Large Language Models (LLMs) have demonstrated remarkable capabilities in open-ended text generation tasks. However, the inherent open-ended nature of these tasks implies that there is always room for improvement in the quality of model responses. To address this challenge, various approaches have been proposed to enhance the performance of LLMs. There has been a growing focus on enabling LLMs to self-improve their response quality, thereby reducing the reliance on extensive human annotation efforts for collecting diverse and high-quality training data. Recently, prompting-based methods have been widely explored among self-improvement methods owing to their effectiveness, efficiency, and convenience. However, those methods usually require explicitly and thoroughly written rubrics as inputs to LLMs. It is expensive and challenging to manually derive and provide all necessary rubrics with a real-world complex goal for improvement (e.g., being more helpfulness and less harmful). To this end, we propose an imPlicit self-ImprovemenT (PIT) framework that implicitly learns the improvement goal from human preference data. PIT only requires preference data that are used to train reward models with no extra human efforts. Specifically, we reformulate the training objective of reinforcement learning from human feedback (RLHF) -- instead of maximizing response quality for a given input, we maximize the quality gap of the response conditioned on a reference response. In this way, PIT is implicitly trained with the improvement goal of better aligning with human preferences. Experiments on two real-world datasets and one synthetic dataset show that our method significantly outperforms prompting-based methods.

Sparse Mixture-of-Experts (MoE) is a neural architecture design that adds learnable parameters to Large Language Models (LLMs) without increasing computational complexity (FLOPs). Instruction tuning is a technique for training LLMs to follow instructions. We advocate combining these two approaches, as we find that MoE models benefit more from instruction tuning than dense models. In particular, we conduct empirical studies across three experimental setups: (i) Direct finetuning on individual downstream tasks devoid of instruction tuning; (ii) Instruction tuning followed by in-context few-shot or zero-shot generalization on downstream tasks; and (iii) Instruction tuning supplemented by further finetuning on individual downstream tasks. In the first scenario, MoE models overall underperform dense models of identical computational capacity. This narrative, however, dramatically changes with the introduction of instruction tuning (in the second and third scenarios), used independently or in conjunction with task-specific finetuning. Our most powerful model, FLAN-MoE-32B, surpasses the performance of Flan-PaLM-62B on four benchmark tasks, while using only a third of the FLOPs. The advancements embodied by FLAN-MoE inspire a reevaluation of the design principles of large-scale, high-performance language models in the framework of task-agnostic learning.

Finetuning language models on a collection of datasets phrased as instructions has been shown to improve model performance and generalization to unseen tasks. In this paper we explore instruction finetuning with a particular focus on (1) scaling the number of tasks, (2) scaling the model size, and (3) finetuning on chain-of-thought data. We find that instruction finetuning with the above aspects dramatically improves performance on a variety of model classes (PaLM, T5, U-PaLM), prompting setups (zero-shot, few-shot, CoT), and evaluation benchmarks (MMLU, BBH, TyDiQA, MGSM, open-ended generation, RealToxicityPrompts). For instance, Flan-PaLM 540B instruction-finetuned on 1.8K tasks outperforms PaLM 540B by a large margin (+9.4% on average). Flan-PaLM 540B achieves state-of-the-art performance on several benchmarks (at time of release), such as 75.2% on five-shot MMLU. We also publicly release Flan-T5 checkpoints,1 which achieve strong few-shot performance even compared to much larger models, such as PaLM 62B. Overall, instruction finetuning is a general method for improving the performance and usability of pretrained language models.

Large Language Models (LLMs) have achieved excellent performances in various tasks.However, fine-tuning an LLM requires extensive supervision.Human, on the other hand, may improve their reasoning abilities by self-thinking without external inputs.In this work, we demonstrate that an LLM is also capable of self-improving with only unlabeled datasets.We use a pre-trained LLM to generate "highconfidence" rationale-augmented answers for unlabeled questions using Chain-of-Though (CoT) prompting and self-consistency, and finetune the LLM using those self-generated solutions as target outputs.We show that without any ground truth label, our approach significantly improves the general reasoning ability of PaLM 540B model (74.4%→82.1% on GSM8K, 90.0%→94.4% on OpenBookQA, and 63.4%→67.9% on ANLI-A3) and can also be adapted to extreme low-resource cases where even training questions and CoT prompts are limited.We conduct ablation studies and show that fine-tuning on diverse reasoning paths is critical for self-improvement.

Jerry Wei, Le Hou, Andrew Lampinen, Xiangning Chen, Da Huang, Yi Tay, Xinyun Chen, Yifeng Lu, Denny Zhou, Tengyu Ma, Quoc Le. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

We study the design decision of publicly available instruction tuning methods, by reproducing and breaking down the development of Flan 2022 (Chung et al., 2022). Through careful ablation studies on the Flan Collection of tasks and methods, we tease apart the effect of design decisions which enable Flan-T5 to outperform prior work by 3-17% across evaluation settings. We find task balancing and enrichment techniques are overlooked but critical to effective instruction tuning, and in particular, training with mixed prompt settings (zero-shot, few-shot, chain-of-thought) actually yields equivalent or stronger (2%) performance in all settings. In further experiments we show Flan-T5 requires less finetuning to converge higher and faster than T5 on single downstream tasks -- motivating instruction-tuned models as more computationally-efficient starting checkpoints for new tasks. Finally, to accelerate research on instruction tuning, we make the Flan 2022 collection of datasets, templates, and methods publicly available.

Le Hou, Richard Yuanzhe Pang, Tianyi Zhou, Yuexin Wu, Xinying Song, Xiaodan Song, Denny Zhou. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Recent shadow detection algorithms have shown initial success on small datasets of images from specific domains. However, shadow detection on broader image domains is still challenging due to the lack of annotated training data, caused by the intense manual labor required for annotating shadow data. In this paper we propose "lazy annotation", an efficient annotation method where an annotator only needs to mark the important shadow areas and some non-shadow areas. This yields data with noisy labels that are not yet useful for training a shadow detector. We address the problem of label noise by jointly learning a shadow region classifier and recovering the labels in the training set. We consider the training labels as unknowns and formulate label recovery as the minimization of the sum of squared leave-one-out errors of a Least Squares SVM, which can be efficiently optimized. Experimental results show that a classifier trained with recovered labels achieves comparable performance to a classifier trained on the properly annotated data. These results motivated us to collect a new dataset that is 20 times larger than existing datasets and contains a large variety of scenes and image types. Naturally, such a large dataset is appropriate for training deep learning methods. Thus, we propose a stacked Convolutional Neural Network architecture that efficiently trains on patch level shadow examples while incorporating image level semantic information. This means that the detected shadow patches are refined based on image semantics. Our proposed pipeline, trained on recovered labels, performs at state-of-the art level. Furthermore, the proposed model performs exceptionally well on a cross dataset task, proving the generalization power of the proposed architecture and dataset.

Deep learning, through the use of neural networks, has demonstrated remarkable ability to automate many routine tasks when presented with sufficient data for training. The neural network architecture (e.g. number of layers, types of layers, connections between layers, etc.) plays a critical role in determining what, if anything, the neural network is able to learn from the training data. The trend for neural network architectures, especially those trained on ImageNet, has been to grow ever deeper and more complex. The result has been ever increasing accuracy on benchmark datasets with the cost of increased computational demands. In this paper we demonstrate that neural network architectures can be automatically generated, tailored for a specific application, with dual objectives: accuracy of prediction and speed of prediction. Using MENNDL- an HPC-enabled software stack for neural architecture search-we generate a neural network with comparable accuracy to state-of-the-art networks on a cancer pathology dataset that is also 16× faster at inference. The speedup in inference is necessary because of the volume and velocity of cancer pathology data; specifically, the previous state-of-the-art networks are too slow for individual researchers without access to HPC systems to keep pace with the rate of data generation. Our new model enables researchers with modest computational resources to analyze newly generated data faster than it is collected.

Detection, segmentation and classification of nuclei are fundamental analysis operations in digital pathology. Existing state-of-the-art approaches demand extensive amount of supervised training data from pathologists and may still perform poorly in images from unseen tissue types. We propose an unsupervised approach for histopathology image segmentation that synthesizes heterogeneous sets of training image patches, of every tissue type. Although our synthetic patches are not always of high quality, we harness the motley crew of generated samples through a generally applicable importance sampling method. This proposed approach, for the first time, re-weighs the training loss over synthetic data so that the ideal (unbiased) generalization loss over the true data distribution is minimized. This enables us to use a random polygon generator to synthesize approximate cellular structures (i.e., nuclear masks) for which no real examples are given in many tissue types, and hence, GAN-based methods are not suited. In addition, we propose a hybrid synthesis pipeline that utilizes textures in real histopathology patches and GAN models, to tackle heterogeneity in tissue textures. Compared with existing state-of-the-art supervised models, our approach generalizes significantly better on cancer types without training data. Even in cancer types with training data, our approach achieves the same performance without supervision cost. We release code and segmentation results on over 5000 Whole Slide Images (WSI) in The Cancer Genome Atlas (TCGA) repository, a dataset that would be orders of magnitude larger than what is available today.

In this paper we propose multi-resolution data fusion methods for deep learning-based high-resolution land cover mapping from aerial imagery. The land cover mapping problem, at country-level scales, is challenging for common deep learning methods due to the scarcity of high-resolution labels, as well as variation in geography and quality of input images. On the other hand, multiple satellite imagery and low-resolution ground truth label sources are widely available, and can be used to improve model training efforts. Our methods include: introducing low-resolution satellite data to smooth quality differences in high-resolution input, exploiting low-resolution labels with a dual loss function, and pairing scarce high-resolution labels with inputs from several points in time. We train models that are able to generalize from a portion of the Northeast United States, where we have high-resolution land cover labels, to the rest of the US. With these models, we produce the first high-resolution (1-meter) land cover map of the contiguous US, consisting of over 8 trillion pixels. We demonstrate the robustness and potential applications of this data in a case study with domain experts and develop a web application to share our results. This work is practically useful, and can be applied to other locations over the earth as high-resolution imagery becomes more widely available even as high-resolution labeled land cover data remains sparse.

Within Neural Networks (NN), the parameters of Adaptive Activation Functions (AAF) control the shapes of activation functions. These parameters are trained along with other parameters in the NN. AAFs have improved performance of Convolutional Neural Networks (CNN) in multiple classification tasks. In this paper, we propose and apply AAFs on CNNs for regression tasks. We argue that applying AAFs in the regression (second-to-last) layer of a NN can significantly decrease the bias of the regression NN. However, using existing AAFs may lead to overfitting. To address this problem, we propose a Smooth Adaptive Activation Function (SAAF) with a piecewise polynomial form which can approximate any continuous function to arbitrary degree of error, while having a bounded Lipschitz constant for given bounded model parameters. As a result, NNs with SAAF can avoid overfitting by simply regularizing model parameters. We empirically evaluated CNNs with SAAFs and achieved state-of-the-art results on age and pose estimation datasets.

Classifying the various shapes and attributes of a glioma cell nucleus is crucial for diagnosis and understanding of the disease. We investigate the automated classification of the nuclear shapes and visual attributes of glioma cells, using Convolutional Neural Networks (CNNs) on pathology images of automatically segmented nuclei. We propose three methods that improve the performance of a previously-developed semi-supervised CNN. First, we propose a method that allows the CNN to focus on the most important part of an image-the image's center containing the nucleus. Second, we inject (concatenate) pre-extracted VGG features into an intermediate layer of our Semi-Supervised CNN so that during training, the CNN can learn a set of additional features. Third, we separate the losses of the two groups of target classes (nuclear shapes and attributes) into a single-label loss and a multi-label loss in order to incorporate prior knowledge of inter-label exclusiveness. On a dataset of 2078 images, the combination of the proposed methods reduces the error rate of attribute and shape classification by 21.54% and 15.07% respectively compared to the existing state-of-the-art method on the same dataset.

Convolutional Neural Networks (CNN) are state-of-the-art models for many image classification tasks. However, to recognize cancer subtypes automatically, training a CNN on gigapixel resolution Whole Slide Tissue Images (WSI) is currently computationally impossible. The differentiation of cancer subtypes is based on cellular-level visual features observed on image patch scale. Therefore, we argue that in this situation, training a patch-level classifier on image patches will perform better than or similar to an image-level classifier. The challenge becomes how to intelligently combine patch-level classification results and model the fact that not all patches will be discriminative. We propose to train a decision fusion model to aggregate patch-level predictions given by patch-level CNNs, which to the best of our knowledge has not been shown before. Furthermore, we formulate a novel Expectation-Maximization (EM) based method that automatically locates discriminative patches robustly by utilizing the spatial relationships of patches. We apply our method to the classification of glioma and non-small-cell lung carcinoma cases into subtypes. The classification accuracy of our method is similar to the inter-observer agreement between pathologists. Although it is impossible to train CNNs on WSIs, we experimentally demonstrate using a comparable non-cancer dataset of smaller images that a patch-based CNN can outperform an image-based CNN.