Deep problems with neural network models of human vision

Behavioral and Brain Sciences, Journal Year: 2022, Volume and Issue: 46

Published: Dec. 1, 2022

Abstract Deep neural networks (DNNs) have had extraordinary successes in classifying photographic images of objects and are often described as the best models biological vision. This conclusion is largely based on three sets findings: (1) DNNs more accurate than any other model taken from various datasets, (2) do job predicting pattern human errors behavioral (3) brain signals response to datasets (e.g., single cell responses or fMRI data). However, these not test hypotheses regarding what features contributing good predictions we show that may be mediated by share little overlap with More problematically, account for almost no results psychological research. contradicts common claim good, let alone best, object recognition. We argue theorists interested developing biologically plausible vision need direct their attention explaining findings. generally, build explain experiments manipulate independent variables designed rather compete making predictions. conclude briefly summarizing promising modeling approaches focus data.

Language: Английский

---

Are Deep Neural Networks Adequate Behavioral Models of Human Visual Perception?

Annual Review of Vision Science, Journal Year: 2023, Volume and Issue: 9(1), P. 501 - 524

Published: March 31, 2023

Deep neural networks (DNNs) are machine learning algorithms that have revolutionized computer vision due to their remarkable successes in tasks like object classification and segmentation. The success of DNNs as has led the suggestion may also be good models human visual perception. In this article, we review evidence regarding current adequate behavioral core recognition. To end, argue it is important distinguish between statistical tools computational understand model quality a multidimensional concept which clarity about modeling goals key. Reviewing large number psychophysical explorations recognition performance humans DNNs, highly valuable scientific but that, today, should only regarded promising-but not yet adequate-computational behavior. On way, dispel several myths surrounding science.

Language: Английский

---

Learning high-level visual representations from a child’s perspective without strong inductive biases

Nature Machine Intelligence, Journal Year: 2024, Volume and Issue: 6(3), P. 271 - 283

Published: March 7, 2024

Language: Английский

---

Deep Problems with Neural Network Models of Human Vision

Published: April 13, 2022

Deep neural networks (DNNs) have had extraordinary successes in classifying photographic images of objects and are often described as the best models biological vision. This conclusion is largely based on three sets findings: (1) DNNs more accurate than any other model taken from various datasets, (2) do job predicting pattern human errors behavioral benchmark (3) brain signals response to datasets (e.g., single cell responses or fMRI data). However, most benchmarks report outcomes observational experiments that not manipulate independent variables, we show good prediction these may be mediated by share little overlap with More problematically, account for almost no results psychological research. contradicts common claim good, let alone best, object recognition. We argue theorists interested developing biologically plausible vision need direct their attention explaining findings. generally, build explain variables designed test hypotheses rather compete data. conclude briefly summarizing promising modelling approaches focus

Language: Английский

---

Faster region based convolution neural network with context iterative refinement for object detection

Measurement Sensors, Journal Year: 2024, Volume and Issue: 31, P. 101025 - 101025

Published: Jan. 8, 2024

In this paper, proposed a novel method to improve the localisation precision of identified objects. We present framework for iteratively enhancing image region recommendations meet ground truth values in research. The Faster R–CNN (FR-CNN) seems be an object recognition deep convolutional network. It gives user impression that network is cohesive and single. can provide accurate timely predictions about whereabouts range first build unified model based on rapid relocate inaccurate area recommendations. Because emphasis detection, it may utilized with wide datasets compatible various FR-CNN architectures. Second, we focus application joint score function variety picture features. This depicts location concealed concerning other data updated structured production loss are only two inputs influence parameters scoring function. join-score iterative context refinement (CIR) used generate our final model, which then classified using Smooth Support Vector Machine (SSVM). measured accuracy mean average after training + CIR SSVM low-cost GPU PASCAL VOC 2012 dataset. Our results 3.6 % more exact than rival learning algorithms average.

Language: Английский

---

Open Iris - An Open Source Framework for Video-Based Eye-Tracking Research and Development

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: March 3, 2024

ABSTRACT Eye-tracking is an essential tool in many fields, yet existing solutions are often limited for customized applications due to cost or lack of flexibility. We present OpenIris, adaptable and user-friendly open-source framework video-based eye-tracking. OpenIris developed C# with modular design that allows further extension customization through plugins different hardware systems, tracking, calibration pipelines. It can be remotely controlled via a network interface from other devices programs. Eye movements recorded online camera stream offline post-processing videos. Example have been track eye motion 3-D, including torsion. Currently implemented binocular pupil tracking pipelines achieve frame rates more than 500Hz. With the framework, we aim fill gap research tools available high-precision high-speed eye-tracking, especially environments require custom not currently well-served by commercial eye-trackers. CCS CONCEPTS Applied computing → Life medical sciences.

Language: Английский

---

Human-like face pareidolia emerges in deep neural networks optimized for face and object recognition

PLoS Computational Biology, Journal Year: 2025, Volume and Issue: 21(1), P. e1012751 - e1012751

Published: Jan. 27, 2025

The human visual system possesses a remarkable ability to detect and process faces across diverse contexts, including the phenomenon of face pareidolia—–seeing in inanimate objects. Despite extensive research, it remains unclear why employs such broadly tuned detection capabilities. We hypothesized that pareidolia results from system’s optimization for recognizing both To test this hypothesis, we used task-optimized deep convolutional neural networks (CNNs) evaluated their alignment with behavioral signatures responses, measured via magnetoencephalography (MEG), related processing. Specifically, trained CNNs on tasks involving combinations identification, detection, object categorization, detection. Using representational similarity analysis, found included categorization training represented faces, real matched objects more similarly responses than those did not. Although these showed similar overall data, closer examination internal representations revealed specific had distinct effects how were layers. Finally, interpretability methods only CNN identification relied face-like features—such as ‘eyes’—to classify stimuli mirroring findings perception. Our suggest human-like may emerge within context generalized categorization.

Language: Английский

---

Fine-grained knowledge about manipulable objects is well-predicted by CLIP

Published: May 11, 2024

Object recognition is an important human ability that relies on distinguishing between similar objects, for example, deciding which kitchen utensil(s) to use at different stages of meal preparation. Recent work describes the fine-grained organization knowledge about manipulable objects via study constituent dimensions are most relevant behavior, vision, manipulation, and function-based object properties. A logical extension this concerns whether or not these uniquely human, can be approximated by deep learning. Here, we show behavioral well-predicted a state-of-the-art multimodal network trained large diverse set image-text pairs - CLIP-ViT part, also generate good predictions behavior previously unseen objects. Moreover, model vastly outperforms comparison networks pre-trained with smaller, image-only training datasets. These results demonstrate impressive capacity approximate knowledge. We discuss possible sources benefit relative other tested models (e.g. pre-training vs. image only pre-training, dataset size, architecture).

Language: Английский

---

Neuronal and behavioral responses to naturalistic texture im­ages in macaque monkeys

Journal of Neuroscience, Journal Year: 2024, Volume and Issue: 44(42), P. e0349242024 - e0349242024

Published: Aug. 28, 2024

The visual world is richly adorned with texture, which can serve to delineate important elements of natural scenes. In anesthetized macaque monkeys, selectivity for the statistical features texture weak in V1, but substantial V2, suggesting that neuronal activity V2 might directly support perception. To test this, we investigated relation between single cell V1 and simultaneously measured behavioral judgments texture. We generated stimuli along a continuum naturalistic phase-randomized noise trained two monkeys judge whether sample more closely resembled one or other extreme. Analysis responses revealed individual neurons carried much less information about naturalness than reports. However, sensitivity neurons, especially those preferring textures, was significantly closer behavior compared V1. firing both predicted perceptual choices response repeated presentations same ambiguous stimulus monkey, despite low neural sensitivity. neither population choice second monkey. conclude supporting perception likely continue develop downstream V2. Further, combined data recorded while performed an orientation discrimination task, our results demonstrate choice-correlated early sensory cortex unstable across observers tasks, untethered from sensitivity, therefore unlikely reflect formation decisions.

Language: Английский

---

Neural correlates of minimal recognizable configurations in the human brain

Cell Reports, Journal Year: 2025, Volume and Issue: 44(3), P. 115429 - 115429

Published: March 1, 2025

Language: Английский

---