQ based on the Shapley value from cooperative game theory.

Mosques are worship places of Allah and must be preserved clean, immaculate, provide all the comforts of the worshippers in them. The prophet's mosque in Medina/ Saudi Arabia is one of the most important mosques for Muslims. It occupies second place after the sacred mosque in Mecca/ Saudi Arabia, which is in constant overcrowding by all Muslims to visit the prophet Mohammad's tomb. This paper aims to propose a smart dome model to preserve the fresh air and allow the sunlight to enter the mosque using artificial intelligence techniques. The proposed model controls domes movements based on the weather conditions and the overcrowding rates in the mosque. The data have been collected from two different resources, the first one from the database of Saudi Arabia weather's history, and the other from Shanghai Technology Database. Congested Scene Recognition Network (CSRNet) and Fuzzy techniques have applied using Python programming language to control the domes to be opened and closed for a specific time to renew the air inside the mosque. Also, this model consists of several parts that are connected for controlling the mechanism of opening/closing domes according to weather data and the situation of crowding in the mosque. Finally, the main goal of this paper has been achieved, and the proposed model has worked efficiently and specifies the exact duration time to keep the domes open automatically for a few minutes for each hour head.

Human action recognition has been a topic of interest across multiple fields ranging from security to entertainment systems. Tracking the motion and identifying the action being performed on a real time basis is necessary for critical security systems. In entertainment, especially gaming, the need for immediate responses for actions and gestures are paramount for the success of that system. We show that Motion History image has been a well established framework to capture the temporal and activity information in multi dimensional detail enabling various usecases including classification. We utilize MHI to produce sample data to train a classifier and demonstrate its effectiveness for action classification across six different activities in a single multi-action video. We analyze the classifier performance and identify usecases where MHI struggles to generate the appropriate activity image and discuss mechanisms and future work to overcome those limitations.

Language model calibration refers to the alignment between the confidence of the model and the actual performance of its responses. While previous studies point out the overconfidence phenomenon in Large Language Models (LLMs) and show that LLMs trained with Reinforcement Learning from Human Feedback (RLHF) are overconfident with a more sharpened output probability, in this study, we reveal that RLHF tends to lead models to express verbalized overconfidence in their own responses. We investigate the underlying cause of this overconfidence and demonstrate that reward models used for Proximal Policy Optimization (PPO) exhibit inherent biases towards high-confidence scores regardless of the actual quality of responses. Building upon this insight, we propose two PPO variants: PPO-M: PPO with Calibrated Reward Modeling and PPO-C: PPO with Calibrated Reward Calculation. PPO-M integrates explicit confidence scores in reward model training, which calibrates reward models to better capture the alignment between response quality and verbalized confidence. PPO-C adjusts the reward score during PPO based on the difference between the current reward and the moving average of past rewards. Both PPO-M and PPO-C can be seamlessly integrated into the current PPO pipeline and do not require additional golden labels. We evaluate our methods on both Llama3-8B and Mistral-7B across six diverse datasets including multiple-choice and open-ended generation. Experiment results demonstrate that both of our methods can reduce calibration error and maintain performance comparable to standard PPO. We further show that they do not compromise model capabilities in open-ended conversation settings.

This paper introduces a novel generalized self-imitation learning ($\textbf{GSIL}$) framework, which effectively and efficiently aligns large language models with offline demonstration data. We develop $\textbf{GSIL}$ by deriving a surrogate objective of imitation learning with density ratio estimates, facilitating the use of self-generated data and optimizing the imitation learning objective with simple classification losses. $\textbf{GSIL}$ eliminates the need for complex adversarial training in standard imitation learning, achieving lightweight and efficient fine-tuning for large language models. In addition, $\textbf{GSIL}$ encompasses a family of offline losses parameterized by a general class of convex functions for density ratio estimation and enables a unified view for alignment with demonstration data. Extensive experiments show that $\textbf{GSIL}$ consistently and significantly outperforms baselines in many challenging benchmarks, such as coding (HuamnEval), mathematical reasoning (GSM8K) and instruction-following benchmark (MT-Bench).