Combinatorial optimization (CO) problems are often NP-hard and thus out of reach for exact algorithms, making them a tempting domain to apply machine learning methods. The highly structured constraints in these problems can hinder either optimization or sampling directly in the solution space. On the other hand, GFlowNets have recently emerged as a powerful machinery to efficiently sample from composite unnormalized densities sequentially and have the potential to amortize such solution-searching processes in CO, as well as generate diverse solution candidates. In this paper, we design Markov decision processes (MDPs) for different combinatorial problems and propose to train conditional GFlowNets to sample from the solution space. Efficient training techniques are also developed to benefit long-range credit assignment. Through extensive experiments on a variety of different CO tasks with synthetic and realistic data, we demonstrate that GFlowNet policies can efficiently find high-quality solutions.

While Transformers have shown remarkable success in natural language processing, their attention mechanism's large memory requirements have limited their ability to handle longer contexts. Prior approaches, such as recurrent memory or retrieval-based augmentation, have either compromised the random-access flexibility of attention (i.e., the capability to select any token in the entire context) or relied on separate mechanisms for relevant context retrieval, which may not be compatible with the model's attention. In this paper, we present a novel approach that allows access to the complete context while retaining random-access flexibility, closely resembling running attention on the entire context. Our method uses a landmark token to represent each block of the input and trains the attention to use it for selecting relevant blocks, enabling retrieval of blocks directly through the attention mechanism instead of by relying on a separate mechanism. Our approach seamlessly integrates with specialized data structures and the system's memory hierarchy, enabling processing of arbitrarily long context lengths. We demonstrate that our method can obtain comparable performance with Transformer-XL while significantly reducing the number of retrieved tokens in each step. Finally, we show that fine-tuning LLaMA 7B with our method successfully extends its context length capacity to over 32k tokens, allowing for inference at the context lengths of GPT-4. We release the implementation of landmark attention and the code to reproduce our experiments at https://github.com/epfml/landmark-attention/.

Lastly, in scenarios where multiple agents are present, the behavioral mixture of agents approach, for example Vinyals et al. (2019) samples the final agent from the Nash distribution of the set of agents, can be utilized. Given that different agents, or experts, may recommend varying actions for an identical state, this results in an intrinsic stochastic policy, taking advantage of the diversity in agent decisions. If the state space is continuous, a common approach is to transform the actions into a normal or beta distribution. We apply one-hot encoding with temperature-scaled softmax. A discrete action space can be represented as a one-hot encoded vector, For instance, if action 2 out of 5 is chosen, its one-hot representation is [0, 1, 0, 0, 0], the scale the one-hot vector to [0, 1/τ, 0, 0, 0]. The higher the temperature coefficient τ, the more spread out the distribution becomes, while a lower temperature coefficient nudges the distribution closer to a deterministic action.

Optimizers are essential components for successfully training deep neural network models. In order to achieve the best performance from such models, designers need to carefully choose the optimizer hyperparameters. However, this can be a computationally expensive and time-consuming process. Although it is known that all optimizer hyperparameters must be tuned for maximum performance, there is still a lack of clarity regarding the individual influence of minor priority hyperparameters, including the safeguard factor $\epsilon$ and momentum factor $\beta$, in leading adaptive optimizers (specifically, those based on the Adam optimizers). In this manuscript, we introduce a new framework based on gradient histograms to analyze and justify important attributes of adaptive optimizers, such as their optimal performance and the relationships and dependencies among hyperparameters. Furthermore, we propose a novel gradient histogram-based algorithm that automatically estimates a reduced and accurate search space for the safeguard hyperparameter $\epsilon$, where the optimal value can be easily found.

Accurate spelling correction is a critical step in modern search interfaces, especially in an era of mobile devices and speech-to-text interfaces. For services that are deployed around the world, this poses a significant challenge for multilingual NLP: spelling errors need to be caught and corrected in all languages, and even in queries that use multiple languages. In this paper, we tackle this challenge using multi-teacher distillation. On our approach, a monolingual teacher model is trained for each language/locale, and these individual models are distilled into a single multilingual student model intended to serve all languages/locales. In experiments using open-source data as well as user data from a worldwide search service, we show that this leads to highly effective spelling correction models that can meet the tight latency requirements of deployed services.

LoRA achieves remarkable resource efficiency and comparable performance when adapting LLMs for specific tasks. Since ChatGPT demonstrated superior performance on various tasks, there has been a growing desire to adapt one model for all tasks. However, the explicit low-rank of LoRA limits the adaptation performance in complex multi-task scenarios. LoRA is dominated by a small number of top singular vectors while fine-tuning decomposes into a set of less important unitary transforms. In this paper, we propose MultiLoRA for better multi-task adaptation by reducing the dominance of top singular vectors observed in LoRA. MultiLoRA scales LoRA modules horizontally and change parameter initialization of adaptation matrices to reduce parameter dependency, thus yields more balanced unitary subspaces. We unprecedentedly construct specialized training data by mixing datasets of instruction follow, natural language understanding, world knowledge, to cover semantically and syntactically different samples. With only 2.5% of additional parameters, MultiLoRA outperforms single LoRA counterparts and fine-tuning on multiple benchmarks and model scales. Further investigation into weight update matrices of MultiLoRA exhibits reduced dependency on top singular vectors and more democratic unitary transform contributions.

Foundation models hold promise for transforming AI in healthcare by providing modular components that are easily adaptable to downstream healthcare tasks, making AI development more scalable and cost-effective. Structured EHR foundation models, trained on coded medical records from millions of patients, demonstrated benefits including increased performance with fewer training labels, and improved robustness to distribution shifts. However, questions remain on the feasibility of sharing these models across different hospitals and their performance for local task adaptation. This multi-center study examined the adaptability of a recently released structured EHR foundation model ($FM_{SM}$), trained on longitudinal medical record data from 2.57M Stanford Medicine patients. Experiments were conducted using EHR data at The Hospital for Sick Children and MIMIC-IV. We assessed both adaptability via continued pretraining on local data, and task adaptability compared to baselines of training models from scratch at each site, including a local foundation model. We evaluated the performance of these models on 8 clinical prediction tasks. In both datasets, adapting the off-the-shelf $FM_{SM}$ matched the performance of GBM models locally trained on all data while providing a 13% improvement in settings with few task-specific training labels. With continued pretraining on local data, label efficiency substantially improved, such that $FM_{SM}$ required fewer than 1% of training examples to match the fully trained GBM's performance. Continued pretraining was also 60 to 90% more sample-efficient than training local foundation models from scratch. Our findings show that adapting shared EHR foundation models across hospitals provides improved prediction performance at less cost, underscoring the utility of base foundation models as modular components to streamline the development of healthcare AI.

In this paper, we consider a heavy inner product identification problem, which generalizes the Light Bulb problem~(\cite{prr89}): Given two sets $A \subset \{-1,+1\}^d$ and $B \subset \{-1,+1\}^d$ with $|A|=|B| = n$, if there are exact $k$ pairs whose inner product passes a certain threshold, i.e., $\{(a_1, b_1), \cdots, (a_k, b_k)\} \subset A \times B$ such that $\forall i \in [k], \langle a_i,b_i \rangle \geq \rho \cdot d$, for a threshold $\rho \in (0,1)$, the goal is to identify those $k$ heavy inner products. We provide an algorithm that runs in $O(n^{2 \omega / 3+ o(1)})$ time to find the $k$ inner product pairs that surpass $\rho \cdot d$ threshold with high probability, where $\omega$ is the current matrix multiplication exponent. By solving this problem, our method speed up the training of neural networks with ReLU activation function.

Large pre-trained models have been instrumental in significant advancements in domains like language and vision making model training for individual downstream tasks more efficient as well as provide superior performance. However, tackling time-series analysis tasks usually involves designing and training a separate model from scratch leveraging training data and domain expertise specific to the task. We tackle a significant challenge for pre-training a general time-series model from multiple heterogeneous time-series dataset: providing semantically useful inputs to models for modeling time series of different dynamics from different domains. We observe that partitioning time-series into segments as inputs to sequential models produces semantically better inputs and propose a novel model LPTM that automatically identifies optimal dataset-specific segmentation strategy leveraging self-supervised learning loss during pre-training. LPTM provides performance similar to or better than domain-specific state-of-art model and is significantly more data and compute efficient taking up to 40% less data as well as 50% less training time to achieve state-of-art performance in a wide range of time-series analysis tasks from multiple disparate domains. Time-series analysis tasks involve important well-studied problems involving time-series datasets such as forecasting (Hyndman & Athanasopoulos, 2018) and classification (Chowdhury et al., 2022) with applications in wide-ranging domains such as retail, meteorology, economics, and health. Recent works (Chen et al., 2021; Wang et al., 2022; Zeng et al., 2023) have shown the efficacy of purely data-driven deep learning models in learning complex domain-specific properties of the time series over traditional statistic and mechanistic models across many domains. However, coming up with a model for a specific application or time-series analysis task is usually non-trivial.

The adaptive leaky integrate-and-fire (ALIF) model is fundamental within computational neuroscience and has been instrumental in studying our brains $\textit{in silico}$. Due to the sequential nature of simulating these neural models, a commonly faced issue is the speed-accuracy trade-off: either accurately simulate a neuron using a small discretisation time-step (DT), which is slow, or more quickly simulate a neuron using a larger DT and incur a loss in simulation accuracy. Here we provide a solution to this dilemma, by algorithmically reinterpreting the ALIF model, reducing the sequential simulation complexity and permitting a more efficient parallelisation on GPUs. We computationally validate our implementation to obtain over a $50\times$ training speedup using small DTs on synthetic benchmarks. We also obtained a comparable performance to the standard ALIF implementation on different supervised classification tasks - yet in a fraction of the training time. Lastly, we showcase how our model makes it possible to quickly and accurately fit real electrophysiological recordings of cortical neurons, where very fine sub-millisecond DTs are crucial for capturing exact spike timing.