Non-invasive neural recording modalities like electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) offer a window into human cognition without the risks of surgery. However, they suffer from an exceptionally low signal-to-noise ratio (SNR). Scalp EEG signals are blurred by the skull and heavily contaminated by muscle movements and environmental interference, while fMRI measurements are distorted by background physiological fluctuations.
When deep learning models attempt to map these raw, noisy signals directly to complex targets—such as reconstructing a perceived image or decoding silent thought—they face a fundamental bottleneck. As noted by Julie D. Golomb in Achieving more human brain-like vision via human EEG representational alignment., this low SNR creates a ceiling on how accurately artificial models can align with biological brain activity. Traditional end-to-end optimization often causes models to overfit to transient artifacts rather than isolating the underlying neural signatures. To overcome this, modern frameworks shift from direct signal mapping to strategies that leverage abstract latent spaces, generative priors, and structured state-space modeling.
Shifting from Raw Signals to Abstract Latent Spaces
Direct reconstruction of neural signals often fails because standard loss functions (like mean squared error) force a model to spend its capacity replicating high-frequency noise and muscle artifacts, which dominate the raw data stream.
To bypass this, Geoffrey Mackellar et al. introduced a self-supervised framework in EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs that operates on a different principle: predicting missing data within an abstract latent space rather than the raw signal space. By using a semantic masking strategy, the model learns to capture the broader contextual structure of the neural dynamics while ignoring localized, high-frequency noise.
Similarly, Bruno Aristimunha demonstrated in Evaluating the structure of cognitive tasks with transfer learning that pre-training models to learn transferable representations across different cognitive tasks helps ground the network in robust neural features, raising the performance ceiling when downstream data is scarce or noisy.
Leveraging Generative Priors as Conditioned Guides
Another breakthrough in overcoming low SNR is changing what we expect the brain signal to provide. In tasks like visual reconstruction, the noisy neural signal does not need to contain the exact pixel-level coordinates or fine-grained textures of an image. Instead, it can act purely as a conditioning prompt for a generative model that already understands the statistical structure of the visual world.
This intuition forms the basis of several modern decoding pipelines. For instance, Simone Palazzo et al. trained a recurrent neural network in Generative Adversarial Networks Conditioned by Brain Signals to extract high-level visual features from noisy EEG signals, using these features to condition a Generative Adversarial Network (GAN). Because the GAN already knows how to generate realistic objects, the noisy EEG signal only needs to provide enough energy to steer the generation toward the correct semantic category.
More recently, this concept has been extended to diffusion models. Dongyang Li et al. developed an Adaptive Thinking Mapper in Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion to map noisy EEG embeddings into the latent space of a guided diffusion model, enabling zero-shot image reconstruction.
In the domain of fMRI, where spatial structures are often blurred by noise, Yasheng SUN et al. designed DreamConnect in Connecting dreams with visual brainstorming instruction, which uses an asynchronous diffusion strategy and language-model guidance to stabilize and manipulate the ambiguous visual structures decoded from brain activity. By relying on generative models to fill in the structural details, these frameworks ensure that neural noise does not degrade the fidelity of the final output.
Structural and Temporal Filtering Mechanisms
Beyond latent spaces and generative priors, handling low SNR requires architectures designed to isolate true neural dynamics from background noise over time. Because brain activity is inherently sequential and multi-scale, standard architectures can easily lose the signal across long sequences.
To address the high-dimensional and noisy nature of dynamic functional network connectivity in fMRI, YUXIANG WEI et al. proposed a hierarchical spatio-temporal state-space model based on the Mamba architecture in Hierarchical Spatio-Temporal State-Space Modeling for fMRI Analysis. This approach allows the model to maintain long-term context and filter out transient noise by tracking state transitions rather than raw correlations.
On the algorithmic side, Niklas Smedemark-Margulies demonstrated in Recursive Estimation of User Intent from Noninvasive Electroencephalography using Discriminative Models that a recursive Bayesian update rule can allow discriminative deep learning models to iteratively refine estimates of user intent from noisy EEG data, preventing the loss of critical information during aggressive preprocessing.
Finally, Javier Fumanal-Idocin et al. showed in Motor-Imagery-Based Brain-Computer Interface Using Signal Derivation and Aggregation Functions that explicitly differentiating EEG signals to isolate neural spikes, combined with fuzzy ensemble aggregation, helps mitigate sensor redundancy and isolate meaningful intent from background interference.