Does prediction error drive one-shot declarative learning?

The role of prediction error (PE) in driving learning is well-established in fields such as classical and instrumental conditioning, reward learning and procedural memory; however, its role in human one-shot declarative encoding is less clear. According to one recent hypothesis, PE reflects the divergence between two probability distributions: one reflecting the prior probability (from previous experiences) and the other reflecting the sensory evidence (from the current experience). Assuming unimodal probability distributions, PE can be manipulated in three ways: (1) the distance between the mode of the prior and evidence, (2) the precision of the prior, and (3) the precision of the evidence. We tested these three manipulations across five experiments, in terms of peoples' ability to encode a single presentation of a scene-item pairing as a function of previous exposures to that scene and/or item. Memory was probed by presenting the scene together with three choices for the previously paired item, in which the two foil items were from other pairings within the same condition as the target item. In Experiment 1, we manipulated the evidence to be either consistent or inconsistent with prior expectations, predicting PE to be larger, and hence memory better, when the new pairing was inconsistent. In Experiments 2a-c, we manipulated the precision of the priors, predicting better memory for a new pairing when the (inconsistent) priors were more precise. In Experiment 3, we manipulated both visual noise and prior exposure for unfamiliar faces, before pairing them with scenes, predicting better memory when the sensory evidence was more precise. In all experiments, the PE hypotheses were supported. We discuss alternative explanations of individual experiments, and conclude the Predictive Interactive Multiple Memory Signals (PIMMS) framework provides the most parsimonious account of the full pattern of results.

Multimodal Word Meaning Induction From Minimal Exposure to Natural Text

By the time they reach early adulthood, English speakers are familiar with the meaning of thousands of words. In the last decades, computational simulations known as distributional semantic models (DSMs) have demonstrated that it is possible to induce word meaning representations solely from word co-occurrence statistics extracted from a large amount of text. However, while these models learn in batch mode from large corpora, human word learning proceeds incrementally after minimal exposure to new words. In this study, we run a set of experiments investigating whether minimal distributional evidence from very short passages suffices to trigger successful word learning in subjects, testing their linguistic and visual intuitions about the concepts associated with new words. After confirming that subjects are indeed very efficient distributional learners even from small amounts of evidence, we test a DSM on the same multimodal task, finding that it behaves in a remarkable human-like way. We conclude that DSMs provide a convincing computational account of word learning even at the early stages in which a word is first encountered, and the way they build meaning representations can offer new insights into human language acquisition.

How stable is perception in #TheDress and #TheShoe?

#TheDress is perceived by some people as black and blue while others perceive it as white and gold. We have previously shown that the first encounter with #TheDress strongly biases its perception. This percept remained stable during the experiment, suggesting a role of one-shot learning. #TheShoe is another image that elicits similar bimodal color percepts. Here, we investigated how percepts change over time in both #TheShoe and #TheDress. First, we show that the important role of one-shot learning, which we found for #TheDress extends to #TheShoe. Similarly to our previous results with the dress, hiding large parts of the image with occluders biased the percept of the shoe. The percept did not change for the majority of observers when the occluders were removed. Second, we investigated if and how percepts switch over a time course of 14 days. We found that although some observers experienced percept switches, the percept was largely stable for most observers.

CancerSiamese: one-shot learning for predicting primary and metastatic tumor types unseen during model training

BACKGROUND

The state-of-the-art deep learning based cancer type prediction can only predict cancer types whose samples are available during the training where the sample size is commonly large. In this paper, we consider how to utilize the existing training samples to predict cancer types unseen during the training. We hypothesize the existence of a set of type-agnostic expression representations that define the similarity/dissimilarity between samples of the same/different types and propose a novel one-shot learning model called CancerSiamese to learn this common representation. CancerSiamese accepts a pair of query and support samples (gene expression profiles) and learns the representation of similar or dissimilar cancer types through two parallel convolutional neural networks joined by a similarity function.

RESULTS

We trained CancerSiamese for cancer type prediction for primary and metastatic tumors using samples from the Cancer Genome Atlas (TCGA) and MET500. Network transfer learning was utilized to facilitate the training of the CancerSiamese models. CancerSiamese was tested for different N-way predictions and yielded an average accuracy improvement of 8% and 4% over the benchmark 1-Nearest Neighbor (1-NN) classifier for primary and metastatic tumors, respectively. Moreover, we applied the guided gradient saliency map and feature selection to CancerSiamese to examine 100 and 200 top marker-gene candidates for the prediction of primary and metastatic cancers, respectively. Functional analysis of these marker genes revealed several cancer related functions between primary and metastatic tumors.

CONCLUSION

This work demonstrated, for the first time, the feasibility of predicting unseen cancer types whose samples are limited. Thus, it could inspire new and ingenious applications of one-shot and few-shot learning solutions for improving cancer diagnosis, prognostic, and our understanding of cancer.

Putting a bug in ML: The moth olfactory network learns to read MNIST

We seek to (i) characterize the learning architectures exploited in biological neural networks for training on very few samples, and (ii) port these algorithmic structures to a machine learning context. The moth olfactory network is among the simplest biological neural systems that can learn, and its architecture includes key structural elements and mechanisms widespread in biological neural nets, such as cascaded networks, competitive inhibition, high intrinsic noise, sparsity, reward mechanisms, and Hebbian plasticity. These structural biological elements, in combination, enable rapid learning. MothNet is a computational model of the moth olfactory network, closely aligned with the moth's known biophysics and with in vivo electrode data collected from moths learning new odors. We assign this model the task of learning to read the MNIST digits. We show that MothNet successfully learns to read given very few training samples (1-10 samples per class). In this few-samples regime, it outperforms standard machine learning methods such as nearest-neighbors, support-vector machines, and neural networks (NNs), and matches specialized one-shot transfer-learning methods but without the need for pre-training. The MothNet architecture illustrates how algorithmic structures derived from biological brains can be used to build alternative NNs that may avoid the high training data demands of many current engineered NNs.

Concept learning through deep reinforcement learning with memory-augmented neural networks

Deep neural networks have shown superior performance in many regimes to remember familiar patterns with large amounts of data. However, the standard supervised deep learning paradigm is still limited when facing the need to learn new concepts efficiently from scarce data. In this paper, we present a memory-augmented neural network which is motivated by the process of human concept learning. The training procedure, imitating the concept formation course of human, learns how to distinguish samples from different classes and aggregate samples of the same kind. In order to better utilize the advantages originated from the human behavior, we propose a sequential process, during which the network should decide how to remember each sample at every step. In this sequential process, a stable and interactive memory serves as an important module. We validate our model in some typical one-shot learning tasks and also an exploratory outlier detection problem. In all the experiments, our model gets highly competitive to reach or outperform those strong baselines.

One-shot learning of view-invariant object representations in newborn chicks

Can newborn brains perform one-shot learning? To address this question, we reared newborn chicks in strictly controlled environments containing a single view of a single object, then tested their object recognition performance across 24 uniformly-spaced viewpoints. We found that chicks can build view-invariant object representations from a single view of an object: a case of one-shot learning in newborn brains. Chicks can also build the same view-invariant object representation from different views of an object, showing that newborn brains converge on common object representations from different sets of sensory inputs. Finally, by rearing chicks with larger numbers of object views, we found that chicks develop enhanced recognition for familiar views. These results illuminate the earliest stages of object recognition, revealing (1) powerful one-shot learning that builds invariant object representations from the first views of an object and (2) view-based learning that enriches object representations, producing enhanced recognition for familiar views.

A dataset for evaluating one-shot categorization of novel object classes

With the advent of deep convolutional neural networks, machines now rival humans in terms of object categorization. The neural networks solve categorization with a hierarchical organization that shares a striking resemblance to their biological counterpart, leading to their status as a standard model of object recognition in biological vision. Despite training on thousands of images of object categories, however, machine-learning networks are poorer generalizers, often fooled by adversarial images with very simple image manipulations that humans easily distinguish as a false image. Humans, on the other hand, can generalize object classes from very few samples. Here we provide a dataset of novel object classifications in humans. We gathered thousands of crowd-sourced human responses to novel objects embedded either with 1 or 16 context sample(s). Human decisions and stimuli together have the potential to be re-used (1) as a tool to better understand the nature of the gap in category learning from few samples between human and machine, and (2) as a benchmark of generalization across machine learning networks.

Segment Any Tissue: One-shot reference guided training-free automatic point prompting for medical image segmentation

Medical image segmentation frequently encounters high annotation costs and challenges in task adaptation. While visual foundation models have shown promise in natural image segmentation, automatically generating high-quality prompts for class-agnostic segmentation of medical images remains a significant practical challenge. To address these challenges, we present Segment Any Tissue (SAT), an innovative, training-free framework designed to automatically prompt the class-agnostic visual foundation model for the segmentation of medical images with only a one-shot reference. SAT leverages the robust feature-matching capabilities of a pretrained foundation model to construct distance metrics in the feature space. By integrating these with distance metrics in the physical space, SAT establishes a dual-space cyclic prompt engineering approach for automatic prompt generation, optimization, and evaluation. Subsequently, SAT utilizes a class-agnostic foundation segmentation model with the generated prompt scheme to obtain segmentation results. Additionally, we extend the one-shot framework by incorporating multiple reference images to construct an ensemble SAT, further enhancing segmentation performance. SAT has been validated on six public and private medical segmentation tasks, capturing both macroscopic and microscopic perspectives across multiple dimensions. In the ablation experiments, automatic prompt selection enabled SAT to effectively handle tissues of various sizes, while also validating the effectiveness of each component. The comparative experiments show that SAT is comparable to, or even exceeds, some fully supervised methods. It also demonstrates superior performance compared to existing one-shot methods. In summary, SAT requires only a single pixel-level annotated reference image to perform tissue segmentation across various medical images in a training-free manner. This not only significantly reduces the annotation costs of applying foundational models to the medical field but also enhances task transferability, providing a foundation for the clinical application of intelligent medicine. Our source code is available at https://github.com/SnowRain510/Segment-Any-Tissue.

Current strategies to address data scarcity in artificial intelligence-based drug discovery: A comprehensive review

Artificial intelligence (AI) has played a vital role in computer-aided drug design (CADD). This development has been further accelerated with the increasing use of machine learning (ML), mainly deep learning (DL), and computing hardware and software advancements. As a result, initial doubts about the application of AI in drug discovery have been dispelled, leading to significant benefits in medicinal chemistry. At the same time, it is crucial to recognize that AI is still in its infancy and faces a few limitations that need to be addressed to harness its full potential in drug discovery. Some notable limitations are insufficient, unlabeled, and non-uniform data, the resemblance of some AI-generated molecules with existing molecules, unavailability of inadequate benchmarks, intellectual property rights (IPRs) related hurdles in data sharing, poor understanding of biology, focus on proxy data and ligands, lack of holistic methods to represent input (molecular structures) to prevent pre-processing of input molecules (feature engineering), etc. The major component in AI infrastructure is input data, as most of the successes of AI-driven efforts to improve drug discovery depend on the quality and quantity of data, used to train and test AI algorithms, besides a few other factors. Additionally, data-gulping DL approaches, without sufficient data, may collapse to live up to their promise. Current literature suggests a few methods, to certain extent, effectively handle low data for better output from the AI models in the context of drug discovery. These are transferring learning (TL), active learning (AL), single or one-shot learning (OSL), multi-task learning (MTL), data augmentation (DA), data synthesis (DS), etc. One different method, which enables sharing of proprietary data on a common platform (without compromising data privacy) to train ML model, is federated learning (FL). In this review, we compare and discuss these methods, their recent applications, and limitations while modeling small molecule data to get the improved output of AI methods in drug discovery. Article also sums up some other novel methods to handle inadequate data.

Cone Beam Computed Tomography Image-Quality Improvement Using "One-Shot" Super-resolution

Cone beam computed tomography (CBCT) images are convenient representations for obtaining information about patients' internal organs, but their lower image quality than those of treatment planning CT images constitutes an important shortcoming. Several proposed CBCT image-quality improvement methods based on deep learning require large amounts of training data. Our newly developed model using a super-resolution method, "one-shot" super-resolution (OSSR) based on the "zero-shot" super-resolution method, requires only small amounts of training data to improve CBCT image quality using only the target CBCT image and the paired treatment planning CT image. For this study, pelvic CBCT images and treatment planning CT images of 30 prostate cancer patients were used. We calculated the root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) to evaluate image-quality improvement and normalized mutual information (NMI) as a quantitative evaluation of positional accuracy. Our proposed method can improve CBCT image quality without requiring large amounts of training data. After applying our proposed method, the resulting RMSE, PSNR, SSIM, and NMI between the CBCT images and the treatment planning CT images were as much as 0.86, 1.05, 1.03, and 1.31 times better than those obtained without using our proposed method. By comparison, CycleGAN exhibited values of 0.91, 1.03, 1.02, and 1.16. The proposed method achieved performance equivalent to that of CycleGAN, which requires images from approximately 30 patients for training. Findings demonstrated improvement of CBCT image quality using only the target CBCT images and the paired treatment planning CT images.

One-shot neuroanatomy segmentation through online data augmentation and confidence aware pseudo label

Recently, deep learning-based brain segmentation methods have achieved great success. However, most approaches focus on supervised segmentation, which requires many high-quality labeled images. In this paper, we pay attention to one-shot segmentation, aiming to learn from one labeled image and a few unlabeled images. We propose an end-to-end unified network that joints deformation modeling and segmentation tasks. Our network consists of a shared encoder, a deformation modeling head, and a segmentation head. In the training phase, the atlas and unlabeled images are input to the encoder to get multi-scale features. The features are then fed to the multi-scale deformation modeling module to estimate the atlas-to-image deformation field. The deformation modeling module implements the estimation at the feature level in a coarse-to-fine manner. Then, we employ the field to generate the augmented image pair through online data augmentation. We do not apply any appearance transformations cause the shared encoder could capture appearance variations. Finally, we adopt supervised segmentation loss for the augmented image. Considering that the unlabeled images still contain rich information, we introduce confidence aware pseudo label for them to further boost the segmentation performance. We validate our network on three benchmark datasets. Experimental results demonstrate that our network significantly outperforms other deep single-atlas-based and traditional multi-atlas-based segmentation methods. Notably, the second dataset is collected from multi-center, and our network still achieves promising segmentation performance on both the seen and unseen test sets, revealing its robustness. The source code will be available at https://github.com/zhangliutong/brainseg.

One-shot stimulus-control associations generalize over different stimulus viewpoints and exemplars

Cognitive control processes are central to adaptive behavior, but how control is applied in a context-appropriate manner is not fully understood. One way to produce context-sensitive control is by mnemonically linking particular control settings to specific stimuli that demanded those settings in a prior encounter. In support of this episodic reinstatement of control hypothesis, recent studies have produced evidence for the formation of stimulus-control associations in one-shot, prime-probe learning paradigms. However, since those studies employed perceptually identical stimuli across prime and probe presentations, it is not yet known how generalizable one-shot stimulus-control associations are. In the current study, we therefore probed whether associations formed between a prime object and the control process of task-switching would generalize to probe objects seen from a different viewpoint (Experiment 1), to different exemplars of the same object type (Experiment 2), and to different members of the object category (Experiment 3). We replicated prior findings of one-shot control associations for identical prime/probe stimuli. Importantly, we additionally found that these episodic control effects are expressed regardless of changes in viewpoint and exemplar, but do not seem to generalize to other category members. These findings elucidate the scope of generalization of the episodic reinstatement of cognitive control.