Minyoung Kim

Minyoung Kim, Timothy Hospedales

We release a new Bayesian neural network library for PyTorch for large-scale deep networks. Our library implements mainstream approximate Bayesian inference algorithms: variational inference, MC-dropout, stochastic-gradient MCMC, and Laplace approximation. The main differences from other existing Bayesian neural network libraries are as follows: 1) Our library can deal with very large-scale deep networks including Vision Transformers (ViTs). 2) We need virtually zero code modifications for users (e.g., the backbone network definition codes do not neet to be modified at all). 3) Our library also allows the pre-trained model weights to serve as a prior mean, which is very useful for performing Bayesian inference with the large-scale foundation models like ViTs that are hard to optimise from scratch with the downstream data alone. Our code is publicly available at: \url{https://github.com/SamsungLabs/BayesDLL}\footnote{A mirror repository is also available at: \url{https://github.com/minyoungkim21/BayesDLL}.}.

Minyoung Kim, Da Li, Timothy Hospedales

We tackle the domain generalisation (DG) problem by posing it as a domain adaptation (DA) task where we adversarially synthesise the worst-case target domain and adapt a model to that worst-case domain, thereby improving the model's robustness. To synthesise data that is challenging yet semantics-preserving, we generate Fourier amplitude images and combine them with source domain phase images, exploiting the widely believed conjecture from signal processing that amplitude spectra mainly determines image style, while phase data mainly captures image semantics. To synthesise a worst-case domain for adaptation, we train the classifier and the amplitude generator adversarially. Specifically, we exploit the maximum classifier discrepancy (MCD) principle from DA that relates the target domain performance to the discrepancy of classifiers in the model hypothesis space. By Bayesian hypothesis modeling, we express the model hypothesis space effectively as a posterior distribution over classifiers given the source domains, making adversarial MCD minimisation feasible. On the DomainBed benchmark including the large-scale DomainNet dataset, the proposed approach yields significantly improved domain generalisation performance over the state-of-the-art.

Issam H. Laradji, Mark Schmidt, Vladimir Pavlovic et al.

Deep Gaussian processes (DGP) have appealing Bayesian properties, can handle variable-sized data, and learn deep features. Their limitation is that they do not scale well with the size of the data. Existing approaches address this using a deep random feature (DRF) expansion model, which makes inference tractable by approximating DGPs. However, DRF is not suitable for variable-sized input data such as trees, graphs, and sequences. We introduce the GP-DRF, a novel Bayesian model with an input layer of GPs, followed by DRF layers. The key advantage is that the combination of GP and DRF leads to a tractable model that can both handle a variable-sized input as well as learn deep long-range dependency structures of the data. We provide a novel efficient method to simultaneously infer the posterior of GP's latent vectors and infer the posterior of DRF's internal weights and random frequencies. Our experiments show that GP-DRF outperforms the standard GP model and DRF model across many datasets. Furthermore, they demonstrate that GP-DRF enables improved uncertainty quantification compared to GP and DRF alone, with respect to a Bhattacharyya distance assessment. Source code is available at https://github.com/IssamLaradji/GP_DRF.

Xinnuo Xu, Minyoung Kim, Royson Lee et al.

Data point selection (DPS) is becoming a critical topic in deep learning due to the ease of acquiring uncurated training data compared to the difficulty of obtaining curated or processed data. Existing approaches to DPS are predominantly based on a bi-level optimisation (BLO) formulation, which is demanding in terms of memory and computation, and exhibits some theoretical defects regarding minibatches. Thus, we propose a novel Bayesian approach to DPS. We view the DPS problem as posterior inference in a novel Bayesian model where the posterior distributions of the instance-wise weights and the main neural network parameters are inferred under a reasonable prior and likelihood model. We employ stochastic gradient Langevin MCMC sampling to learn the main network and instance-wise weights jointly, ensuring convergence even with minibatches. Our update equation is comparable to the widely used SGD and much more efficient than existing BLO-based methods. Through controlled experiments in both the vision and language domains, we present the proof-of-concept. Additionally, we demonstrate that our method scales effectively to large language models and facilitates automated per-task optimization for instruction fine-tuning datasets.

Royson Lee, Minyoung Kim, Fady Rezk et al.

Federated learning (FL) has enabled the training of multilingual large language models (LLMs) on diverse and decentralized multilingual data, especially on low-resource languages. To improve client-specific performance, personalization via the use of parameter-efficient fine-tuning (PEFT) modules such as LoRA is common. This involves a personalization strategy (PS), such as the design of the PEFT adapter structures (e.g., in which layers to add LoRAs and what ranks) and choice of hyperparameters (e.g., learning rates) for fine-tuning. Instead of manual PS configuration, we propose FedP$^2$EFT, a federated learning-to-personalize method for multilingual LLMs in cross-device FL settings. Unlike most existing PEFT structure selection methods, which are prone to overfitting low-data regimes, FedP$^2$EFT collaboratively learns the optimal personalized PEFT structure for each client via Bayesian sparse rank selection. Evaluations on both simulated and real-world multilingual FL benchmarks demonstrate that FedP$^2$EFT largely outperforms existing personalized fine-tuning methods, while complementing other existing FL methods.

Minyoung Kim, Timothy M. Hospedales

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.

Minyoung Kim, Timothy Hospedales

We propose a novel hierarchical Bayesian approach to Federated Learning (FL), where our model reasonably describes the generative process of clients' local data via hierarchical Bayesian modeling: constituting random variables of local models for clients that are governed by a higher-level global variate. Interestingly, the variational inference in our Bayesian model leads to an optimisation problem whose block-coordinate descent solution becomes a distributed algorithm that is separable over clients and allows them not to reveal their own private data at all, thus fully compatible with FL. We also highlight that our block-coordinate algorithm has particular forms that subsume the well-known FL algorithms including Fed-Avg and Fed-Prox as special cases. Beyond introducing novel modeling and derivations, we also offer convergence analysis showing that our block-coordinate FL algorithm converges to an (local) optimum of the objective at the rate of $O(1/\sqrt{t})$, the same rate as regular (centralised) SGD, as well as the generalisation error analysis where we prove that the test error of our model on unseen data is guaranteed to vanish as we increase the training data size, thus asymptotically optimal.

Minyoung Kim, Timothy Hospedales

The meta learning few-shot classification is an emerging problem in machine learning that received enormous attention recently, where the goal is to learn a model that can quickly adapt to a new task with only a few labeled data. We consider the Bayesian Gaussian process (GP) approach, in which we meta-learn the GP prior, and the adaptation to a new task is carried out by the GP predictive model from the posterior inference. We adopt the Laplace posterior approximation, but to circumvent the iterative gradient steps for finding the MAP solution, we introduce a novel linear discriminant analysis (LDA) plugin as a surrogate for the MAP solution. In essence, the MAP solution is approximated by the LDA estimate, but to take the GP prior into account, we adopt the prior-norm adjustment to estimate LDA's shared variance parameters, which ensures that the adjusted estimate is consistent with the GP prior. This enables closed-form differentiable GP posteriors and predictive distributions, thus allowing fast meta training. We demonstrate considerable improvement over the previous approaches.

Minyoung Kim, Vladimir Pavlovic

Variational autoencoder (VAE) is a very successful generative model whose key element is the so called amortized inference network, which can perform test time inference using a single feed forward pass. Unfortunately, this comes at the cost of degraded accuracy in posterior approximation, often underperforming the instance-wise variational optimization. Although the latest semi-amortized approaches mitigate the issue by performing a few variational optimization updates starting from the VAE's amortized inference output, they inherently suffer from computational overhead for inference at test time. In this paper, we address the problem in a completely different way by considering a random inference model, where we model the mean and variance functions of the variational posterior as random Gaussian processes (GP). The motivation is that the deviation of the VAE's amortized posterior distribution from the true posterior can be regarded as random noise, which allows us to take into account the uncertainty in posterior approximation in a principled manner. In particular, our model can quantify the difficulty in posterior approximation by a Gaussian variational density. Inference in our GP model is done by a single feed forward pass through the network, significantly faster than semi-amortized methods. We show that our approach attains higher test data likelihood than the state-of-the-arts on several benchmark datasets.

Behnam Gholami, Pritish Sahu, Minyoung Kim et al.

Domain Adaptation (DA), the process of effectively adapting task models learned on one domain, the source, to other related but distinct domains, the targets, with no or minimal retraining, is typically accomplished using the process of source-to-target manifold alignment. However, this process often leads to unsatisfactory adaptation performance, in part because it ignores the task-specific structure of the data. In this paper, we improve the performance of DA by introducing a discriminative discrepancy measure which takes advantage of auxiliary information available in the source and the target domains to better align the source and target distributions. Specifically, we leverage the cohesive clustering structure within individual data manifolds, associated with different tasks, to improve the alignment. This structure is explicit in the source, where the task labels are available, but is implicit in the target, making the problem challenging. We address the challenge by devising a deep DA framework, which combines a new task-driven domain alignment discriminator with domain regularizers that encourage the shared features as task-specific and domain invariant, and prompt the task model to be data structure preserving, guiding its decision boundaries through the low density data regions. We validate our framework on standard benchmarks, including Digits (MNIST, USPS, SVHN, MNIST-M), PACS, and VisDA. Our results show that our proposed model consistently outperforms the state-of-the-art in unsupervised domain adaptation.

Mark-Oliver Stehr, Minyoung Kim, Carolyn L. Talcott et al.

In spite of the rapidly increasing number of applications of machine learning in various domains, a principled and systematic approach to the incorporation of domain knowledge in the engineering process is still lacking and ad hoc solutions that are difficult to validate are still the norm in practice, which is of growing concern not only in mission-critical applications. In this note, we introduce Probabilistic Approximate Logic (PALO) as a logic based on the notion of mean approximate probability to overcome conceptual and computational difficulties inherent to strictly probabilistic logics. The logic is approximate in several dimensions. Logical independence assumptions are used to obtain approximate probabilities, but by averaging over many instances of formulas a useful estimate of mean probability with known confidence can usually be obtained. To enable efficient computational inference, the logic has a continuous semantics that reflects only a subset of the structural properties of classical logic, but this imprecision can be partly compensated by richer theories obtained by classical inference or other means. Computational inference, which refers to the construction of models and validation of logical properties, is based on Stochastic Gradient Descent (SGD) and Markov Chain Monte Carlo (MCMC) techniques and hence another dimension where approximations are involved. We also present the Logical Imagination Engine (LIME), a prototypical implementation of PALO based on TensorFlow. Albeit not limited to the biological domain, we illustrate its operation in a quite substantial bioinformatics machine learning application concerned with network synthesis and analysis in a recent DARPA project.

Minyoung Kim, Pritish Sahu, Behnam Gholami et al.

In unsupervised domain adaptation, it is widely known that the target domain error can be provably reduced by having a shared input representation that makes the source and target domains indistinguishable from each other. Very recently it has been studied that not just matching the marginal input distributions, but the alignment of output (class) distributions is also critical. The latter can be achieved by minimizing the maximum discrepancy of predictors (classifiers). In this paper, we adopt this principle, but propose a more systematic and effective way to achieve hypothesis consistency via Gaussian processes (GP). The GP allows us to define/induce a hypothesis space of the classifiers from the posterior distribution of the latent random functions, turning the learning into a simple large-margin posterior separation problem, far easier to solve than previous approaches based on adversarial minimax optimization. We formulate a learning objective that effectively pushes the posterior to minimize the maximum discrepancy. This is further shown to be equivalent to maximizing margins and minimizing uncertainty of the class predictions in the target domain, a well-established principle in classical (semi-)supervised learning. Empirical results demonstrate that our approach is comparable or superior to the existing methods on several benchmark domain adaptation datasets.