image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 13:Qualitative results of Task I.C (Removing specific real-scenario degradation based on given prompt in driving). Prompt (1)-(5) are illustrated in Table10. Our method adeptly removes various degradations in response to specific prompts, resulting in generated images that are fundamentally consistent with the ground truth.
|
2403.10520v1
| 13 |
|
Figure 14:Qualitative results of Task II: Real-world bad
weather removal. We showcase the performance of methods
across three real-world nature datasets, each encompassing distinct weather conditions: rain streak, raindrop, and snow. Cases (1) to (2) feature rain streaks; cases (3) to (5) showcase raindrops; cases (6) to (8) are characterized by snow. Please zoom in to view the details.
|
2403.10520v1
| 14 |
|
Figure 15:Additional qualitative results of Task III.A : Image restoration in multi-degradation removal. Cases are: (1) rain streak; (2) rain streak + snow + moderate haze + raindrop; (3) flare + reflection + shadow; (4) fence + watermark; (5) rain streak + shadow + fence; (6) rain streak + snow + moderate haze + raindrop + flare + shadow + fence + reflection + watermark. Ours proposed CBDNet surpasses BIDeN in all cases, effectively restoring images with low visibility.
|
2403.10520v1
| 15 |
|
Figure 16:Additional qualitative results of Task III.B: Degradation masks reconstruction in multi-degradation removel. The input images contains all nine degradations. BIDeN is not able to effectively separate the various degradation components, which is reflected in the failed reconstruction of the degradation masks, such as snow, flare and shadow. Our proposed CBDNet successfully separates and reconstructs all components.
|
2403.10520v1
| 16 |
|
Figure 17:Additional qualitative results of Task III.C: Controllable blind image decomposition in multi-degradation removal. Prompt (1): remove the haze and the raindrop; Prompt (2): generate the foggy scene; Prompt (3): retain the scene image and the snow; Prompt (4): keep the background and all weather components; Prompt (5): compose image with watermark; Prompt (6): restore scene, flare and shadow together; Prompt (7): remove watermark, reflection and shadow.
|
2403.10520v1
| 17 |
|
Figure 1:Few-shot results averaged across eight test sets, including ILSVRC-2012[14,44]. We use cached features from an L/16 model[16]pretrained on JFT-3B[56](left) or WebLI[5]following a sigmoid language-image pretraining (SigLIP)[57](right). Our method,i.e., a multi-head attention pooling[30]head trained with weight decay (MAPwd) and frozen feature augmentation (FroFA), shows significant gains across all shots with respect to a weight-decayed MAP,i.e., MAPwd, or an L2-regularized linear probe baseline, both without FroFA.
|
2403.10519v2
| 1 |
|
Figure 5:Average top-1 accuracy for patch dropout FroFAonourILSVRC-2012 test set. We use the L/16 JFT-3B base setup (cf.Sec.5). We sweep across a base sweep (cf.Sec.4.4) to first find the best setting onourILSVRC-2012 validation set for each number of patches (cf.Sec.A3.2). Shaded areas indicate standard errors collected via sampling each shot five times.
|
2403.10519v2
| 12 |
|
Figure 6:Top-1 accuracy for channel variants (c/c2) of brightness FroFAonourILSVRC-2012 test set. We use the L/16 JFT-3B base setup (cf.Sec.5). We sweep across a base sweep (cf.Sec.4.4) to first find the best setting onourILSVRC-2012 validation set for each brightness level (cf.Sec.A3.2). Shaded areas indicate standard errors collected via sampling each shot five times
|
2403.10519v2
| 13 |
|
Figure 1:Overview of VideoAgent.Given a long-form video,VideoAgentiteratively searches and aggregates key information to answer the question. The process is controlled by a large language model (LLM) as the agent, with the visual language model (VLM) and contrastive language-image model (CLIP) serving as tools.
|
2403.10517v1
| 1 |
|
Figure 2:Detailed view of VideoAgent’s iterative process.Each round starts with the state, which includes previously viewed video frames. The large language model then determines subsequent actions by answering prediction and self-reflection. If additional information is needed, new observations are acquired in the form of video frames.
|
2403.10517v1
| 2 |
|
Figure 3:(Left) Frame efficiency compared to uniform sampling and previous methods.X-axis is in log scale. Our method achieves exceptional frame efficiency for long-form video understanding.(Right) Number of frames for different types of NExT-QA questions.Min, mean, max, distribution are plotted.VideoAgentselects more frames on questions related to temporal reasoning than causal reasoning and descriptive questions.
|
2403.10517v1
| 4 |
|
Figure 4:Case study on NExT-QA.VideoAgentaccurately identifies missing information in the first round, bridges the information gap in the second round, and thereby makes the correct prediction.
|
2403.10517v1
| 7 |
|
Figure 5:Case study on hour-long videos.VideoAgentaccurately identifies the key frame during the second iteration, subsequently making an accurate prediction. Conversely, GPT-4V, when relying on 48 uniformly sampled frames up to its maximum context length, does not get successful prediction. However, by integrating the frame pinpointed byVideoAgent, GPT-4V is able to correctly answer the question.
|
2403.10517v1
| 8 |
|
Figure 2:The FeatUp training architecture. FeatUp learns to upsample features through a consistency loss on low resolution “views” of a model’s features that arise from slight transformations of the input image.
|
2403.10516v2
| 2 |
|
Figure 6:FeatUp can upsample the features of any backbone, even convnets with aggressive nonlinear pooling.
|
2403.10516v2
| 6 |
|
Figure 10:Ablation of FeatUp’s training hyper-parameters. We are robust to a range of jitter values, though features degrade with large changes in max pad.
|
2403.10516v2
| 10 |
|
Figure 11:Qualitative ablation study of the TV and magnitude regularizers. FeatUp is fairly robust to the settng of these parameters.
|
2403.10516v2
| 11 |
|
Figure 12:Visualizing higher PCA components with FeatUp. FeatUp upsamples entire feature maps, so their higher-order principal components also remain in the same space as the original features and are upsampled precisely. Higher components tend to separate more fine-grained object categories like the skater from the skateboard, and the trees from the background, and the clouds from the sky. Note that each subobject’s features are upsampled precisely to the object it represents.
|
2403.10516v2
| 12 |
|
Figure 13:Visualization of downsampler salience and weight and bias kernels for two images. Note how fine-grained objects have higher salience and regions around important objects (like the sky between the hands and the skateboard) have lower salience. This allows the network to capture nonlinear behavior where embeddings from salient regions dominate the embeddings of other regions.
|
2403.10516v2
| 13 |
|
Figure 14:An example predicted uncertainty map for a set of ViT features. White areas have higher uncertainty. In this figure, we can see that nonlinear artifacts like the spurious pink tokens are marked with high uncertainty as they change location depending on the given evaluation. These tokens might serve some other role in the network, such as class-token-like information aggregation. We do not see these types of effects in DINO or convolutional networks.
|
2403.10516v2
| 14 |
|
Figure 16:We evaluate how floating point operations scale with various factors. In varying the upsampling factor, feature dimension, and target spatial dimension, FeatUp (JBU) remains competitive in GFLOP usage. For each experiment, the attributes not studied are kept constant (upsampling factor = 2, feature dimension = 256, starting spatial dimension = 8x8).
|
2403.10516v2
| 16 |
|
Figure 17:Analysis of peak memory usage (left) and inference time (right) for various forward-pass upsamplers. FeatUp (JBU) is competitive with SAPA and resize-conv across upsampling factors and is more efficient than CARAFE for smaller factors. The large image and strided approaches become infeasible at large upsampling factors we only show metrics for these methods up to8×8\times8 ×upsampling.
|
2403.10516v2
| 17 |
|
Figure 22:Left: Though FeatUp’s implicit network can capture fine detail such as the soccer ball or window frame, it can still produce some halo effects (see soccer player). Additionally, because the method relies on the input image’s spatial signal, certain patterns unrelated to object semantics can be transferred to the feature map (see rug pattern), though this is a rare occurrence. Right: FeatUp’s JBU network is not as sensitive to fine detail as the implicit network, instead capturing broader contours.
|
2403.10516v2
| 22 |
|
Figure 1:Results of our proposed multi-person
and temporal transformer architecture for joint gaze following and social gaze prediction, namely Looking at Humans (LAH), Looking at Each Other (LAEO), and Shared Attention (SA). For each person, the social gaze task is listed with the associated person ID (e.g.in frame 1 for person 2, they are in SA with person 4). More qualitative results can be found in the supplementary.
|
2403.10511v1
| 1 |
|
Figure 2:Proposed architecture for multi-person temporal gaze following and social gaze prediction. First, the Person Module (left) processes the set of head crops and bounding boxes to extract a sequence of person tokens for each person. In parallel, the ViT tokenizer processes the sequence of frames to extract frame tokens. Next, the Interaction Module (middle) jointly processes the person and frame tokens, iteratively updating them through people-scene interactions and spatio-temporal social interactions. Finally, the Prediction Module (right) processes the resulting frame and person tokens to infer a sequence of gaze heatmaps and in-out gaze labels for each person, as well as pair-wise social gaze labels for LAH, LAEO, and SA.
|
2403.10511v1
| 2 |
|
Figure 3:An illustration of the few cases where the predicted gaze point does not match with the predicted LAH label.
The uncertainty in the gaze target is reflected in the heatmap, while the uncertainty in the LAH target is reflected in the LAH scores.
|
2403.10511v1
| 3 |
|
Figure 0:Example egocentric visual (top-left) and whole-body tactile (right) observations when the humanoid interacts with a package in thetruckenvironment. In the right figure, the two cameras on the robot head are highlighted in green, while continuous tactile pressure readings are indicated in shades of red (strong pressure) and yellow (mild pressure). Note that for ease of visualization, we are not showing shear forces and tactile readings on the back of the robot, which are also implemented in our environment.
|
2403.10506v2
| 1 |
|
Figure 1:(a) System of spheres; (b) System of spheres with contractionϕα÷subscriptsuperscriptitalic-ϕ𝛼\phi^{\div}_{\alpha}italic_ϕ start_POSTSUPERSCRIPT ÷ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_α end_POSTSUBSCRIPTshaded.
|
2403.10502v1
| 1 |
|
Figure 2:System of spheres for severe withdrawal showingϕα÷÷subscriptsuperscriptitalic-ϕ𝛼\phi^{\mathbin{{\div}\mkern-10.0mu{\div}}}_{\alpha}italic_ϕ start_POSTSUPERSCRIPT ÷ ÷ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_α end_POSTSUBSCRIPTshaded.
|
2403.10502v1
| 2 |
|
Figure 3:System of spheres whereϕα∧β÷κ≤ℙβsubscriptℙsubscriptsuperscriptitalic-ϕsubscript𝜅𝛼𝛽𝛽\phi^{\div_{\kappa}}_{\alpha\land\beta}\leq_{\mathbb{P}}\betaitalic_ϕ start_POSTSUPERSCRIPT ÷ start_POSTSUBSCRIPT italic_κ end_POSTSUBSCRIPT end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_α ∧ italic_β end_POSTSUBSCRIPT ≤ start_POSTSUBSCRIPT blackboard_P end_POSTSUBSCRIPT italic_βholds.
|
2403.10502v1
| 3 |
|
Figure 3:Our ImageNet-T samples. We show the gold class (upper) and the target class (lower) of each sample.
|
2403.10499v1
| 9 |
|
(b)Percentage of overlapped images and corresponding accuracies on ImageNetV2 and Stylized ImageNet. Left y-axis: Accuracy. Right y-axis: Percentage.
|
2403.10499v1
| 18 |
|
Figure 10:Summary of producing an automated prompt for CLIP-Auto. A prompt is created using a template that combines a set of trigger tokens with the label text. The trigger tokens are shared across all classes and decided via a gradient-based search. At each search step, we use the current trigger tokens and each label text to synthesize a linear classifier, then compute the gradients for candidate trigger tokens and update the trigger tokens with that lead to the smallest loss. After iteratively repeating this process, the trigger tokens converge and are returned to create a prompt.
|
2403.10499v1
| 33 |
|
(a)Seven natural distribution shifts. Examples are from ImageNetV2, ImageNet-A, ImageNet-R, ImageNet Sketch, ImageNet-Vid, ObjectNet, and Youtube-BB.
|
2403.10499v1
| 37 |
|
(a)Seven natural distribution shifts. Examples are from ImageNetV2, ImageNet-A, ImageNet-R, ImageNet Sketch, ImageNet-Vid, ObjectNet, and Youtube-BB.
|
2403.10499v1
| 38 |
|
(b)Three synthetic distribution shifts. Examples are from ImageNet-C, video frames of ImageNet-P, and Stylized ImageNet.
|
2403.10499v1
| 39 |
|
Figure 1:Polynomial barrier certificate computed using SOSTOOLS depicted forϕ=0ϕ0\upphi=0roman_ϕ = 0. The unsafe regions𝕏usubscript𝕏𝑢\mathbb{X}_{u}blackboard_X start_POSTSUBSCRIPT italic_u end_POSTSUBSCRIPTare indicated in red and the initial region𝕏0subscript𝕏0\mathbb{X}_{0}blackboard_X start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPTin black. The two gray planes represent the level setsη𝜂\etaitalic_ηandγ𝛾\gammaitalic_γ.
|
2403.10497v1
| 1 |
|
Figure 2:Robust safety probability via BCs generated for radiiε𝜀\varepsilonitalic_ε, averaged over ten samples of sizeN=104𝑁superscript104N=10^{4}italic_N = 10 start_POSTSUPERSCRIPT 4 end_POSTSUPERSCRIPTand a constantB¯=0.1¯𝐵0.1\bar{B}=0.1over¯ start_ARG italic_B end_ARG = 0.1.
|
2403.10497v1
| 2 |
|
Figure 1:(a) shows an example of dialogue hallucination generated by an LVLM (e.g., LLaVALiu et al. (2023c)) for a test example in ScienceQA dataset; (b) shows the average performance drop of LLaVA and AIT on EvalDial for VQA and Captioning tasks with prepended adversarial dialogues.
|
2403.10492v3
| 1 |
|
Figure 2:Overview of dialogue hallucinations onEvalDial. A test example on ScienceQA that LLaVA originally answers correctly becomes hallucinated after three types of prepended dialogues, i.e., General, Random, and Adversarial.
|
2403.10492v3
| 2 |
|
Figure 3:shows the overview of AQG, generating an adversarial dialogueXdialogueadvsubscriptsuperscriptX𝑎𝑑𝑣dialogue\text{X}^{adv}_{\texttt{dialogue}}X start_POSTSUPERSCRIPT italic_a italic_d italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT dialogue end_POSTSUBSCRIPT(in yellow box) to hallucinate the answerXaadvsubscriptsuperscriptX𝑎𝑑𝑣a\text{X}^{adv}_{\texttt{a}}X start_POSTSUPERSCRIPT italic_a italic_d italic_v end_POSTSUPERSCRIPT start_POSTSUBSCRIPT a end_POSTSUBSCRIPT(in green box) by incorporating an extra LVLM into the optimization of adversarial attack
|
2403.10492v3
| 3 |
|
Table 3:Comparison of AQG with different attacking methods. The lower, the more effective in attacking.
|
2403.10492v3
| 6 |
|
Figure 6:Embedding distribution analysis for three types of dialogues. (a) illustrates the TSNE plot of each dialogue example with its corresponding target image, question, and answer on the joint text-image embedding space of InstructBLIP. (b) shows the density plot of three types of dialogue in terms of the average cosine similarity to the corresponding target image, question, and answer.
|
2403.10492v3
| 8 |
|
Figure 10:Examples of generated adversarial questions using GPT-4 and AQG on VQA and Captioning tasks.
|
2403.10492v3
| 17 |
|
Figure 1:(Top) An illustration of the vanilla multimodal transformer fusion architecture in the case of two input sources, A and B. (Bottom) Our proposed JMT fusion (in red) relies on joint multimodal representations.
|
2403.10488v3
| 1 |
|
Figure 2:An overview of the proposed joint multimodal transformer model for A-V fusion. The audio and visual modalities are cross-attended using transformer blocks. The JMT block also takes in the joint representation (shown with red arrows). The output of the cross-attended features is concatenated, and an FC layer is used for valence/arousal prediction.
|
2403.10488v3
| 2 |
|
Figure 3:Illustration of the proposed joint multimodal transformer architecture used for the Biovid pain estimation task. The blue branch shows the visual backbone, and the yellow branch is the physiological backbone. The joint representation is shown with a red block. The three feature vectors are fed into the joint transformer block.
|
2403.10488v3
| 3 |
|
Figure 4:Visualization of attention weights for visual and physiological modalities on the Biovid heat pain database. The facial frames are taken 1400 msec each.
|
2403.10488v3
| 4 |
|
Figure 1:Episode reward comparison between competitive and non-competitiveWalker2denvironment. The performance ofWalker2ds trained with competition can reach 120%percent\%%of the baseline.
|
2403.10487v1
| 1 |
|
Figure 2:Framework for learning knowledge from comparative information. Wherea𝑎aitalic_adenotes actions,s𝑠sitalic_srepresents the proprioceptive state, ando𝑜oitalic_ois the competitive observation.
|
2403.10487v1
| 2 |
|
Figure 3:The proposed Self-Interest Competition EnvironmentsMultiagent-Race. Three walker2d robots racing is illustrated as an example.
|
2403.10487v1
| 3 |
|
Figure 1:An illustration of the nullspace in three cases (projection function case, left top; linear function case, left bottom; vision transformer case, right).
For the functions in these three cases, there exists somenullspace, and theoutputof the function with respect to theinputwill remain
the same no matter how much perturbation is introduced
to theinputalong thenullspace.
Also, thenullspaceis function-specific (model-specific) and will not vary for different samples.
|
2403.10476v1
| 1 |
|
(b)ℓ2subscriptℓ2\ell_{2}roman_ℓ start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTnorm of learned noise under different regularization strengths
|
2403.10476v1
| 8 |
|
Figure 4:Change trend of multiple metrics with training steps."Adversarial" is the average performance on "FGSM" and "DamageNet" settings, "OOD" is the average score of the six OOD datasets, and "avg" is the total average. All values are divided by their initial values to show the trend more clearly.
|
2403.10476v1
| 9 |
|
(b)Influence ofϵitalic-ϵ\epsilonitalic_ϵnoise under convex combination with differentα1,α2subscript𝛼1subscript𝛼2\alpha_{1},\alpha_{2}italic_α start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , italic_α start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT
|
2403.10476v1
| 12 |
|
Figure 1:Stages of Tumour metastasis illustrated in AJCC TNM staging system. Carcinoma-in-situ (CIS), Ta and T1 are non-muscle invasive stages and T2 - T4 are muscle invasive stages. CIS: primary stage where tumour is confined to inner bladder lining. Ta: tumour limited to epithelium. T1: tumour reaches the lamina propria. Stage III (T2): tumour invades into bladder wall muscle. T3: tumour spreads to the fat around the bladder. Stage IV (T4): tumour spreads to nearby pelvic organs/tissues.
|
2403.10586v3
| 1 |
|
Figure 2:Distribution of major risk factors for bladder cancer. Tobacco smoking accounts for approximately 50% of cases, occupational exposures contribute to 18%, and other factors (including exposure to arsenic, chronic bladder inflammation, previous radiation or chemotherapy, diet, and genetic predisposition) comprise the remaining 32%. Data synthesized from multiple sources(24,26,29,25,27).
|
2403.10586v3
| 2 |
|
Figure 3:This graph shows the overall growth in bladder cancer recurrence studies and the exponential rise in ML-based approaches. While the total number of bladder cancer studies has increased linearly over the past two decades, the adoption of ML methods has grown at a much faster, exponential rate. This trend suggests that ML approaches are becoming increasingly relevant in the field, with future research likely to continue emphasizing advanced computational methods.
|
2403.10586v3
| 3 |
|
Figure 6:The distribution of machine learning models used in NMIBC recurrence prediction studies included in this review. More complex models, such as NNs and SVMs, dominate the research landscape, while simpler, interpretable models like logistic regression are used less frequently
|
2403.10586v3
| 6 |
|
Figure 1:The U-matrix is the main output of a Self Organizing Map. Instead of visualizing the4444-dimensional pointer at each of the16×16161616\times 1616 × 16nodes of the unit layer, the U-matrix indicates the mean distance between each node’s pointer and the pointers of all its neighboring nodes. Nodes in the corners have3333neighbors, the other nodes along the fringe have5555, and nodes in the middle have8888neighbors. The single items used to train the SOM (techno music) are shown as colored dots, whereby different colors represent different techno music styles.
|
2404.00016v2
| 1 |
|
Figure 3:The SOMson interface with the visualizations on the left and the sonification parameters on the right. Buttons allow switching between U-matrix and the component planes, and showing/hiding the training data.
|
2404.00016v2
| 3 |
|
Figure 4:Bezold effect: Colors and shades may appear different depending on their surrounding colors. Here, all three notes have the same (single) color, even though the one on the left may appear darker, and the one in the middle may seem to have a color gradient.
|
2404.00016v2
| 4 |
|
Figure 5:The Bezold effect also holds for lightness: All notes have the exact same (single) color and lightness level, even though the one on the left may appear darker and the one in the middle appears as if it had a lightness gradient.
|
2404.00016v2
| 5 |
|
Figure 1:the GSN diagram is the start of a holistic safety case6. The decomposition above would occur in step 2 (specifying unacceptable outcomes). ‘G’ labeled rectangles represent subclaims (i.e. goals). ‘S’ labeled parallelograms indicate justification strategies. For an example of a full safety case, see section6.
|
2403.10462v2
| 1 |
|
Figure 2:As AI systems become more powerful, developers will likely increasingly rely on arguments toward the right in the plot above.
|
2403.10462v2
| 2 |
|
Figure 3:building block arguments for making safety cases. See section5for detailed descriptions of each building block argument and justifications for their labels. The arguments that rank highest on practicality, strength, and scalability aremonitoring,externalized reasoning, andtestbeds.
|
2403.10462v2
| 3 |
|
Figure 4:The diagram above is a ‘risk matrix’ used in the aviation industry(ICAO,,2016). If boxes in red are checked, risk is considered unacceptably high.
|
2403.10462v2
| 4 |
|
Figure 5:We recommend that regulators review risk cases along with safety cases, effectively putting AI systems “on trial.”
|
2403.10462v2
| 5 |
|
Figure 6:We motivate our four argument categories with an analogy to how one might justify that a robber won’t steal a diamond.
|
2403.10462v2
| 6 |
|
Figure 7:the diagram above illustrates how developers could evaluate easier versions of tasks to demonstrate a ‘margin of safety.’ If AI systems can only perform the second (yellow) task, then the “margin of safety” is the difference in the difficulty of this task and autonomously executing a cyberattack in the wild (dark red).
|
2403.10462v2
| 7 |
|
Figure 8:Caption: The diagram above illustrates an example monitoring scheme where a watchdog model (right) monitors the agent model (left). In this setup, a control argument could be constructed by conservatively assuming that the agent and its users are pursuing their best strategy for causing a catastrophe (untrustworthy). The watchdog is assumed to be reliable, which must be justified with a trustworthiness argument.
|
2403.10462v2
| 8 |
|
Figure 9:the diagram above depicts a potential setup where three different models would need to behave badly to cause a catastrophe.
|
2403.10462v2
| 9 |
|
Figure 10:The diagram above shows how not-alignment-faking arguments could be bootstrapped up to strongly superhuman AI systems. This particular strategy is called iterated amplification(Christianoet al.,,2018). AI systems can be ‘amplified’ by creating additional copies of them or speeding them up (see sectionA.3.6).
|
2403.10462v2
| 10 |
|
Figure 11:The key above indicates what each shape in a GSN diagram represents. A ‘goal’ is a claim to be justified.
|
2403.10462v2
| 11 |
|
Figure 12:Zoom in to read the claims and evidence depicted in the GSN tree. The colors in the diagram above show how different parts of the argument correspond to steps in our six-step framework (GSN diagrams are not typically color-coded in this way).
|
2403.10462v2
| 12 |
|
Figure 13:Developers will likely respond to failed catastrophe attempts to guard against subsequent attempts, which is a key dynamic to consider when estimating risk from unilateral subsystems.
|
2403.10462v2
| 13 |
|
Figure 14:the “one and done” response policy described previously implies that a catastrophe occurs if and only if a catastrophe attempt succeeds before an attempt is caught.
|
2403.10462v2
| 14 |
|
Figure 15:the above is a hypothetical setup where a watchdog model filters the actions of an agent. To make an argument this subsystem is safe, developers first place probabilities that different combinations of models reliably behave as intended. They then iterate through these combinations and evaluate whether the unsafe models could cause a catastrophe by conservatively assuming they (and human users) will pursue their best available strategies for doing so.
|
2403.10462v2
| 15 |
|
Figure 16:The figure lists examples of strategies that AI systems could pursue collectively at a large scale to cause a catastrophe (right) as well as reasons AI systems might pursue them.
|
2403.10462v2
| 16 |
|
Figure 5:Accuracy of Fine Tuning vs EWC on the original training set (without adversarial perturbations) during Adversarial Continuous Training. The solid black line represents the 10 training epochs before deployment, while the dotted lines represent the accuracy after each adversarial training session.
|
2403.10461v2
| 5 |
|
Figure 6:Effect of Continuous Adversarial Training: Accuracy Comparison on (Dtestsubscript𝐷testD_{\text{test}}italic_D start_POSTSUBSCRIPT test end_POSTSUBSCRIPT) of perturbed Enron SPAM (Fine Tuning vs EWC)
|
2403.10461v2
| 6 |
|
Figure 1:Quantum circuit illustrating the inversion test.ϕ(x)italic-ϕ𝑥\phi(x)italic_ϕ ( italic_x )represents the quantum feature map. The probability of the all-zero is estimated as the kernel.
|
2404.00015v3
| 1 |
|
Figure 2:Basic examples of a translation between individual in the population and its circuit representation.
|
2404.00015v3
| 2 |
|
Figure 5:Fidelity kernels obtained using 1000 individuals, 50 generations and 5 qubit for data mapping for the 2000 down-sampled dataset.
|
2404.00015v3
| 7 |
|
Figure 1:Examples on solving noisy image inverse problems on ImageNet validation set using our proposed method without task specific model finetuning or training.
|
2403.10585v1
| 1 |
|
Figure 2:Evolution of the reconstruction loss‖𝐲−𝒜(𝝁i(𝐱i))‖2subscriptnorm𝐲𝒜subscript𝝁𝑖subscript𝐱𝑖2\|{\bf y}-\mathcal{A}({\bm{\mu}}_{i}({\bf x}_{i}))\|_{2}∥ bold_y - caligraphic_A ( bold_italic_μ start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ( bold_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ) ) ∥ start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTof the DPS and DPG method.
|
2403.10585v1
| 2 |
|
Figure 4:Results on solving linear noisy inverse problems (inpainting, super-resolution and Gaussian deblurring) on ImageNet Dataset. The input image is distorted by random Gaussian noiseσ𝐲=0.05subscript𝜎𝐲0.05\sigma_{\bf y}=0.05italic_σ start_POSTSUBSCRIPT bold_y end_POSTSUBSCRIPT = 0.05.
|
2403.10585v1
| 4 |
|
Figure 8:Evolution of the reconstruction error in the image restoration process of super-resolution (left), Gaussian Deblurring (middle) and Motion Deblurring (Right).
|
2403.10585v1
| 8 |
|
Figure 1:Overview of our system design. The process begins with aWeb Crawler, which consists of a Recursive Crawler and a Filter to gather raw data. This raw data is sanitized and stored in an S3 storage bucket. TheDataset Generatorrelies on an Annotation Model to prompt annotations and filter datapoints, generating finetune datasets for all models in the system, i.e. Annotation, Embedding, Reranking, and Core Models. For query processing, theContext Retrieverembeds the user question and retrieves relevant contexts, which are re-ranked and rewritten if necessary. TheGeneration modulethen utilizes the Core Model and a Prompt Template to generate a system answer through a QA Chain, resulting in an answer that leverages both retrieved information and generative capabilities for accurate and context-aware responses.
|
2403.10446v1
| 1 |
|
Figure 2:Hierarchical knowledge base file system keeps the structural information of the relation between original files, enrich the semantic providing to the retriever.
|
2403.10446v1
| 2 |
|
Figure 3:Overview of RAG QA Pipeline, which can be divided into retrieval phase and generation phase. In the retrieval phase, the retriever fetches top 5 reference chunks with maximum similarity in terms ofmmrscore, which is then sent to reranking model to prioritizing the most relevant information for the given user inquery. In the generation phase, the generative model takes the rewritten prompt as input and completes the answer.
|
2403.10446v1
| 3 |
|
Figure 4:Recall, F1 Score, Cosine Similarity and BELU on our local test question-answer dataset under different settings. Note the data in the chart is normalized between 0 and 1 for better visibility. For original experiment output please refer to Table1.
|
2403.10446v1
| 4 |
|
Figure 1:One iteration of speculative decoding (Algorithm3). The prefixes and verified tokens are in blue, the unverified tokens from the draft model are in red, and the tokens sampled from the residual distribution are underlined.
|
2403.10444v2
| 1 |
|
Figure 3:Average block efficiency (BE) and wall clock speedup (WS) across all datasets for token verification (TokenV) and block verification (BlockV) with differentγ𝛾\gammaitalic_γ. The large model is PALM-2-S and the drafter model is either PALM-2-XXS (XXS) or PALM-2-XXXS (XXXS).
|
2403.10444v2
| 4 |
|
Figure 4:Average relative improvement of block verification over token verification in block efficiency (BE) and wall clock speedup (WS) across all datasets for different drafters and draft lengths.
|
2403.10444v2
| 6 |
|
Figure 1.The Data Ethics Emergency Drill logo based on the icon for meeting point. The project website can be found atwww.gooddeed.ai.
|
2403.10438v1
| 1 |
|
Figure 2.Overview of Drill Studies - The cut off circles for scenario crafting represent that these activities were incomplete before the third drill. The dashed arrows pointing from the follow up interviews to the drills represent that participants from all drill iterations were included in the follow up.
|
2403.10438v1
| 2 |
|
Figure 7.Average Ratings of each study for perceived realism of the scenario, ability to provoke new thoughts and to elicit a useful discussion. The bars show the mean ratings and the black lines mark ranges of lowest rating and highest rating.
|
2403.10438v1
| 6 |
|
Figure 1:A multilayer representation of a collective intelligence system.A multilayer representation can be used to untangle the processes within the complex system of human and AI agents. It consists of three interdependent layers: cognition, physical, and information. External factors and a changing environment can also influence the entire system’s emergent collective intelligence.
|
2403.10433v4
| 1 |
|
Figure 2:AI-CI cases by application area. Distribution of AI-enhanced collective intelligence cases by application area based on dataset curated by Supermind Design.
|
2403.10433v4
| 2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.