Paper deep dive
HyFI: Hyperbolic Feature Interpolation for Brain-Vision Alignment
Sangmin Jo, Wootaek Jeong, Da-Woon Heo, Yoohwan Hwang, Heung-Il Suk
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 99%
Last extracted: 3/26/2026, 1:38:03 AM
Summary
HyFI (Hyperbolic Feature Interpolation) is a novel framework for brain-to-image decoding that addresses the modality gap and feature entanglement by interpolating semantic and perceptual visual features within hyperbolic space. By leveraging the geometric properties of the Lorentz model, HyFI fuses and compresses visual information to better align with the limited expressiveness of neural signals, achieving state-of-the-art performance on THINGS-EEG and THINGS-MEG benchmarks.
Entities (5)
Relation Signals (3)
HyFI → improvesperformanceon → THINGS-EEG
confidence 100% · HyFI achieves state-of-the-art performance in zero-shot brain-to-image retrieval... on THINGS-EEG
HyFI → improvesperformanceon → THINGS-MEG
confidence 100% · HyFI achieves state-of-the-art performance in zero-shot brain-to-image retrieval... on THINGS-MEG
HyFI → utilizes → Lorentz model
confidence 100% · we adopt the Lorentz (hyperboloid) model due to its strong empirical performance in multi-modal tasks.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recent progress in artificial intelligence has encouraged numerous attempts to understand and decode human visual system from brain signals. These prior works typically align neural activity independently with semantic and perceptual features extracted from images using pre-trained vision models. However, they fail to account for two key challenges: (1) the modality gap arising from the natural difference in the information level of representation between brain signals and images, and (2) the fact that semantic and perceptual features are highly entangled within neural activity. To address these issues, we utilize hyperbolic space, which is well-suited for considering differences in the amount of information and has the geometric property that geodesics between two points naturally bend toward the origin, where the representational capacity is lower. Leveraging these properties, we propose a novel framework, Hyperbolic Feature Interpolation (HyFI), which interpolates between semantic and perceptual visual features along hyperbolic geodesics. This enables both the fusion and compression of perceptual and semantic information, effectively reflecting the limited expressiveness of brain signals and the entangled nature of these features. As a result, it facilitates better alignment between brain and visual features. We demonstrate that HyFI achieves state-of-the-art performance in zero-shot brain-to-image retrieval, outperforming prior methods with Top-1 accuracy improvements of up to +17.3% on THINGS-EEG and +9.1% on THINGS-MEG.
Tags
Links
- Source: https://arxiv.org/abs/2603.22721v1
- Canonical: https://arxiv.org/abs/2603.22721v1
Trouble viewing inline? Open PDF directly →
Full Text
75,544 characters extracted from source content.
Expand or collapse full text
HyFI: Hyperbolic Feature Interpolation for Brain-Vision Alignment Sangmin Jo, Wootaek Jeong, Da-Woon Heo, Yoohwan Hwang, Heung-Il Suk Corresponding author Abstract Recent progress in artificial intelligence has encouraged numerous attempts to understand and decode human visual system from brain signals. These prior works typically align neural activity independently with semantic and perceptual features extracted from images using pre-trained vision models. However, they fail to account for two key challenges: (1) the modality gap arising from the natural difference in the information level of representation between brain signals and images, and (2) the fact that semantic and perceptual features are highly entangled within neural activity. To address these issues, we utilize hyperbolic space, which is well-suited for considering differences in the amount of information and has the geometric property that geodesics between two points naturally bend toward the origin, where the representational capacity is lower. Leveraging these properties, we propose a novel framework, Hyperbolic Feature Interpolation (HyFI), which interpolates between semantic and perceptual visual features along hyperbolic geodesics. This enables both the fusion and compression of perceptual and semantic information, effectively reflecting the limited expressiveness of brain signals and the entangled nature of these features. As a result, it facilitates better alignment between brain and visual features. We demonstrate that HyFI achieves state-of-the-art performance in zero-shot brain-to-image retrieval, outperforming prior methods with Top-1 accuracy improvements of up to +17.3%17.3\% on THINGS-EEG and +9.1%9.1\% on THINGS-MEG. Code — https://github.com/ku-milab/HyFI Introduction Understanding how the human brain encodes information has long been a central topic in neuroscience, and has recently attracted growing attention in artificial intelligence. Specifically, the field of brain decoding aims to infer internal cognitive states or external sensory experiences from recorded brain activity (Naselaris et al. 2011; Oota et al. 2023). It offers insights into how the brain represents the external world and enables brain-computer interface (BCI) systems (Ko et al. 2021). In recent years, brain decoding techniques based on machine learning have been successfully applied across diverse cognitive domains such as vision, audition, and language (Wang and Ji 2022; Défossez et al. 2023; Scotti et al. 2024). Among these, visual brain decoding has received particular attention, given that vision is the dominant sensory modality in humans and plays a crucial role in perception and cognition (Mathis et al. 2024). Figure 1: (a) The human visual system processes perceptual and semantic information, and some degradation occurs when neural activity is recorded. (b) Previous works aligned semantic and perceptual features through separate pathways, overlooking their entanglement in brain signals. (c) In contrast, hyperbolic interpolation merges perceptual and semantic features with lower complexity, enhancing alignment with brain signals. Visual brain decoding has been extensively studied using neuroimaging modalities such as functional magnetic resonance imaging (fMRI), electroencephalography (EEG), and magnetoencephalography (MEG). In particular, fMRI has long been a dominant modality in this field due to its superior spatial resolution. However, its limited temporal resolution and bulky equipment make it less suitable for real-world applications. In contrast, EEG and MEG offer high temporal resolution, making it particularly suitable for BCI. Motivated by these advantages, we focus on brain decoding tasks based on EEG and MEG signals. Recent brain decoding studies have increasingly adopted dual-pathway frameworks to capture both perceptual details and semantic representations from neural signals. This approach reflects the human visual system, which processes both perceptual and semantic features, as shown in Fig. 1(a). Perceptual features refer to low-level visual attributes extracted in early visual areas (e.g., V1), such as orientation, color, and edge information (Miyawaki et al. 2008). Semantic features denote high-level conceptual representations encoded in cortical regions, such as object identity and category (DiCarlo et al. 2012). In recent approaches, semantic features are typically captured by aligning brain activity with image embeddings from pre-trained vision-language models (VLMs) like CLIP (Scotti et al. 2023). In parallel, perceptual features are often decoded by aligning representations derived from variational autoencoders (VAEs) (Shen et al. 2025). These efforts have advanced brain decoding by integrating brain-inspired models with multi-modal representations (Li et al. 2025). Despite these advances, current approaches still face two key limitations. First, aligning brain signals with image embeddings remains challenging—a problem commonly referred to as the modality gap (Liang et al. 2022). This issue is known to arise from an inherent information imbalance between modalities (Schrodi et al. 2024). In brain decoding, this imbalance is particularly pronounced, as neural signals contain substantially less semantic information than image embeddings for visual tasks (Chen et al. 2024; Wu et al. 2025). The limitation is driven by human attentional bottlenecks, restricted visual working memory (Cavanagh and Alvarez 2005; Dux and Marois 2009), and the low signal-to-noise ratio and resolution of neural recordings (Srinivasan et al. 2007; Naselaris et al. 2011). Second, many existing approaches that separately align perceptual and semantic features fail to reflect the fact that neural activity encodes features in a highly entangled and interactive manner, as they are not processed independently (Pollen 1999; Naselaris et al. 2009). As a result, attempts to align these features independently may lead to suboptimal performance. To address these limitations, we adopt hyperbolic space for brain-vision alignment. Unlike Euclidean space, hyperbolic geometry with negative curvature offers two key advantages: (1) geodesics between two points naturally bend toward the origin, (2) representational capacity decreases near the origin due to the exponential expansion of the space with radius. Building on these insights, we introduce a novel Hyperbolic Feature Interpolation (HyFI) method that interpolates semantic and perceptual visual features in hyperbolic space. This allows semantic and perceptual features to be effectively integrated and compressed during interpolation, making it well suited for brain signals with limited information and entangled semantic-perceptual components. As a result, the interpolated representations become better aligned with brain activity. Our main contributions are as follows: • We propose a hyperbolic interpolation method that effectively integrates and compresses semantic and perceptual visual features, explicitly accounting for the limited information capacity and entangled nature of brain signals. • Our method consistently improves performance across combinations of visual and brain encoders, demonstrating broad applicability. • It achieves state-of-the-art (SOTA) performance on two public brain decoding benchmarks, with 68.2%\% Top-1 accuracy on THINGS-EEG and 35.8%\% on THINGS-MEG, outperforming previous methods by +17.3%\% and +9.1%\%, respectively. Related Works Visual Brain Decoding Visual brain decoding has received increasing attention for its potential to uncover the mechanisms of human cognition and to enable practical BCI (Kay et al. 2008; Yang et al. 2024). With the rise of large-scale VLMs, recent studies have begun to leverage those representations by aligning semantic features with CLIP (Scotti et al. 2023; Song et al. 2024). A similar trend is observed in EEG-based visual decoding. However, the inherent modality gap between neural signals and pre-trained visual embeddings remains a major challenge. To mitigate this, Li et al. (2024) utilized a diffusion prior to map brain features into image space, Zhang et al. (2025) employed multi-modal fusion to increase shared information, and Wu et al. (2025) reduced visual complexity using Gaussian blur. These methods overlook the fact that semantic and perceptual representations are often entangled in neural activity. In contrast, our approach uses hyperboloid interpolation to fuse both features while reducing complexity, enabling better alignment. Hyperbolic Representation Learning Hyperbolic space has attracted attention in representation learning for its ability to model hierarchical data due to its negative curvature (Nickel and Kiela 2017; Chamberlain et al. 2017). This property has led to successful applications across various modalities with inherent hierarchies, including graphs (Liu et al. 2019), text (Dhingra et al. 2018; Tifrea et al. 2019), and images (Atigh et al. 2022). Recent works have further extended hyperbolic geometry to the multi-modal domain. For instance, Desai et al. (2023) propose hyperbolic vision-language models where image embeddings are constrained to lie within a concept cone defined by the text embeddings. Similarly, Pal et al. (2024) extend this framework to model hierarchical relations across multiple levels, including cropped-text, cropped-image, original text, and original image representations. Building on these insights, we extend hyperbolic representation learning to the domain of brain decoding. Specifically, we leverage the geodesic property of hyperbolic space to unify and compress semantic and perceptual visual information, providing a new perspective on aligning neural signals with rich visual stimuli. Figure 2: (a) The semantic image vsx_v^s and perceptual image vpx_v^p are encoded by CLIP and projected via a linear layer, and then lifted onto the hyperboloid via the exponential map. Using a learned weight t derived from the semantic image features, the two image features are interpolated on the hyperbolic manifold. Similarly, EEG inputs are encoded and projected onto the same hyperbolic space. Contrastive learning is then performed on the hyperboloid to bring paired EEG-image representations closer. (b) A schematic view of the hyperbolic embedding space. The interpolated representation ^v z_v lies along the geodesic between the semantic feature vsz_v^s and the perceptual feature vpz_v^p. Contrastive learning then pulls the EEG feature bz_b toward the target ^v z_v. Preliminaries We formulate our approach on hyperbolic space, a Riemannian manifold with constant negative curvature, where volume grows exponentially with radius. This property makes it well suited for representing hierarchical structures and addressing modality imbalance in multi-modal learning (Le et al. 2019; Peng et al. 2021). Following prior works (Desai et al. 2023; Pal et al. 2024), we adopt the Lorentz (hyperboloid) model due to its strong empirical performance in multi-modal tasks. Definition The Lorentz model nL^n of n-dimensional hyperbolic space with constant negative curvature is realized as the “upper sheet” of a two-sheeted hyperboloid in (n+1)(n+1)-dimensional Minkowski space (Cannon et al. 1997). Concretely, a point ∈ℝn+1p ^n+1 is represented as =(p0,~)p=(p_0, p), where p0>0p_0>0 denotes the time component and ~∈ℝn p ^n are spatial coordinates. The Lorentz manifold is defined as: n=∈ℝn+1:⟨,⟩=−1κ,p0>0,L^n= \p ^n+1: ,p _L=- 1κ,\ p_0>0 \, (1) where −κ∈ℝ-κ is the curvature of the space. The Lorentzian inner product for two vectors ,∈np,q ^n is defined as: ⟨,⟩=−p0q0+⟨~,~⟩, ,q _L=-p_0q_0+ p, q _E, (2) where ⟨⋅,⋅⟩ ·,· _E denotes the standard Euclidean dot product. Geodesics A geodesic is the shortest curve connecting two points on the manifold. In the Lorentz model, the geodesic distance between two points ,∈np,q ^n is defined as: d(,)=1/κ⋅cosh−1(−κ⟨,⟩).d_L(p,q)= 1/κ· ^-1 (-κ ,q _L ). (3) Exponential and Logarithmic Map The exponential map defines a smooth mapping from the tangent space onto the Lorentz manifold. For a point ∈np ^n, the tangent space is defined as: Tn=∈ℝn+1∣⟨,⟩=0.T_pL^n= \v ^n+1 ,v _L=0 \. (4) Given a tangent vector ∈Tnv∈ T_pL^n, the exponential map traces the geodesic from p in the direction of v and is parameterized as γ(t)=expκ(t)γ(t)= _p^κ(tv), where t∈[0,1]t∈[0,1]. It is explicitly defined as: expκ(t)=cosh(tκ‖)+sinh(tκ‖)κ‖, ^κ_p(tv)= (t κ\|v\|_L )p+ (t κ\|v\|_L ) κ\|v\|_Lv, (5) where ‖=⟨,⟩||v||_L= ,v _L. Conversely, a point ∈nq ^n on the hyperboloid can be projected onto the tangent space via the logarithmic map logκ(⋅):n→Tn _p^κ(·):L^n→ T_pL^n, as follows: logκ()=cosh−1(−κ⟨,⟩)(κ⟨,⟩)2−1(+κ⟨,⟩). _p^κ(q)= ^-1(-κ ,q _L) (κ ,q _L)^2-1 (q+κ ,q _Lp ). (6) In practice, the point p is commonly set to the time origin =(1/κ,0,…,0)⊤∈nO=( 1/κ,0,…,0) ^n. Under this setting, a vector =[0,enc]∈ℝn+1v=[0,v_enc] ^n+1 lies in the tangent space at O and can be mapped onto the hyperboloid via exponential map expκ(⋅) _O^κ(·), where encv_enc denotes the encoder output. Method Problem Formulation Given the brain signals space bX_b and the visual stimuli space vX_v, the goal of visual brain decoding is to map brain signals b∈bx_b _b into a shared space ℋH aligned with v∈vx_v _v. To this end, we learn a brain encoder fb:b→ℋf_b:X_b and a visual encoder Wfv:v→ℋWf_v:X_v , where fvf_v is a frozen backbone of pre-trained VLM and W is a linear layer. Specifically, we use n-dimensional Lorentz space nL^n as a semantically aligned space ℋH. Embeddings in nL^n are obtained via the exponential map expκ(⋅) _O^κ(·) at the time origin O. The visual embedding is defined as v=expκ(αv⋅Wfv(v))z_v= _O^κ( _v· Wf_v(x_v)) and the brain embedding b=expκ(αb⋅fb(b))z_b= _O^κ( _b· f_b(x_b)), where αv _v and αb _b are learnable projection scalars that reduce the norm of the embeddings to keep them near the origin O. An overview of our overall framework is illustrated in Fig 2(a). Hyperbolic Brain-Vision Contrastive Learning To align neural embeddings with corresponding visual embedding, we utilize contrastive learning in hyperbolic space. Unlike Euclidean space, hyperbolic space provides a natural embedding space for aligning modalities with asymmetric information capacity and has shown empirical success in multi-modal representation learning (Desai et al. 2023; Pal et al. 2024; Mandica et al. 2025). Given a batch of EEG-image pairs (b,i,v,i)i=1B\(z_b,i,z_v,i)\_i=1^B, hyperbolic contrastive learning is formulated as: ℒ(v,b)=−∑i∈Blogexp(d(v,i,b,i)/τ)∑k=1,k≠iBexp(d(v,i,b,k)/τ),L(z_v,z_b)=- _i∈ B (d_L(z_v,i,z_b,i)/τ )Σ _k=1,\,k≠ i^B (d_L(z_v,i,z_b,k)/τ ), (7) where B denote the batch, d_L denote the negative Lorentz distance, τ is temperature parameter. Hyperbolic Feature Interpolation With hyperbolic space established, we aim to learn visual representations aligned with brain signals by capturing their inherent properties. In particular, we fuse and compress semantic and perceptual visual features. Our approach is motivated by two key observations: (1) semantic and perceptual visual features are often entangled in neural activity, (2) and brain signals inherently contain less information than natural images. To address the first property, we describe how semantic and perceptual features are extracted from an image, then present a hyperbolic interpolation method. Finally, we show that this interpolation naturally leads to information compression, thereby addressing the second observation. Extracting Semantic and Perceptual Features We first apply image-level augmentations to obtain the semantic and perceptual visual inputs. The semantic image vsx_v^s is generated via fovea blur, simulating peripheral vision to preserve semantics and enhance alignment with brain signals (Wu et al. 2025). The perceptual image vpx_v^p is obtained by applying Gaussian blurring to suppress high-frequency components and retain coarse structure. This augmentation amplifies perceptual attributes in the CLIP embeddings. These augmentations and their effects are presented in Fig. 3. These inputs are then projected into hyperbolic space as semantic visual features vs=expκ(αv⋅Wsfv(vs))z_v^s= _O^κ( _v· W_sf_v(x_v^s)) and perceptual visual features vp=expκ(αv⋅Wpfv(vp))z_v^p= _O^κ( _v· W_pf_v(x_v^p)), respectively. Here, Ws,Wp∈ℝd×dW_s,W_p ^d× d are learnable matrices and d denotes the CLIP embedding dimension. Figure 3: Examples of image augmentations and retrieval results. The semantic image vsx_v^s and perceptual image vpx_v^p are generated via fovea blur and Gaussian blur, respectively. Retrieval results using CLIP embedding show that semantic queries return category-relevant matches (e.g., fruits), while perceptual queries retrieve images with similar low-level attributes such as color and shape. Hyperbolic Interpolation To perform interpolation of these features, we approximate the geodesic in the Lorentz model using the exponential map. The perceptual feature vpz_v^p is projected onto the tangent space at vsz_v^s using the logarithmic map, i.e., logvsκ(vp)∈Tvsn ^κ_z_v^s(z_v^p)∈ T_z_v^sL^n. This tangent vector is scaled and mapped back to the hyperbolic space via the exponential map. The resulting interpolated visual representation ^v z_v follows the geodesic from vsz_v^s to vpz_v^p: ^v=γvs→vp(t)=expvsκ(t⋅logvsκ(vp)), z_v= _z_v^s _v^p(t)= ^κ_z_v^s (t· ^κ_z_v^s(z_v^p) ), (8) where t∈[0,1]t∈[0,1] is the interpolation coefficient. We compute the interpolation coefficient t dynamically to reflect image-specific variation in the relative importance of semantic and perceptual features: t=σ(Wtfv(vs)),t=σ(W_tf_v(x_v^s)), (9) where Wt∈ℝ1×dW_t ^1× d is a learnable matrix and σ denotes the sigmoid function. Compression Effects The proposed interpolation mechanism concurrently compresses and fuses semantic and perceptual features. To analyze how this compression arises, we revisit the geodesic formulation in the Lorentz model. We reformulate the geodesic in Eq. (8) as follows: γ→(t)=sinh((1−t)β)sinh(β)+sinh(tβ)sinh(β), _p (t)= ((1-t)β ) (β)p+ (tβ ) (β)q, (10) where β=κ⋅d(,)β= κ· d_L(p,q). A detailed derivation is provided in the appendix. Unlike linear interpolation (1−t)+t(1-t)p+tq in Euclidean space, the hyperbolic interpolation weights sinh((1−t)β)sinh(β) ((1-t)β) (β) and sinh(tβ)sinh(β) (tβ) (β) are strictly smaller than (1−t)(1-t) and t, respectively. This causes the interpolated points to lie closer to the origin. To understand how this contraction relates to the geometry of hyperbolic space, we revisit the hyperboloid constraint in Eq. (1), which gives: p0=1/κ+‖~‖2,p_0= 1/κ+\| p\|^2, (11) A smaller p0p_0 constrains ‖~‖\| p\| more tightly, thus reducing the expressive capacity of embedding. This is consistent with prior work (Ganea et al. 2018; Khrulkov et al. 2020), suggesting that points near the origin represent more abstract concepts. Final Objective Function Finally, we train our model by aligning the interpolated visual representations with the brain embeddings in hyperbolic space. The final hyperbolic contrastive loss is defined as: ℒHCL=ℒ(^v,b)+ℒ(b,^v).L_HCL=L( z_v,z_b)+L(z_b, z_v). (12) Method Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 Subject 6 Subject 7 Subject 8 Subject 9 Subject 10 Average T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 Intra-subject: train and test on one subject BraVL 6.1 17.9 4.9 14.9 5.6 17.4 5.0 15.1 4.0 13.4 6.0 18.2 6.5 20.4 8.8 23.7 4.3 14.0 7.0 19.7 5.8 17.5 NICE 13.2 39.5 13.5 40.3 14.5 42.7 20.6 52.7 10.1 31.5 16.5 44.0 17.0 42.1 22.9 56.1 15.4 41.6 17.4 45.8 16.1 43.6 ATM-S 25.6 60.4 22.0 54.5 25.0 62.4 31.4 60.9 12.9 43.0 21.3 51.1 30.5 61.5 38.8 72.0 30.4 51.5 29.1 63.5 28.5 60.4 Cog-cap 31.4 79.7 31.4 77.8 38.2 85.7 40.4 85.8 24.4 66.3 34.8 78.8 34.7 81.0 48.1 88.6 37.4 79.4 35.6 79.3 35.6 80.2 UBP 41.2 70.5 51.2 80.9 51.2 82.0 51.1 76.9 42.2 72.8 57.5 83.5 49.0 79.9 58.6 85.8 45.1 76.2 61.5 88.2 50.9 79.7 HyFI 60.6 85.3 65.9 94.0 69.5 93.9 66.5 89.8 55.0 86.0 74.4 95.0 68.4 91.3 78.9 96.9 66.0 90.6 77.0 96.4 68.2 91.9 Inter-subject: leave one subject out for test BraVL 2.3 8.0 1.5 6.3 1.9 6.7 2.1 8.1 2.2 7.6 1.6 6.4 2.3 8.5 1.8 7.0 1.4 5.9 1.7 6.7 1.5 5.6 NICE 7.6 22.8 5.9 20.5 6.0 22.3 6.3 20.7 4.4 18.3 5.6 22.2 5.6 19.7 6.3 22.0 5.7 17.6 8.4 28.3 6.2 21.4 NICE-G 5.9 21.4 6.4 22.7 5.5 20.1 6.1 21.0 4.7 19.5 6.2 22.5 5.9 19.1 7.3 25.3 6.2 18.3 6.2 26.3 5.9 21.6 ATM-S 10.5 26.8 7.1 24.8 11.9 33.8 14.7 39.4 7.0 23.9 11.1 35.8 16.1 43.5 15.0 40.3 4.9 22.7 20.5 46.5 11.8 33.7 UBP 11.5 29.7 15.5 40.0 9.8 27.0 13.0 32.3 8.8 33.8 11.7 31.0 10.2 23.8 12.2 32.2 15.5 40.5 16.0 43.5 12.4 33.4 HyFI 16.2 35.8 20.0 47.7 7.5 26.7 18.8 41.3 9.7 27.5 15.5 33.8 11.0 34.5 13.2 30.3 13.3 38.8 25.3 55.7 15.1 37.2 Table 1: Top-1 (T-1) and top-5 (T-5) accuracy (%\%) results in 200-way zero-shot brain-to-image retrieval on THINGS-EEG, reported for intra- and inter-subject settings. Bold and underline indicate the best and second-best results, respectively. Method Subject 1 Subject 2 Subject 3 Subject 4 Average T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 T-1 T-5 Intra-subject: train and test on one subject NICE 8.7 30.5 21.8 56.6 16.5 49.7 10.3 32.3 14.3 42.3 UBP 15.0 38.0 46.0 80.5 27.3 59.0 18.5 43.5 26.7 55.2 HyFI 17.6 40.1 63.8 91.1 38.0 76.9 23.7 50.2 35.8 64.6 Inter-subject: leave one subject out for test UBP 2.0 5.7 1.5 17.2 2.7 10.5 2.5 8.0 2.2 10.4 HyFI 2.7 6.0 4.3 17.0 3.5 12.5 1.0 7.6 3.2 11.5 Table 2: Top-1 and top-5 accuracy (%\%) results in 200-way zero-shot brain-to-image retrieval on THINGS-MEG. Experiments and Results Datasets THINGS-EEG We used the THINGS-EEG dataset (Gifford et al. 2022), a large-scale EEG benchmark collected under the Rapid Serial Visual Presentation (RSVP) paradigm, which provides image-EEG paired data from 10 subjects viewing rapid sequences of images. The training set consists of 1,654 object concepts, each represented by 10 distinct images, with each image repeated 4 times per subject. The test set includes 200 concepts with one image per concept and each image repeated 80 times per subject. Following prior preprocessing protocols (Song et al. 2024; Wu et al. 2025), we averaged trials, downsampled EEG signals to 250 Hz, and selected 17 channels over occipital and parietal regions. THINGS-MEG We additionally utilized the THINGS-MEG dataset (Hebart et al. 2023), which contains recordings from four participants using 271 MEG channels. The training set includes 1,854 object concepts, each presented once with 12 different images, while the test set consists of 200 concepts, each associated with a single image repeated 12 times. To ensure fair comparison, we follow the same preprocessing pipeline as described in (Song et al. 2024; Wu et al. 2025); more details are provided in the appendix. Implementation Details We used AdamW (Loshchilov and Hutter 2017) with learning rate 3×10−43× 10^-4, weight decay 1×10−41× 10^-4, and batch size 1024. We trained the model for 50 epochs. The curvature parameter κ was set to 1 at initialization and optimized during training. To control feature norms, we initialize the scaling factors αv _v and αb _b to 1/d 1/d, following (Desai et al. 2023). We empirically observed that fixing αv=1 _v=1 led to better training stability for ResNet-based and hyperbolic vision encoder, due to the already exhibit low norm of their features. All experiments were conducted on a GPU, GTX 1080 Ti (12GB). We applied fovea blur and Gaussian blur. Augmentation details are in the appendix. Vision Encoders We evaluated a range of visual backbones, including RN50, RN101, ViT-B/16, ViT-B/32, ViT-L/14, and ViT-H/14. In addition, we also used two recent hyperbolic vision-language models, MERU (Desai et al. 2023) and HyCoCLIP (Pal et al. 2024), which are pretrained to capture hierarchical relationships between images and text. We adopted RN50 as the default vision encoder. Brain Encoders We applied our method using a variety of EEG encoders to demonstrate its generality across different architectures. Specifically, we adopted models that are either widely used in the literature or recently proposed in neural encoding studies, including ShallowNet (Schirrmeister et al. 2017), EEGNet (Lawhern et al. 2018), and TSConv (Li et al. 2024), as well as EEGProject (Wu et al. 2025). We used EEGProject as default EEG encoder. Results We compare our method with several recent neural decoding approaches for brain-to-image retrieval, including BraVL (Du et al. 2023), NICE (Song et al. 2024), ATM (Li et al. 2024), CogCap (Zhang et al. 2025), and UBP (Wu et al. 2025). Detailed baseline descriptions are in the appendix. To evaluate decoding performance, we perform 200-way zero-shot retrieval on THINGS-EEG and THINGS-MEG, following prior work (Du et al. 2023; Wu et al. 2025). This evaluates alignment quality and generalization to novel concepts. The results in Tables 1 and 2. We averaged over 5 runs, and all improvements are statistically significant (p<0.01p<0.01). On the THINGS-EEG dataset, our method achieves a top-1 accuracy of 68.2%\% and top-5 accuracy of 91.9%\%, outperforming the previous SOTA (UBP) by +17.3%\% and +12.2%\%, respectively. On the THINGS-MEG dataset, HyFI achieves a top-1 accuracy of 35.8%\% and top-5 accuracy of 64.6%\%, improving upon UBP by +9.1%\% and +9.4%\%, respectively. Furthermore, we conduct a qualitative comparison of retrieval results with UBP, as shown in Fig. 4. For fair comparison, we used subject 4 with near-average performance. As illustrated in the figure, our method preserves both semantic and perceptual coherence, whereas previous approaches fail to maintain this consistency. This indicates that our method effectively fuses perceptual and semantic features, thereby maintaining coherence in both aspects. More examples of retrieved image are provided in the appendix. Ablation Study Effect of Hyperbolic Geometry and Feature Interpolation To evaluate the contributions of hyperbolic space and feature interpolation, we conduct an ablation study and results in Table 3. First, we observe that aligning brain and image features is more effective in hyperbolic space than in CLIP space (Euclidean space). This supports the suitability of hyperbolic space for modeling the information imbalance between brain and visual modalities. Next, we compare interpolating features in the original CLIP space and in hyperbolic space. Even in the CLIP space, fusing low-level visual features improves performance, highlighting the benefit of integrating perceptual information. However, performing interpolation in hyperbolic space leads to the best results. This finding indicates that hyperbolic interpolation effectively reduces redundancy and representational complexity, while facilitating feature fusion. Interpolation Hyperbolic THINGS-EEG THINGS-MEG T-1 T-5 T-1 T-5 – – 49.4 81.0 23.1 47.8 – ✓ 54.3 82.5 28.8 55.1 ✓ – 59.7 86.8 25.2 51.8 ✓ ✓ 68.2 91.9 35.8 64.6 Table 3: Ablation study: effect of interpolation and hyperbolic space. The row with only interpolation checked refer to interpolation in CLIP space. Method Architecture Model THINGS-EEG THINGS-MEG w/o HyFI Ours w/o HyFI Ours CLIP CNN RN50 49.4 68.2 23.1 35.8 RN101 44.4 62.1 21.0 34.8 ViT ViT-B/16 36.0 41.6 18.6 23.3 ViT-B/32 42.5 46.9 20.0 25.1 ViT-L/14 30.1 34.1 13.0 21.0 ViT-H/14 42.5 43.8 18.7 27.9 MERU ViT ViT-S/16 43.7 52.4 21.6 31.1 ViT-B/16 31.7 48.8 15.7 24.6 ViT-L/16 23.1 30.9 11.0 19.6 HyCoCLIP ViT ViT-S/16 44.3 51.6 21.2 27.9 ViT-B/16 35.6 45.6 18.1 27.0 Table 4: Top-1 image retrieval accuracy (%\%) across visual encoders. The “w/o HyFI” represents alignment in CLIP space; MERU and HyCoCLIP use hyperbolic alignment. Figure 4: Qualitative comparison of image retrieval results. Our method retrieves semantically and perceptually coherent images, while the previous method often suffers from color or semantic inconsistencies. Brain Encoder THINGS-EEG THINGS-MEG w/o HyFI Ours w/o HyFI Ours ShallowNet 36.6 50.5 15.1 23.2 EEGNet 36.2 51.6 19.5 26.4 TSConv 41.1 57.1 21.9 30.5 EEGProject 49.4 68.2 23.1 35.8 Table 5: Top-1 image retrieval accuracy (%) using different brain encoders and CLIP-RN50. Effect of Vision and Brain Encoders We investigate the impact of different vision and brain encoders on brain-to-image retrieval performance, the results summarized in Table 4 and 5. For this analysis, we use EEGProject as the brain encoder and compare various pretrained vision encoders. We observe that CNN-based backbones yield stronger alignment with EEG signals compared to transformer-based models. Interestingly, lightweight architectures outperform deeper models. These results suggest that compact visual representations may be more compatible with neural signals. Importantly, our method consistently outperforms the baseline across all vision architectures. We also evaluate the effect of different brain encoders while fixing the vision encoder to CLIP-RN50. We observe that our method consistently improves performance across all brain encoder architectures. This highlights the general applicability and robustness of our approach, regardless of the specific choice of brain encoder. Analysis Feature Visualization To investigate the effect of feature interpolation in hyperbolic space, we visualize the distribution of distances from the root, as shown in Fig. 5. Following (Desai et al. 2023), we define the root as the mean of all embeddings in CLIP space and the time origin O in hyperbolic space. As shown in Fig. 5, (a) in the CLIP space, the interpolated image embeddings lie between the semantic and perceptual features, whereas (b) in the hyperbolic space, the interpolated embeddings are located closer to the time origin O. This behavior reflects the nature of hyperbolic interpolation, which bends toward the origin and results in a tighter bound on the spatial components, as described in Eq. (11). It suggests that effectively compressing and integrating semantic and perceptual features leads to better alignment with brain signals. Notably, the EEG embeddings lie farther from the origin, as their high variability requires regions with fewer constraints (i.e., away from the time origin). Figure 5: Distributions of embedding distances from the root in (a) CLIP space and (b) hyperbolic space. Interpolated image embeddings lie closer to the root in hyperbolic space, unlike in CLIP space. Analysis of interpolation coefficient To understand how the model adaptively integrates semantic and perceptual information, we analyze the learned interpolation coefficient t∈[0,1]t∈[0,1]. The distribution of coefficients is shown in Fig. 6. We observe that t is predominantly distributed below 0.5, indicating that the model tends to focus more on semantic features during interpolation. To further interpret this behavior, we also examine images corresponding to both low and high t values in the test set. Images with lower t values typically contain objects that are iconic examples of their higher-level categories (e.g., banana–fruit, cheetah–mammal, and van-car). In contrast, images with higher t values tend to exhibit salient low-level visual attributes such as orientation and color. Figure 6: Distribution of the interpolation coefficient t and example images with low and high t values. Analyzing the effect of image augmentation To investigate the effect of image augmentations on a pre-trained VLM, we conduct image-to-image retrieval using perceptual and semantic images as queries in CLIP-RN50. The results are shown in Fig. 7. When using the semantic images as queries, the retrieved results tend to be semantically consistent. In contrast, perceptual images bias the retrieval toward low-level visual attributes such as color and object orientation. Additional results with other VLMs are in the appendix. Figure 7: Comparison of Top-3 retrieval results using (a) semantic and (b) perceptual images. Semantic queries tend to retrieve conceptually similar images (e.g., plants, fruits, animals). In contrast, perceptual queries retrieve images with shared low-level visual feature such as color and orientation. Conclusion We propose Hyperbolic Feature Interpolation (HyFI), a framework that interpolates semantic and perceptual visual features in hyperbolic space for improved alignment with brain signals. Leveraging the geodesic curvature toward the origin, HyFI enables effective fusion and compression of visual representations. This leads to better alignment with neural activity by reflecting both limited brain information and feature entanglement. Experiments on THINGS-EEG and THINGS-MEG show that HyFI achieves SOTA performance on zero-shot brain-to-image retrieval task. Acknowledgments This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) under two projects: (No. RS-2019-I190079, Artificial Intelligence Graduate School Program at Korea University) and (No. RS-2024-00457882, National AI Research Lab Project). References M. G. Atigh, J. Schoep, E. Acar, N. Van Noord, and P. Mettes (2022) Hyperbolic image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 4453–4462. Cited by: Hyperbolic Representation Learning. J. W. Cannon, W. J. Floyd, R. Kenyon, W. R. Parry, et al. (1997) Hyperbolic geometry. Flavors of geometry 31 (59-115), p. 2. Cited by: Definition. P. Cavanagh and G. A. Alvarez (2005) Tracking multiple targets with multifocal attention. Trends in cognitive sciences 9 (7), p. 349–354. Cited by: Introduction. B. P. Chamberlain, J. Clough, and M. P. Deisenroth (2017) Neural embeddings of graphs in hyperbolic space. arXiv preprint arXiv:1705.10359. Cited by: Hyperbolic Representation Learning. I. Chami, A. Gu, D. P. Nguyen, and C. Ré (2021) Horopca: hyperbolic dimensionality reduction via horospherical projections. In International Conference on Machine Learning, p. 1419–1429. Cited by: Feature Visualization. H. Chen, L. He, Y. Liu, and L. Yang (2024) Visual neural decoding via improved visual-eeg semantic consistency. arXiv preprint arXiv:2408.06788. Cited by: Introduction. A. Défossez, C. Caucheteux, J. Rapin, O. Kabeli, and J. King (2023) Decoding speech perception from non-invasive brain recordings. Nature Machine Intelligence 5 (10), p. 1097–1107. Cited by: Introduction. K. Desai, M. Nickel, T. Rajpurohit, J. Johnson, and S. R. Vedantam (2023) Hyperbolic image-text representations. In International Conference on Machine Learning, p. 7694–7731. Cited by: Hyperbolic Representation Learning, Preliminaries, Hyperbolic Brain-Vision Contrastive Learning, Vision Encoders, Implementation Details, Feature Visualization. B. Dhingra, C. J. Shallue, M. Norouzi, A. M. Dai, and G. E. Dahl (2018) Embedding text in hyperbolic spaces. arXiv preprint arXiv:1806.04313. Cited by: Hyperbolic Representation Learning. J. J. DiCarlo, D. Zoccolan, and N. C. Rust (2012) How does the brain solve visual object recognition?. Neuron 73 (3), p. 415–434. Cited by: Introduction. C. Du, K. Fu, J. Li, and H. He (2023) Decoding visual neural representations by multimodal learning of brain-visual-linguistic features. IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (9), p. 10760–10777. Cited by: Results, Results, 1st item. P. E. Dux and R. Marois (2009) The attentional blink: a review of data and theory. Attention, Perception, & Psychophysics 71 (8), p. 1683–1700. Cited by: Introduction. O. Ganea, G. Becigneul, and T. Hofmann (2018) Hyperbolic entailment cones for learning hierarchical embeddings. In International Conference on Machine Learning, p. 1646–1655. Cited by: Compression Effects. A. T. Gifford, K. Dwivedi, G. Roig, and R. M. Cichy (2022) A large and rich eeg dataset for modeling human visual object recognition. NeuroImage 264, p. 119754. Cited by: THINGS-EEG, THINGS-EEG. M. N. Hebart, O. Contier, L. Teichmann, A. H. Rockter, C. Y. Zheng, A. Kidder, A. Corriveau, M. Vaziri-Pashkam, and C. I. Baker (2023) THINGS-data, a multimodal collection of large-scale datasets for investigating object representations in human brain and behavior. Elife 12, p. e82580. Cited by: THINGS-MEG, THINGS-EEG. K. N. Kay, T. Naselaris, R. J. Prenger, and J. L. Gallant (2008) Identifying natural images from human brain activity. Nature 452 (7185), p. 352–355. Cited by: Visual Brain Decoding. V. Khrulkov, L. Mirvakhabova, E. Ustinova, I. Oseledets, and V. Lempitsky (2020) Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 6418–6428. Cited by: Compression Effects. W. Ko, E. Jeon, S. Jeong, and H. Suk (2021) Multi-scale neural network for eeg representation learning in bci. IEEE Computational Intelligence Magazine 16 (2), p. 31–45. Cited by: Introduction. V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance (2018) EEGNet: a compact convolutional neural network for eeg-based brain–computer interfaces. Journal of neural engineering 15 (5), p. 056013. Cited by: Brain Encoders. M. Le, S. Roller, L. Papaxanthos, D. Kiela, and M. Nickel (2019) Inferring concept hierarchies from text corpora via hyperbolic embeddings. arXiv preprint arXiv:1902.00913. Cited by: Preliminaries. D. Li, C. Wei, S. Li, J. Zou, H. Qin, and Q. Liu (2024) Visual decoding and reconstruction via eeg embeddings with guided diffusion. arXiv preprint arXiv:2403.07721. Cited by: Visual Brain Decoding, Brain Encoders, Results, 3rd item. H. Li, H. Wu, and B. Chen (2025) NeuralDiffuser: neuroscience-inspired diffusion guidance for fmri visual reconstruction. IEEE Transactions on Image Processing. Cited by: Introduction. V. W. Liang, Y. Zhang, Y. Kwon, S. Yeung, and J. Y. Zou (2022) Mind the gap: understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems 35, p. 17612–17625. Cited by: Introduction. Q. Liu, M. Nickel, and D. Kiela (2019) Hyperbolic graph neural networks. Advances in neural information processing systems 32. Cited by: Hyperbolic Representation Learning. I. Loshchilov and F. Hutter (2017) Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101. Cited by: Implementation Details. P. Mandica, L. Franco, K. Kallidromitis, S. Petryk, and F. Galasso (2025) Hyperbolic learning with multimodal large language models. In European Conference on Computer Vision, p. 382–398. Cited by: Hyperbolic Brain-Vision Contrastive Learning. M. W. Mathis, A. P. Rotondo, E. F. Chang, A. S. Tolias, and A. Mathis (2024) Decoding the brain: from neural representations to mechanistic models. Cell 187 (21), p. 5814–5832. Cited by: Introduction. Y. Miyawaki, H. Uchida, O. Yamashita, M. Sato, Y. Morito, H. C. Tanabe, N. Sadato, and Y. Kamitani (2008) Visual image reconstruction from human brain activity using a combination of multiscale local image decoders. Neuron 60 (5), p. 915–929. Cited by: Introduction. T. Naselaris, K. N. Kay, S. Nishimoto, and J. L. Gallant (2011) Encoding and decoding in fmri. Neuroimage 56 (2), p. 400–410. Cited by: Introduction, Introduction. T. Naselaris, R. J. Prenger, K. N. Kay, M. Oliver, and J. L. Gallant (2009) Bayesian reconstruction of natural images from human brain activity. Neuron 63 (6), p. 902–915. Cited by: Introduction. M. Nickel and D. Kiela (2017) Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems 30. Cited by: Hyperbolic Representation Learning. S. R. Oota, Z. Chen, M. Gupta, R. S. Bapi, G. Jobard, F. Alexandre, and X. Hinaut (2023) Deep neural networks and brain alignment: brain encoding and decoding (survey). arXiv preprint arXiv:2307.10246. Cited by: Introduction. A. Pal, M. van Spengler, G. M. D. di Melendugno, A. Flaborea, F. Galasso, and P. Mettes (2024) Compositional entailment learning for hyperbolic vision-language models. arXiv preprint arXiv:2410.06912. Cited by: Hyperbolic Representation Learning, Preliminaries, Hyperbolic Brain-Vision Contrastive Learning, Vision Encoders. W. Peng, T. Varanka, A. Mostafa, H. Shi, and G. Zhao (2021) Hyperbolic deep neural networks: a survey. IEEE Transactions on pattern analysis and machine intelligence 44 (12), p. 10023–10044. Cited by: Preliminaries. D. A. Pollen (1999) On the neural correlates of visual perception. Cerebral cortex 9 (1), p. 4–19. Cited by: Introduction. R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter, K. Eggensperger, M. Tangermann, F. Hutter, W. Burgard, and T. Ball (2017) Deep learning with convolutional neural networks for eeg decoding and visualization. Human brain mapping 38 (11), p. 5391–5420. Cited by: Brain Encoders. S. Schrodi, D. T. Hoffmann, M. Argus, V. Fischer, and T. Brox (2024) Two effects, one trigger: on the modality gap, object bias, and information imbalance in contrastive vision-language models. arXiv preprint arXiv:2404.07983. Cited by: Introduction. P. Scotti, A. Banerjee, J. Goode, S. Shabalin, A. Nguyen, A. Dempster, N. Verlinde, E. Yundler, D. Weisberg, K. Norman, et al. (2023) Reconstructing the mind’s eye: fmri-to-image with contrastive learning and diffusion priors. Advances in Neural Information Processing Systems 36, p. 24705–24728. Cited by: Introduction, Visual Brain Decoding. P. S. Scotti, M. Tripathy, C. K. T. Villanueva, R. Kneeland, T. Chen, A. Narang, C. Santhirasegaran, J. Xu, T. Naselaris, K. A. Norman, et al. (2024) MindEye2: shared-subject models enable fmri-to-image with 1 hour of data. In Proceedings of the 41st International Conference on Machine Learning, p. 44038–44059. Cited by: Introduction. G. Shen, D. Zhao, X. He, L. Feng, Y. Dong, J. Wang, Q. Zhang, and Y. Zeng (2025) Neuro-vision to language: enhancing brain recording-based visual reconstruction and language interaction. Advances in Neural Information Processing Systems 37, p. 98083–98110. Cited by: Introduction. Y. Song, B. Liu, X. Li, N. Shi, Y. Wang, and X. Gao (2024) Decoding natural images from eeg for object recognition. In The Twelfth International Conference on Learning Representations, Cited by: Visual Brain Decoding, THINGS-EEG, THINGS-MEG, Results, 2nd item, THINGS-EEG, THINGS-EEG. R. Srinivasan, W. R. Winter, J. Ding, and P. L. Nunez (2007) EEG and meg coherence: measures of functional connectivity at distinct spatial scales of neocortical dynamics. Journal of neuroscience methods 166 (1), p. 41–52. Cited by: Introduction. A. Tifrea, G. Becigneul, and O. Ganea (2019) Poincare glove: hyperbolic word embeddings. In International Conference on Learning Representations, Cited by: Hyperbolic Representation Learning. Z. Wang and H. Ji (2022) Open vocabulary electroencephalography-to-text decoding and zero-shot sentiment classification. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 36, p. 5350–5358. Cited by: Introduction. H. Wu, Q. Li, C. Zhang, Z. He, and X. Ying (2025) Bridging the vision-brain gap with an uncertainty-aware blur prior. arXiv preprint arXiv:2503.04207. Cited by: Introduction, Visual Brain Decoding, Extracting Semantic and Perceptual Features, THINGS-EEG, THINGS-MEG, Brain Encoders, Results, Results, 5th item, Fovea Blur, THINGS-EEG, THINGS-EEG, Ablation Study of Augmentation. H. Yang, J. Gee, and J. Shi (2024) Brain decodes deep nets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 23030–23040. Cited by: Visual Brain Decoding. K. Zhang, L. He, X. Jiang, W. Lu, D. Wang, and X. Gao (2025) Cognitioncapturer: decoding visual stimuli from human eeg signal with multimodal information. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, p. 14486–14493. Cited by: Visual Brain Decoding, Results, 4th item. Appendix of HyFI: Hyperbolic Feature Interpolation for Brain-Vision Alignment Image Augmentation In this section, we describe Gaussian blur and fovea blur to generate perceptual and semantic images, respectively. We also present our parameter search strategy for identifying the optimal augmentation configuration. Gaussian Blur Gaussian blur removes high-frequency details such as textures and sharp edges, while preserving coarse visual structures like color and shape. Formally, the Gaussian-blurred image xblurx_blur is computed as: xblur(i,j)=∑m=−k∑n=−kkx(i−m,j−n)⋅G(m,n),x_blur(i,j)= _m=-k^k _n=-k^kx(i-m,j-n)· G(m,n), (1) where x(i,j)x(i,j) is the pixel value of input image, r=2k+1r=2k+1 represents the radius of the Gaussian kernel. The Gaussian kernel defined as: G(m,n)=12πσ2exp(−m2+n22σ2),G(m,n)= 12πσ^2 (- m^2+n^22σ^2 ), (2) with σ controlling the strength of the blur. Fovea Blur To simulate human vision, we adopt the fovea blur method described in prior work (Wu et al. 2025). x~v=δ⋅x+(1−δ)⋅xblur, x_v=δ· x+(1-δ)· x_blur, (3) where, δ is the blending factor. To simulate the fovea effect, the blending weight δ(i,j)δ(i,j) is defined as a function of the distance from the center: δ(i,j)=exp(−λ⋅d(i,j)L),δ(i,j)= (- λ· d(i,j)L ), (4) where d(i,j)d(i,j) denotes the Euclidean distance between pixel (i,j)(i,j) and the fovea, L is the maximum possible distance in the image, and λ is a hyperparameter that controls the rate of decay. To investigate the effect of augmentation, we systematically varied the augmentation parameters. Let rpr_p and rsr_s denote the Gaussian kernel sizes for perceptual and semantic image generation, respectively. The best Top-1 accuracy on image retrieval task was achieved when rp=31r_p=31, rs=51r_s=51, and λ=3λ=3, as show in Fig. 8. Figure 8: Top-1 accuracy heatmap from a parameter search over perceptual and semantic visual augmentation configurations. Geodesic Interpolation in Lorentz Model Definition of the Exponential Map Following the geodesic definition introduced in the main text, a geodesic γ(t)γ(t) starting at ∈np ^n with an initial tangent vector 0∈Tnv_0∈ T_pL^n can be parameterized as: γ(t)=cosh(tκ‖0‖)+sinh(tκ‖0‖)κ‖0‖0.γ(t)= (t κ\|v_0\|_L )p+ (t κ\|v_0\|_L ) κ\|v_0\|_Lv_0. (5) For notational convenience, let us define β0=κ‖0‖ _0= κ\|v_0\|_L. Then, for t≥0t≥ 0, κ‖t0‖=tβ0 κ\|tv_0\|_L=t _0. Substituting tβ0t _0 into the exponential map: γ(t) γ(t) =cosh(tβ0)+sinh(tβ0)β00. = (t _0)p+ (t _0) _0v_0. (6) This formula describes the geodesic extending from p in the direction of 0v_0. Reformulation Geodesic To find the geodesic connecting two points p and q in nL^n, we use the logarithm map logκ(⋅) ^κ_p(·). We define the point q as the exponential map of a tangent vector 0v_0 at p (i.e., =expκ(0)q= ^κ_p(v_0)). From Eq.(6) with t=1t=1, we have =cosh(β0)+sinh(β0)β00q= ( _0)p+ ( _0) _0v_0. Taking the Lorentzian inner product of both sides with p: ⟨,⟩ ,q _L =⟨,cosh(β0)+sinh(β0)β00⟩ = , ( _0)p+ ( _0) _0v_0 _L (7) =cosh(β0)⟨,⟩+sinh(β0)β0⟨,0⟩ = ( _0) ,p _L+ ( _0) _0 ,v_0 _L (8) By the Lorentzian constraint, ⟨,⟩=−1/κ ,p _L=-1/κ and ⟨,0⟩=0 ,v_0 _L=0 (since 0∈Tnv_0∈ T_pL^n). Thus: ⟨,⟩=−1κcosh(β0) ,q _L=- 1κ ( _0) (9) From the geodesic distance definition in the main text, d(,)=1/κ⋅cosh−1(−κ⟨,⟩)d_L(p,q)= 1/κ· ^-1 (-κ ,q _L ). Let β=κ⋅d(,)β= κ· d_L(p,q). Then, cosh(β)=−κ⋅⟨,⟩ (β)=-κ· ,q _L. Comparing the two expressions for cosh(β0) ( _0) and cosh(α) (α), we find β0=β _0=β. Thus, the parameter β0 _0 associated with the tangent vector 0=log()v_0= _p(q) is precisely β=κ⋅d(,)β= κ· d_L(p,q). Now, we solve for 0v_0 from =cosh(β)+sinh(β)β0q= (β)p+ (β)βv_0: −cosh(β) - (β)p =sinh(β)β0 = (β)βv_0 (10) 0 _0 =βsinh(β)(−cosh(β)) = β (β)(q- (β)p) (11) This vector 0v_0 is exactly what is given by the logarithmic map logκ() ^κ_p(q) in the main text. Finally, substituting this 0v_0 back into the geodesic definition Eq.(6): γ(t) γ(t) =cosh(tβ)+sinh(tβ)β0 = (tβ)p+ (tβ)βv_0 (12) =cosh(tβ)+sinh(tβ)β(βsinh(β)(−cosh(β))) = (tβ)p+ (tβ)β ( β (β)(q- (β)p) ) (13) =cosh(tβ)+sinh(tβ)sinh(β)−sinh(tβ)cosh(β)sinh(β) = (tβ)p+ (tβ) (β)q- (tβ) (β) (β)p (14) =(cosh(tβ)−sinh(tβ)cosh(β)sinh(β))+sinh(tβ)sinh(β) = ( (tβ)- (tβ) (β) (β) )p+ (tβ) (β)q (15) Using the hyperbolic trigonometric identity sinh(A−B)=sinh(A)cosh(B)−cosh(A)sinh(B) (A-B)= (A) (B)- (A) (B), the coefficient in front of p in Eq. (15) can be rewritten as: cosh(tβ)−sinh(tβ)cosh(β)sinh(β) (tβ)- (tβ) (β) (β) (16) =cosh(tβ)sinh(β)−sinh(tβ)cosh(β)sinh(β) = (tβ) (β)- (tβ) (β) (β) (17) =−sinh(tβ−β)sinh(β)=sinh(β−tβ)sinh(β)=sinh((1−t)β)sinh(β) = - (tβ-β) (β)= (β-tβ) (β)= ((1-t)β) (β) (18) Thus, the geodesic from p to q is expressed in terms of hyperbolic sine functions: γ(t)=sinh((1−t)β)sinh(β)+sinh(tβ)sinh(β),t∈[0,1]. γ(t)= ((1-t)β) (β)p+ (tβ) (β)q, t∈[0,1]. (19) Convexity of Geodesics in the Hyperbolic Space Problem Setup Let ,∈np,q ^n be two points on the hyperboloid model separated by a hyperbolic distance D>0D>0, and define β=κ⋅Dβ= κ· D. Let ~ p and ~ q be the spatial components of p and q, respectively, belonging to ℝnR^n. The spatial part of the geodesic interpolation is: γspatial(t) _spatial(t) =a~+b~, =a p+b q, where a=sinh((1−t)β)sinh(β) a= ((1-t)β) (β) and b=sinh(tβ)sinh(β). and b= (tβ) (β). (20) In contrast, the straight-line (Euclidean) interpolation of the spatial coordinates is: Mspatial(t)=(1−t)~+t~. M_spatial(t)=(1-t) p+t q. (21) Our goal is to prove that for every t∈(0,1)t∈(0,1), the hyperbolic interpolation lies strictly closer to the origin in ℝnR^n (with the standard Euclidean norm ∥⋅∥\|·\|) than the Euclidean interpolation, unless ~ p and ~ q are collinear with the origin. That is, we aim to show: ‖γspatial(t)‖<‖Mspatial(t)‖. \| _spatial(t)\|<\|M_spatial(t)\|. (22) Main Proof The proof proceeds in three steps. First, we establish key properties of the coefficients a and b relative to their Euclidean counterparts. Second, we establish their sum property. Finally, we use these properties to compare the norms. Key Inequalities for Coefficients a and b Lemma 1 (Coefficient Inequality): For β>0β>0 and t∈(0,1)t∈(0,1), the coefficients a and b satisfy: a=sinh((1−t)β)sinh(β) a= ((1-t)β) (β) <(1−t) <(1-t) (23) b=sinh(tβ)sinh(β) b= (tβ) (β) <t <t (24) Proof Consider the function f(x)=sinh(x)xf(x)= (x)x for x>0x>0. Its derivative is f′(x)=xcosh(x)−sinh(x)x2f (x)= x (x)- (x)x^2. Let g(x)=xcosh(x)−sinh(x)g(x)=x (x)- (x). Its derivative is g′(x)=cosh(x)+xsinh(x)−cosh(x)=xsinh(x)g (x)= (x)+x (x)- (x)=x (x). For x>0x>0, g′(x)>0g (x)>0, so g(x)g(x) is strictly increasing. Since g(0)=0g(0)=0, it follows that g(x)>0g(x)>0 for all x>0x>0. Therefore, f′(x)>0f (x)>0 for x>0x>0, implying that f(x)=sinh(x)xf(x)= (x)x is a strictly increasing function for x>0x>0. Now, we apply this property to a and b: For a: Since t∈(0,1)t∈(0,1) and β>0β>0, we have 0<(1−t)β<β0<(1-t)β<β. Because f(x)f(x) is strictly increasing: f((1−t)β) f((1-t)β) <f(β) <f(β) (25) sinh((1−t)β)(1−t)β ((1-t)β)(1-t)β <sinh(β)β < (β)β (26) Multiplying both sides by (1−t)(1-t) (which is positive): sinh((1−t)β)sinh(β)<(1−t) ((1-t)β) (β)<(1-t) (27) So, a<(1−t)a<(1-t). Similarly, for b: Since 0<tβ<β0<tβ<β: f(tβ) f(tβ) <f(β) <f(β) (28) sinh(tβ)tβ (tβ)tβ <sinh(β)β < (β)β (29) Multiplying both sides by t (which is positive): sinh(tβ)sinh(β)<t (tβ) (β)<t (30) So, b<tb<t. ■ Convexity and Sum of Coefficients Let us analyze the sum of the coefficients, S(t)=a+bS(t)=a+b. S(t)=sinh((1−t)β)+sinh(tβ)sinh(β). S(t)= ((1-t)β)+ (tβ) (β). (31) We will show that S(t)S(t) is a strictly convex function for t∈(0,1)t∈(0,1). The second derivative of S(t)S(t) with respect to t is: S′(t) S (t) =−βcosh((1−t)β)+βcosh(tβ)sinh(β), = -β ((1-t)β)+β (tβ) (β), (32) S′(t) S (t) =β2sinh((1−t)β)+β2sinh(tβ)sinh(β). = β^2 ((1-t)β)+β^2 (tβ) (β). (33) Since β>0β>0 and t∈(0,1)t∈(0,1), both (1−t)β(1-t)β and tβtβ are positive. Therefore, sinh((1−t)β)>0 ((1-t)β)>0 and sinh(tβ)>0 (tβ)>0, which implies S′(t)>0S (t)>0. Because S(t)S(t) is strictly convex, for any t∈(0,1)t∈(0,1), its value is strictly less than the value on the line segment connecting its endpoints S(0)S(0) and S(1)S(1). We evaluate the function at the endpoints: S(0) S(0) =sinh(β)+sinh(0)sinh(β)=1, = (β)+ (0) (β)=1, (34) S(1) S(1) =sinh(0)+sinh(β)sinh(β)=1. = (0)+ (β) (β)=1. (35) By strict convexity, for all t∈(0,1)t∈(0,1), we have S(t)<max(S(0),S(1))=1S(t)< (S(0),S(1))=1. This means: a+b<1. a+b<1. (36) This result also follows directly from Lemma 1, since a<(1−t)a<(1-t) and b<tb<t implies a+b<(1−t)+t=1a+b<(1-t)+t=1. However, the convexity argument provides an elegant alternative. Comparison of Norms We want to compare ‖γspatial(t)‖\| _spatial(t)\| and ‖Mspatial(t)‖\|M_spatial(t)\|. Let =~P= p and =~Q= q for simplicity. Let’s consider the Euclidean triangle T with vertices at the origin 0, P, and Q. • The geodesic interpolation γspatial(t)=a+b _spatial(t)=aP+bQ. From Lemma 1, a>0a>0 and b>0b>0. From the sum of coefficients property, a+b<1a+b<1. This means γspatial(t) _spatial(t) is a strict convex combination of 0, P, and Q (since the coefficient for 0 is 1−(a+b)1-(a+b), which is strictly positive). Therefore, γspatial(t) _spatial(t) lies strictly in the open interior of the triangle T. • The Euclidean interpolation Mspatial(t)=(1−t)+tM_spatial(t)=(1-t)P+tQ. For t∈(0,1)t∈(0,1), (1−t)>0(1-t)>0, t>0t>0, and their sum is (1−t)+t=1(1-t)+t=1. Therefore, Mspatial(t)M_spatial(t) lies strictly on the line segment connecting P and Q. This segment is the side of the triangle T that is opposite to the origin 0. Now, consider the case where P and Q are non-collinear with the origin. In this scenario, the triangle T is non-degenerate. A fundamental geometric property of Euclidean space states that for a non-degenerate triangle with one vertex at the origin, any point in the open interior of the triangle is strictly closer to the origin than any point on the opposite side (the side that does not include the origin). This holds true unless the opposite side itself passes through the origin. Since we are in the non-collinear case, the segment connecting P and Q does not pass through the origin. This geometric property thus proves our initial claim from Eq. (22): ‖γspatial(t)‖<‖Mspatial(t)‖|| _spatial(t)||<||M_spatial(t)||. This inequality is strict for all t∈(0,1)t∈(0,1), unless ~ p and ~ q are collinear with the origin. Degenerate Collinear Case If ~ p and ~ q are collinear with the origin (i.e., ~=c~ q=c p for some scalar c), the ”triangle” formed by 0, ~ p, and ~ q degenerates into a line segment passing through the origin. In this radial case, both γspatial(t) _spatial(t) and Mspatial(t)M_spatial(t) lie on this line segment. The spatial components become: γspatial(t) _spatial(t) =(a+bc)~ =(a+bc) p (37) Mspatial(t) M_spatial(t) =((1−t)+tc)~ =((1-t)+tc) p (38) The proof still holds true: from the coefficient inequalities (a<(1−t)a<(1-t) and b<tb<t), it follows that ‖a+bc‖<‖(1−t)+tc‖\|a+bc\|<\|(1-t)+tc\| because the coefficients a and b are strictly smaller than their Euclidean counterparts (1−t)(1-t) and t. This ensures that γspatial(t) _spatial(t) remains strictly closer to the origin than Mspatial(t)M_spatial(t) for t∈(0,1)t∈(0,1), unless t=0t=0 or t=1t=1. The only edge case where the norms might be equal is if the points ~ p and ~ q are the zero vector, which corresponds to the hyperbolic points being at the origin of their respective tangent spaces, but this is a specific degenerate scenario generally excluded by the problem’s context of separated points. Relationship Between Spatial Norm and Time Component As defined in Eq. (1) in the main text, any point =(0,~)∈np=(p_0, p) ^n must satisfy the constraint ⟨,⟩=−1/κ ,p _L=-1/κ. Using the definition of the Lorentzian inner product from Eq. (2) in the main text, this constraint can be expanded for a single point p: −02+⟨~,~⟩E=−1/κ. -p_0^2+ p, p _E=-1/κ. (39) Here, ⟨~,~⟩E p, p _E is the standard Euclidean dot product, which is equivalent to the squared Euclidean norm ‖~‖2|| p||^2 of the spatial component ~∈ℝn p ^n. By rearranging the equation to solve for the time component 0p_0 (given 0>0p_0>0), we obtain a direct relationship: 0=‖~‖2+1/κ.p_0= || p||^2+1/κ. (40) This equation mathematically demonstrates that the time component 0p_0 is a monotonically increasing function of the spatial component’s norm ‖~‖|| p||. Therefore, a point being “closer to the spatial origin” (i.e., having a smaller ‖~‖|| p||) is equivalent to its time component 0p_0 being smaller. This brings it closer to the Lorentz manifold’s origin point =(1κ,0,…,0)T∈nO=( 1 κ,0,…,0)^T ^n, which is the point with the minimum possible time component. The conclusion of our proof that ‖γspatial(t)‖<‖Mspatial(t)‖|| _spatial(t)||<||M_spatial(t)|| thus directly implies that the hyperbolic geodesic interpolation lies closer to the manifold’s origin than its Euclidean counterpart. Dataset To study visual brain decoding, we utilized paired datasets comprising brain signals and corresponding visual stimuli. Specifically, we used the widely adopted THINGS-EEG and THINGS-MEG datasets. Both datasets were collected under the Rapid Serial Visual Presentation (RSVP) paradigm, which presents images in rapid succession while recording time-resolved neural responses. THINGS-EEG The THINGS-EEG dataset (Gifford et al. 2022) comprises EEG–image pairs collected from 10 subjects. EEG signals were recorded using a 64-channel system at a sampling rate of 1000 Hz and filtered to the range [0.1, 100] Hz. The training set consists of 1,654 object concepts, each associated with 10 distinct images, and each image repeated 4 times per subject (total 1,654 × 10 × 4 samples). The test set includes 200 concepts, each represented by a single image repeated 80 times per subject (total 200 × 80 samples). To ensure fair comparison with prior work, we follow the preprocessing protocol established in (Song et al. 2024; Wu et al. 2025). The raw EEG signals are downsampled to 250 Hz, and 17 channels located over occipital and parietal regions—areas associated with visual processing—are selected. All repetitions for each image are averaged to enhance the signal-to-noise ratio (SNR), resulting in 16,540 training samples and 200 test samples per subject. THINGS-EEG The THINGS-MEG dataset (Hebart et al. 2023) contains MEG–image pairs acquired from four participants using a 271-channel. During each trial, a visual stimulus was presented for 500 ms, followed by an inter-stimulus interval consisting of a blank screen lasting 1000 ± 200 ms. The training set includes 1,854 object concepts, each associated with 12 unique images presented once per subject. The test set comprises 200 novel concepts, each represented by a single image repeated 12 times. We follow the pre-proesccing pipeline described in (Song et al. 2024; Wu et al. 2025). MEG signals are bandpass filtered between 0.1 and 100 Hz and subsequently downsampled to 200 Hz. To enhance signal quality, all repetitions are averaged. Furthermore, we exclude the 200 concepts from the training set, consistent with prior experimental protocols. Model Params Emb dim Semantic Perceptual Norm Std Norm Std RN50 38.32 M 1024 233 0.05 277 0.06 RN101 56.26 M 512 282 0.09 325 0.11 ViT/B-16 86.19 M 512 1409 0.48 1404 0.48 ViT/B-32 87.85 M 512 1447 0.49 1473 0.51 ViT/L-14 303.97 M 768 2399 0.67 2535 0.71 ViT/H-14 632.08 M 1024 2835 0.68 2587 0.62 MERU (ViT-S) 21.66 M 512 160 0.05 162 0.05 MERU (ViT-B) 89.79 M 512 320 0.11 319 0.11 MERU (ViT-L) 303.30M 512 414 0.14 411 0.14 HyCoClip (ViT-S) 21.66 M 512 80 0.03 79 0.03 HyCoClip (ViT-B) 89.79 M 512 118 0.04 118 0.04 Table 6: Comparison of vision encoders and feature statistics. Baseline Descriptions • BraVL (Du et al. 2023) introduces a Mixture-of-Experts (MoE) framework that jointly models brain, visual, and linguistic modalities for neural decoding. The model leverages multiple expert pathways to capture complementary information across modalities. • NICE (Song et al. 2024) adopts a self-supervised contrastive learning framework, incorporating dual spatial attention modules to enhance EEG feature representations. • ATM (Li et al. 2024) proposes a dedicated EEG encoder named Adaptive Thinking Mapper, which integrates positional encoding with temporal-spatial EEG features to better model brain dynamics. • Cog-cap (Zhang et al. 2025) presents a multi-modal architecture that captures cross-modal information—such as text, depth, and image representations—by processing EEG signals through multiple expert encoders. • UBP (Wu et al. 2025) is a recent state-of-the-art approach that introduces uncertainty-aware fovea blur to enhance the alignment between brain signals and visual features. Semantic Perceptual Top-1 ACC Top-5 ACC Original Fovea blur 53.8 83.7 SAM 56.6 85.4 Gaussian noise 61.4 89.8 Low resolution 63.5 89.6 Gaussian blur 65.3 91.4 Fovea blur SAM 60.5 88.5 Gaussian noise 66.3 92.7 Low resolution 65.6 91.2 Gaussian blur 68.2 91.9 Table 7: Image retrieval accuracy (%\%) with various semantic and perceptual augmentations. Figure 9: Examples of different augmentations. SAM refer to SAM-based augmentation that segments the image and fills each region with its mean color. SAM refers to an augmentation that uses the SAM to segment the image and fills each region with its mean color. Ablation Study of Augmentation To obtain effective perceptual and semantic features, we applied various augmentations to a diverse set of images to obtain more effective perceptual and semantic visual features. Specifically, we consider fovea blur, Gaussian blur, Gaussian noise, low-resolution, and a segment anything model(SAM)-based augmentation that segments the image and fills each region with its mean color. Examples of these augmentations are shown in Fig. 9. We find that applying fovea blur to semantic images yields better performance, which aligns with prior findings (Wu et al. 2025) suggesting that simulating human visual attention through augmentation can be beneficial. Among the perceptual augmentations, Gaussian blur consistently outperforms other augmentations. Based on these observations, we use fovea blur and Gaussian blur for semantic and perceptual images, respectively. Interpolation Coefficient To adaptively integrate perceptual and semantic features, we introduce an interpolation coefficient t. In our main text, t is computed from the semantic feature vsx_v^s as follows: t=σ(Wtfv(vs)),t=σ(W_tf_v(x_v^s)), (41) where σ denotes the sigmoid function, Wt∈ℝ1×dW_t ^1× d is a learnable weight vector, d is the feature dimension, and fvf_v is the pre-trained visual encoder. Table 8 reports the retrieval performance obtained by varying the input feature used to compute the interpolation coefficient t. Using EEG features resulted in the lowest performance, while different image-based features showed comparable results with no substantial differences. Input feature Top-1 ACC Top-5 ACC EEG 54.5 84.8 Perceptual 67.1 91.8 Original 67.6 91.8 Semantic 68.2 91.9 Table 8: Retrieval accuracy with different strategies for obtaining the interpolation coefficient t. Feature Visualization To further explore the structure of the learned features, we visualize them on the Poincaré ball using HoronPCA (Chami et al. 2021), as shown in Fig. 10. The interpolated visual features lie between the semantic and perceptual visual features and tend to cluster near the origin. Notably, while perceptual and semantic features appear close in the Poincaré ball, their true distance exceeds that to the interpolated feature. This reflects the negative curvature of hyperbolic space, where geodesics through the origin are shorter than lateral paths at constant radius. In contrast, EEG features are embedded far from the origin, naturally reflecting their high variability in the hyperbolic space. Method Model ShallowNet EEGNet TSConv EEGProject Base Ours Base Ours Base Ours Base Ours CLIP RN50 36.6 50.5 36.2 51.6 41.1 57.1 49.4 68.2 RN101 34.0 45.8 35.6 46.4 35.9 53.4 44.4 62.1 ViT-B/16 28.3 36.7 30.7 35.1 31.4 40.5 36.0 41.6 ViT-B/32 30.7 36.8 33.7 37.0 36.6 40.8 42.5 46.9 ViT-L/14 19.8 22.8 19.1 27.0 20.9 27.4 30.1 34.1 ViT-H/14 19.5 28.3 28.4 33.1 30.9 35.4 42.5 43.8 MERU ViT-S/16 31.8 44.9 35.4 45.3 38.9 40.0 43.7 52.4 ViT-B/16 24.0 33.6 28.1 39.1 31.5 42.2 31.7 48.6 ViT-L/16 16.7 28.8 21.0 27.8 23.8 33.1 23.1 30.9 HyCoCLIP ViT-S/16 35.1 47.9 34.4 40.6 41.3 47.5 44.3 51.6 ViT-B/16 27.0 37.4 29.1 34.9 32.9 40.6 35.6 45.6 Table 9: Top-1 accuracy (%) for various combinations of brain encoders and vision models on brain-to-image retrieval in the THING-EEG dataset. Base indicates performance using either contrastive alignment (CLIP) or hyperbolic contrastive alignment (MERU and HyCoCLIP), without our proposed interpolation method. Image Retrieval Results We present the top-5 retrieval results on the THINGS-EEG dataset for both successful and failure cases, categorized into three semantic groups—animals, food, and artifacts—in Fig. 11 and Fig. 12. In successful cases, retrieved images align semantically with the stimulus and share perceptual traits like orientation and color. Failure cases rely on superficial cues, such as background color, missing the target concept. Augmentation Effect in Various VLMs We analyze how our image augmentations behave across different vision-language models (VLMs). Specifically, we obtain embeddings of augmented images and retrieve nearby original images in the embedding space to examine which image representations the augmentations are closest to. The results are shown in Fig. 13. In CLIP-RN50, images with fovea blur predominantly retrieved semantically aligned images, whereas images with Gaussian blur tended to retrieve perceptually similar images, such as those sharing similar orientation or color. However, this effect was less pronounced in other VLMs; in particular, CLIP-ViT-L/14 exhibited minimal sensitivity to perceptual differences. These observations suggest that our method is more effective when applied to ResNet-based CLIP models. Figure 10: Feature distributions visualized on the Poincaré ball using HoronPCA. Interpolated visual features lie between perceptual and semantic features. Figure 11: Top-5 image retrieval results on the THINGS-EEG dataset for successful and failure cases. Figure 12: Top-5 image retrieval results on the THINGS-EEG dataset for successful and failure cases. Figure 13: Image retrieval results for various VLMs