While the provided corpus includes explicit references to the brain-to-text generative decoding work of Tang et al. (2023) [1, 2], it does not contain information on Takagi et al.'s "MindEye" or Chen's 2024 brain-to-video research. Therefore, this response will focus on the principles of generative decoding, semantic reconstruction, and naturalistic neuroimaging as detailed in the corpus.
Current State & What is Decodable: The decoding of naturalistic stimuli has evolved from predicting simple visual features to reconstructing continuous, high-level semantics. Voxel-wise encoding models, combined with the predictive power of Large Language Models (LLMs), have enabled the semantic reconstruction of continuous language from non-invasive fMRI recordings [1, 2]. Deep language algorithms (like GPT-2) and visual Convolutional Neural Networks (CNNs) exhibit activation patterns that linearly map onto human brain responses [3, 4]. We can now decode a wide array of representations from the brain: from specific semantic categories (e.g., tools, animals, places) [5], to complex semantic relationships (e.g., "whole-part" or "object-attribute") [6, 7], to continuous narrative intervals [8, 9], and even the affective dimensions of spontaneous daydreams (self-relevance and valence) [10, 11].
Fidelity Benchmarks: Rather than exact pixel-for-pixel or word-for-word accuracy, current benchmarks primarily rely on Representational Similarity Analysis (RSA) and noise ceilings to quantify fidelity. RSA measures how well the distances between representations in an AI model (like Word2Vec or GPT-2) match the distances between neural representations in the brain [12, 13]. For instance, when mapping GPT-2 to fMRI data of subjects listening to spoken stories, the algorithm's activations predict cortical responses with specific "brain scores" (Pearson correlations); in the superior temporal sulcus, these scores can capture up to 60% of the maximum explainable signal (the noise ceiling) [14, 15]. In visual decoding, CNNs are benchmarked against the inferior temporal (IT) cortex's ability to categorize highly variable objects, matching both human behavioral performance and neural unit responses [16-18].
Limits: Despite these successes, generative decoding is constrained by spatiotemporal limits and structural differences between AI and the brain. For example, MR-based eye tracking (DeepMReye) can successfully decode general gaze directions from fMRI signals but lacks fine spatial precision and suffers from a "center bias," failing to accurately predict exact coordinate locations [19-21]. Furthermore, voxel-wise fMRI has low temporal resolution (~1.5 seconds), which severely limits the ability to decode fast, sublexical representations like individual phonemes [22].
Decoding Modality-Independent Semantic Meaning: The brain does not store meaning purely as sensory echoes; it abstracts it. The corpus demonstrates that we can decode modality-independent semantic meaning. Semantic processing engages a widespread network of "heteromodal" or "transmodal" association areas—such as the Default Mode Network (DMN), angular gyrus, and temporal poles—that integrate highly convergent, supramodal input [23-25]. Furthermore, word embeddings mapped to fMRI show that abstract semantic relationships (like conceptual progression from a "part" to a "whole") are encoded by the co-occurring activation of the DMN and deactivation of the frontoparietal network [26]. This cortical pattern encodes the relationship entirely independently of the individual words or sensory features involved [26].
Personal/Autobiographical Interpretations: Neural decoding can capture personal interpretations that are not physically present in the stimulus. Subjective beliefs and schemas powerfully shape cortical representations, sometimes overriding the external input. In the "Same Story, Different Story" experiment, participants were given different contextual primes (e.g., a cheating wife vs. a paranoid husband) before listening to the exact same audio narrative [27, 28]. The neural responses in the DMN, language areas, and mirror neuron system clustered strictly based on the subjects' subjective interpretations, and the neural distance between groups perfectly correlated with the magnitude of their interpretive differences [29-31]. Similarly, when watching a movie with a twist ending (The Sixth Sense), discovering the twist updates the previously encoded event memories; during recall, the DMN reflects the newly integrated personal understanding rather than the naive encoding [32-34]. Models trained on personal narratives have also successfully decoded the "self-relevance" and "valence" of unconstrained, spontaneous thoughts (daydreams) during resting states [10, 11].
Unconscious Content: Decoding can bypass conscious awareness. Multivariate pattern analysis (MVPA) of EEG data during a masked perceptual decision-making task reveals two parallel processing streams. A late, global neural pattern corresponds to the cognitive maintenance of a stimulus and is highly correlated with the subject's decision confidence and conscious reportability [35-37]. However, the brain simultaneously maintains an "off-diagonal" pattern restricted to posterior electrode sites that reflects the sensory maintenance of category-specific information (e.g., faces vs. houses), which operates entirely independently of decision confidence [35, 36, 38]. This indicates that specific semantic content is decodable even when it is not reportable by or accessible to the participant.
Privacy Implications: Because generative decoding techniques can read spontaneous mind-wandering states [39, 40], track highly idiosyncratic emotional daydreams [41, 42], map political/social interpretations [43, 44], and extract unconscious sensory representations [35], the privacy implications are profound. These tools bypass voluntary communication, allowing direct access to the internal, unedited microworlds of individuals [44, 45].
What Decoding Teaches Us About Cortical Format: Decoding successes prove that the brain does not use isolated, anatomically segregated regions to store specific concepts (a modular format) [46]. Instead, meaning is represented by spatially overlapping, distributed cortical patterns [5, 47]. Furthermore, these networks are highly dynamic. The DMN, traditionally considered a "task-negative" network for resting, actively reconfigures its inter-regional correlations to track narrative coherence and event boundaries over time [48-50].
Semantic Disentanglement & Predictive Coding: By utilizing Generative AI (like GPT-2), neuroscientists can disentangle how the cortex processes language. Predictive coding theory suggests the brain continuously forecasts future inputs [51, 52]. When researchers decomposed GPT-2's predictions into separate syntactic and semantic forecast windows, they found a striking hierarchical split in the brain. Temporal cortices predominantly predict short-term, shallow syntactic representations, whereas frontoparietal areas (the top of the language hierarchy) actively predict long-term, highly contextualized semantic representations [53-55]. Thus, AI clarifies that "meaning" in higher cortex is fundamentally anticipatory.
Symbol Grounding and Abstraction: Generative models clarify how symbols (words) are grounded in the brain. When Word2Vec embeddings are mapped to the cortex, they reveal that the transition from concrete concepts (encoded outside the DMN) to abstract concepts (encoded within the DMN) corresponds to specific network shifts [56]. The DMN plays a central, active role in the abstraction of concepts, redefining its function as a core hub for relational and abstract semantic processing [56, 57].
Multimodal Integration: Comparing deep neural networks (DNNs) to human brains exposes crucial truths about multimodal integration. Human fMRI and EEG fusion during naturalistic audiovisual perception reveals an early, asymmetrical cross-modal leakage: acoustic information is represented in early visual regions, while visual information is restricted to visual cortices [58-60]. However, when researchers tested a state-of-the-art audiovisual DNN featuring strictly separated audio and video branches (only integrating at the highest layers), the AI completely failed to capture this early cross-modal interaction [61, 62]. This failure teaches us that biological "meaning" is inherently and deeply multisensory from the earliest stages of perception; to build biologically plausible AI that truly mirrors cortical representation, models require early cross-modal structural connections [63].
[1] (src:67a91294) A unique challenge for data analysis posed by brain recordings during naturalistic tasks (e.g. movie-watching) is the presence of linearly and nonlinearly correlated confounding variables. These limit the effectiveness of standard statistical tools such as t-tests. However, encoding models may prese...
[2] (src:67a91294) 64. Huth AG, de Heer WA, Griffiths TL, Theunissen FE, Gallant JL. 2016 Natural speech reveals the semantic maps that tile human cerebral cortex. Nature 532, 453–458. (doi:10.1038/nature17637) 65. Hamilton LS, Huth AG. 2020 The revolution will not be controlled: natural stimuli in speech neuroscience...
[3] (src:4927f9ca) computational neuroscience | computer vision | array electrophysiology Retinal images of real-world objects vary drastically due to changes in object pose, size, position, lighting, nonrigid de- formation, occlusion, and many other sources of noise and var-iation. Humans effortlessly recognize objec...
[4] (src:a24af881) Results Deep language models map onto brain activity First, we quantified the similarity between deep language models and the brain, when these two systems are inputted with the same stories. For this, we used the Narratives dataset39 and analysed the fMRI of 304 individuals listening to short stori...
[5] (src:b55f6f4f) Distributed cortical patterns encoded semantic categories. Since the encoding model was generalizable to new words and sentences, we further used it to predict cortical responses to >9000 words from nine categories: tool, human, plant, animal, place, communication, emotion, change, quantity (Supplem...
[6] (src:b55f6f4f) of the representational geometry further highlights the distinction across semantic relations in terms of their bilateral (a)symmetry and engagement of individual ROIs (Supplementary Fig. 8). However, several nominal (human-defined) relations, e.g., similar, contrast, object-nonattribute, and cause-...
[7] (src:b55f6f4f) 0 1.5 Z Toola d g h i e f b cAnimal Plant Human PlaceCommunication Quantity Change Emotion Fig. 4 Cortical representations of semantic categories. For each category, the color indicates the mean of the normalized response (or the z score) averaged across word samples in the category (Supplementary T...
[8] (src:7b279216) or using the full ‘fingerprint’ of the correlation pattern, which is a 10� 10 matrix of pairwise correlations between brain regions (right bars). (d) Confusion matrices of FC (left) and ISFC (right) classification across the four conditions. Features for classification were the entire fingerprints (...
[9] (src:7b279216) subjects were used for each condition (resting state, word scramble, paragraph scramble, intact story). and replication subject groups (Fig. 4c, black and grey lines). This more temporally fine-grained analysis also reproduced the pattern of FC results, in which average FC patterns showed little var...
[10] (src:16cdf350) The contents and dynamics of spontaneous thought are important factors for personality traits and mental health. However, assessing spontaneous thoughts is challenging due to their unconstrained nature, and directing participants’ attention to report their thoughts may fundamentally alter them. Here...
[11] (src:16cdf350) Discussion In this study, we developed multivariate pattern- based predictive models of self- relevance and valence that can be used to decode affective dimensions of spontaneous thoughts. For this, we con-ducted an fMRI experiment using a narrative- reading task, in which we showed personal stories...
[12] (src:72df40b6) A Toolbox for Representational Similarity Analysis Hamed Nili1*, Cai Wingfield2, Alexander Walther1, Li Su1,3, William Marslen-Wilson3, Nikolaus Kriegeskorte1* 1 MRC Cognition and Brain Sciences Unit, Cambridge, United Kingdom, 2 Department of Computer Science, University of Bath, Bath, United Kingd...
[13] (src:72df40b6) comprehensively, including not only what information is present, but also the format, in which the information is represented. In addition, we would like to use activity measurements to test computational models of brain information processing [12]. One approach to these challenges is representation...
[14] (src:a24af881) Received: 31 March 2022 Accepted: 15 December 2022 Published online: 2 March 2023 Check for updates 1Meta AI, Paris, France. 2Université Paris-Saclay, Inria, Commissariat à l’Énergie Atomique et aux Énergies Alternatives, Paris, France. 3Laboratoire des systèmes perceptifs, Département d’études cogn...
[15] (src:a24af881) score computation. X consists of the average fMRI recordings of the other individuals who listened to the same stories as individual s. X and Y have the same dimensionality and the bold delay is assumed to be comparable across individuals, so we did not apply a FIR to X. Thus, Rnoise ceiling(s) = co...
[16] (src:4927f9ca) intermediate-level (e.g., V4) neurons followed by simple addi-tional nonlinearities (14, 16, 29). Models were selected for evaluation by one of three proce- dures: (i) random sampling of the uniform distribution over parameter space (Fig. 1A; n = 2,016, green dots); (ii) opti-mization for performanc...
[17] (src:4927f9ca) classifiers on the IT neural population (Fig. 2B, green bars) and the V4 neural population (n= 128, hatched green bars). To ex-pose a key axis of recognition difficulty, we computed perfor-mance results at three levels of object view variation, from low (fixed orientation, size, and position) to hig...
[18] (src:4927f9ca) 100 80 60 40 20 LN LN ... LN LN ... LN LN LN ... LN LN LN ... . . . . . . Spatial Convolution over Image Input A B Fig. 2. Neural-like models via performance optimization. (A) We (1) used high-throughput computational methods to optimize the parameters of a hierarchical CNN with linear-nonlinear (LN...
[19] (src:3fa91849) Comparison of camera-based and MR-based eye tracking Magnetic resonance (MR)-based eye tracking has been explored for decades as an alternative to traditional camera-based systems17,22–24. This approach eliminates the need for additional eye tracking hardware, as it relies solely on the fMRI data to...
[20] (src:3fa91849) Although the re-trained MR-based models performed well on the calibration data, they exhibited limitations when applied to the rest of the experiments. Specifically, while the DeepMReye models success-fully recovered general gaze directions (e.g., looking up, down, left, or right), demonstrating the...
[21] (src:3fa91849) In contrast, the camera-based eye tracking method, enhanced with MoCET, consistently outperformed MR-based methods, main-taining both precise gaze direction and spatial accuracy of gaze coordinates across the entire experiment (ps <0:001 for all compar-isons). We observed that MR-based models tended...
[22] (src:a24af881) we highlight that this issue also prevails in language models, where word sequences, but arguably not their meaning, rapidly become unpredictable. Our results suggests that predicting multiple levels of representations over multiple temporal scopes may be critical to address the indeterminate nature...
[23] (src:1fcbb92e) high-level integrative processes. All are known to receive extensively processed, multimodal and supramodal input. Recent studies show that even cortical regions formerly considered ‘‘unimodal’’ receive multisensory inputs (Schroeder and Foxe 2004; Cappe and Barone 2005), blurring the traditional di...
[24] (src:1fcbb92e) and ‘‘amodal’’ cortex, where input from multiple modalities is more nearly balanced and highly convergent. For continuity with previous work, we refer to these latter regions as heteromodal, though alternative terms such as supramodal or amodal are perhaps equally valid. The human semantic system th...
[25] (src:a0cff8d4) Edited by Peter L. Strick, University of Pittsburgh, Pittsburgh, PA, and approved September 9, 2016 (received for review May 27, 2016) Understanding how the structure of cognition arises from the topographical organization of the cortex is a primary goal in neuroscience. Previous work has described ...
[26] (src:b55f6f4f) For a given word pair, their relation vector could be further projected onto the cortex through the encoding model. For an initial exploration, we applied this analysis to 178 word pairs that all shared a whole-part relationship. For example, in four word pairs, (hand, finger), (zoo, animal), (hour,...
[27] (src:85345bd6) Understanding interactions between characters in a story activates many brain regions, including regions implicated in thinking about the mental states of other people (Adolphs, 2009; Fletcher et al., 1995; Mar, 2011). The mentalizing network—which overlaps with the default-mode network (DMN; Mars e...
[28] (src:85345bd6) Stimuli and experimental design Subjects listened to an adapted version of the J. D. Salin-ger short story “Pretty Mouth and Green My Eyes.” The adapted version was shorter than the original and included some sentences that were not present in the original text. It was read by a professional actor a...
[29] (src:85345bd6) Abstract Differences in people’s beliefs can substantially impact their interpretation of a series of events. In this functional MRI study, we manipulated subjects’ beliefs, leading two groups of subjects to interpret the same narrative in different ways. We found that responses in higher-order brai...
[30] (src:85345bd6) Keywords narrative, interpretation, context, theory of mind, neuroimaging Received 1/6/16; Revision accepted 11/10/16 308 Yeshurun et al. Although the DMN is known to be involved in the processing of other people’s mental states, it remains unclear whether patterns of activity within the mentaliz-in...
[31] (src:85345bd6) Correlation of the neural responses in differentiating voxels with changes in the interpretation of characters’ beliefs, emotions, and intentions Which aspects of the interpretation (if any) drive the context-dependent changes in response time courses within these differentiating voxels? To answer t...
[32] (src:4fbb427f) Neural representations of naturalistic events are updated as our understanding of the past changes Asieh Zadbood1*, Samuel Nastase2, Janice Chen3, Kenneth A Norman2, Uri Hasson2 1Department of Psychology, Columbia University, New York, United States; 2Princeton Neuroscience Institute and Department ...
[33] (src:4fbb427f) Research article Neuroscience Our study design hinges on the hypothesis that participants who received the twist and are aware that the doctor is a ghost might have distinct neural representations of the events from those who encoded the movie while ignorant of the twist. Importantly, we predicted t...
[34] (src:4fbb427f) We tested this hypothesis in two ways. First, we predicted that (Figure 3A, prediction legend) the neural pattern similarity between recall in the ‘twist’ group and encoding in the ‘spoiled’ group (RG to MG) would be higher than the pattern similarity between recall in the ‘no- twist’ group and enco...
[35] (src:352b6b9d) Abstract Several influential theories of consciousness attempt to explain how, when and where conscious perception arises in the brain. The extent of conscious perception of a stimulus is often probed by asking subjects to provide confidence estimations in their choices in challenging perceptual dec...
[36] (src:352b6b9d) perception (Dehaene et al., 2006; Lamme, 2006; Del Cul et al., 2007). Crucially, off-diagonal decoding performance (trained at 150–200 ms; Fig. 3C) revealed a strikingly different temporal profile of results. Decoding was similarly higher early in time than later in time (F(1,22) � 17.93, MSE � 0.01...
[37] (src:352b6b9d) Several influential theories of consciousness try to ex-plain how, when and where conscious perception emerges from brain activity, and how this differs from processing unconscious information (Rees et al., 2002; Tononi and Koch, 2008; Haynes, 2009; Dehaene and Changeux, 2011; Kunde et al., 2012; va...
[38] (src:352b6b9d) (off-diagonal) pattern, showing sensory maintenance of category-specific information that was unrelated to decision confidence, may indicate more local recurrent processes within visual (sensory) cortices. Although in-triguing, future studies are necessary to confirm this in-terpretation of the pres...
[39] (src:84095659) Distinct individual differences in default mode network connectivity relate to off-task thought and text memory during reading Meichao Zhang1*, Nicola Savill2, Daniel S. Margulies 3, Jonathan Smallwood1 & Elizabeth Jefferies1 Often, as we read, we find ourselves thinking about something other than ...
[40] (src:84095659) In Experiment 1, we employed a sentence-reading task in the scanner, in which participants were asked to passively view meaningful sentences or nonwords, presented on an item-by-item basis. This task was expected to activate the visual-to-semantic pathway, implicated in reading comprehension11,50 an...
[41] (src:16cdf350) personal story | spontaneous thought | functional magnetic resonance imaging | brain decoding | affective neuroscience Our mind never rests. Even during quiet periods or sleep, our mind spontaneously wanders from the past to the future and from one concept to another (1–3). Spontaneous thoughts may...
[42] (src:16cdf350) Given that the contents and dynamics of spontaneous thought could provide rich information about individuals’ mental and brain health, the ability to decode some aspects of spontaneous thought directly from neuroimaging data would be useful. In this study, we focused on two content dimensions of spo...
[43] (src:5471b5d9) processes underlying political cognition. In adapting fMRI para-digms to use real-world naturalistic political content, we study the biased processing of political content in a setting where we can be more confident of ecological validity. Using this approach, we identified a neural signature of bia...
[44] (src:85345bd6) 318 Yeshurun et al. modulated by interpretation. Our results are consistent with the proposal that vlPFC is part of a system that facili-tates the construction of knowledge about people by relating past experiences with the personality of the per-son (Ranganath & Ritchey, 2012). We found that shared...
[45] (src:16cdf350) Third, with our self- relevance and valence models, we were able to decode the respective content dimension scores during free- thinking and resting. Different from recent efforts to decode semantic features directly from brain activity (25, 67, 68), we targeted the affective dimensions of thought, ...
[46] (src:b55f6f4f) applied it to thousands of new words. Our results suggest that both semantic categories and relations are represented by spatially overlapping cortical patterns, instead of anatomically segregated regions. Semantic relations that reflect conceptual progression from concreteness to abstractness are r...
[47] (src:b55f6f4f) Discussion Using fMRI data from subjects listening to natural story stimuli, we established a predictive model to map the cortical repre-sentations of semantic categories and relations. We found that semantic categories were not represented by segregated cortical regions but instead by distributed a...
[48] (src:7b279216) I n everyday settings, such as watching a movie or listening to a lecture, it is necessary to accumulate and integrate information over many minutes. We have previously identified a set of high-order brain areas, including the temporal parietal junction, angular gyrus, precuneus, posterior cingulate...
[49] (src:7b279216) The ISFC approach uncovered two novel functional characteristics of DMN correlation patterns. First, DMN correlation patterns were less reliable when the story was scrambled at the paragraph level, and even less so when the story was scrambled at the word level. This suggests that DMN correlations w...
[50] (src:7b279216) The ISFC patterns were specific for different moments in time and also highly reproducible across two independent groups of subjects. Interestingly, the reproducibility of ISFC patterns was observed both when the mean ISFC across all nodes was high and when it was low: that is, reproducibility was h...
[51] (src:a24af881) nature human behaviour Article https://doi.org/10.1038/s41562-022-01516-2 Evidence of a predictive coding hierarchy in the human brain listening to speech Charlotte Caucheteux 1,2 , Alexandre Gramfort1,2 & Jean-Rémi King 1,3 Considerable progress has recently been made in natural language proc...
[52] (src:a24af881) Predictive coding theory25–27 offers a potential explanation to these shortcomings; while deep language models are mostly tuned to predict the very next word, this framework suggests that the human brain makes predictions over multiple timescales and levels of repre-sentations across the cortical hi...
[53] (src:a24af881) Heschl’s gyri: Δk* = 2.5 ± 0.3, P < 0.001) and observed in both the left and right hemispheres (Fig. 3b). Together, these results suggest that the long-range predictions of frontoparietal cortices are more contextualized and of higher level than the short-term predictions of low-level brain regions....
[54] (src:a24af881) Overall, these results reveal multiple levels of predictions in the brain in which the superior temporal cortex predominantly pre-dicts short-term, shallow and syntactic representations whereas the inferior-frontal and parietal areas predominantly predict long-term, contextual, high-level and semant...
[55] (src:a24af881) colour-coded as in Fig. 2c). b, Same as a but with k* averaged across the voxels of nine regions of interest, in the left (circle) and right (triangle) hemispheres. Scores were averaged across individuals (n = 304) and the boxplot summarizes the distribution of the effect obtained on ten distinct an...
[56] (src:b55f6f4f) Our results suggest that DMN is involved in cortical processing of not only concepts but also semantic relations. This finding underscores the fact that DMN plays an active role in language and cognition10,48–51, rather than only a task-negative and default mode of brain function32. In particular, s...
[57] (src:b55f6f4f) concepts. In particular, the default mode network plays a central role in semantic processing for abstraction of concepts. https://doi.org/10.1038/s41467-020-15804-w OPEN 1 Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, USA. 2Department of Mathemati...
[58] (src:b0437311) https://doi.org/10.1038/s42003-024-07434-5 Neural processing of naturalistic audiovisual events in space and time Check for updates Yu Hu 1,2 & Yalda Mohsenzadeh 1,2,3 Our brain seamlessly integrates distinct sensory information to form a coherent percept. However, when real-world audiovisual events...
[59] (src:b0437311) To address this, we employed naturalistic video stimuli that capture common audiovisual events, and examined neural representations using multivariate pattern analysis on both functional magnetic resonance ima-ging (fMRI) and electroencephalogram(EEG)data.Weused computational models to characterize ...
[60] (src:b0437311) visual and acoustic features and high-level categorical and semantic infor-mation as well as when they were processed. By examining neural processes involved for different types of information, our results suggest two different stages of cross-modal interactions, with their associated brain areas, t...
[61] (src:b0437311) The searchlight results (Fig. 3) confirm that the categorical and semantic information wasmainly represented in high-level visual, auditory, and multisensory regions. Because these multisensory areas in the superior temporal cortex are established regions to integrate audiovisual input54, our result...
[62] (src:b0437311) representations and achieve higher task performance113. To evaluate the hierarchical correspondence between the DNN model and the brain, we extracted model activations from seven blocks of each model branch and used them to construct RDMs. We observed that representations in early layers of video an...
[63] (src:b0437311) Currently, DNNmodels serve as the best models of the human visual or auditory system140–144. However, their similarity with human brain responses in multisensory perception is less explored145. Generally, the match between DNN models and the brain depends on multiple factors, such as the training da...
Sources used: 16 documents
- 67a91294-ea55-4ee2-b94f-1676af02528a
- 4927f9ca-1d17-4fe4-8e75-86ba9a89c908
- a24af881-c199-40d5-81a3-cf38ddba7e81
- b55f6f4f-2b04-43c8-ae32-7b4e9444b046
- 7b279216-6743-406c-9009-741ebfeed269
- 16cdf350-d14a-4c5b-9e40-adad31798c82
- 72df40b6-ffb6-472b-8fa6-0ccf2bc869cc
- 3fa91849-c249-4c08-9d85-9648a1cda227
- 1fcbb92e-963b-4822-8c84-f9e4a6eece51
- a0cff8d4-d48c-48bf-8910-4039b750706f