Source-linked AI summary
Interpreting Encoding and Decoding Models
Nikolaus Kriegeskorte, Pamela K. Douglas
TL;DR
Interpreting fitted encoding and decoding models is difficult because significant performance and model weights do not straightforwardly identify computational mechanisms or feature contributions. The paper clarifies how model direction, generalization, and representational analyses constrain interpretation, concluding that multiple models must be compared to advance theory.
Problem
The paper addresses how to interpret empirical encoding and decoding results without treating significant single-model performance as evidence for the underlying computational process.
Method
The paper analyzes encoding and decoding models through their representational predictions, model direction, generalization levels, and complementary analysis methods.
Results
Significant model performance does not establish that a model captures brain computation, while fitted weights and chosen feature sets are not generally straightforwardly interpretable.
Takeaways & Limitations
Theoretical progress requires testing and inferentially comparing multiple models rather than relying on one significant encoding or decoding result.
Takeaways & Limitations
Decoding models generally reveal explicit information but cannot capture the complex nonlinear computations underlying brain processing.
Abstract
from arXiv · showhide
Encoding and decoding models are widely used in systems, cognitive, and computational neuroscience to make sense of brain-activity data. However, the interpretation of their results requires care. Decoding models can help reveal whether particular information is present in a brain region in a format the decoder can exploit. Encoding models make comprehensive predictions about representational spaces. In the context of sensory systems, encoding models enable us to test and compare brain-computational models, and thus directly constrain computational theory. Encoding and decoding models typically include fitted linear-model components. Sometimes the weights of the fitted linear combinations are interpreted as reflecting, in an encoding model, the contribution of different sensory features to the representation or, in a decoding model, the contribution of different measured brain responses to a decoded feature. Such interpretations can be problematic when the predictor variables or their noise components are correlated and when priors (or penalties) are used to regularize the fit. Encoding and decoding models are evaluated in terms of their generalization performance. The correct interpretation depends on the level of generalization a model achieves (e.g. to new response measurements for the same stimuli, to new stimuli from the same population, or to stimuli from a different population). Significant decoding or encoding performance of a single model (at whatever level of generality) does not provide strong constraints for theory. Many models must be tested and inferentially compared for analyses to drive theoretical progress.
Encoding and decoding: concepts with caveats
Encoding and decoding models are useful for interpreting information processing in brain regions, but their labels and causal interpretations depend on the chosen region and processing direction. A region-centered encoder–decoder division can also omit causal pathways in the brain’s recurrent, skipping network.
- Conceptual foundations: Encoding and decoding models help investigate what information neural activity represents and how information is transformed across stages of brain processing.The paper focuses on interpreting empirical results from fitted models applied to brain-activity data.
- Caveats: Whether a processing step is called encoding or decoding depends on the selected region of interest, not on an inherent difference in processing.Processing between regions X and Y can be decoding relative to X but encoding relative to Y.
- Caveats: Dividing processing around one region can miss causal paths because brain regions interact through skipping connections and recurrent signaling rather than simple chains.The primate visual hierarchy is described as a network with about a third of all possible pairwise inter-area connections.
- Caveats: If models are interpreted as brain-computation models, encoding and decoding should follow causal direction toward neural or motor representations rather than treating decoding as inversion back to stimuli.Inverse-decoding models can still reveal information present in a brain region, but they cannot be interpreted as process models of brain function.
Decoding models: revealing information and its format
Decoding models reveal whether a brain region contains particular information in a format the decoder can exploit, but they reveal computational products rather than the mechanisms that generate them. A linear classifier illustrates this by converting measured brain-activity patterns into stimulus labels through a weighted sum and threshold.
- Decoding models: revealing information and its format: Decoding can test whether a brain region contains a particular kind of information in a particular format, but it does not establish the brain’s computational mechanism.Decoding reveals the products of computation, not the process that generates them.
- Decoding models: revealing information and its format: A linear classifier takes a measured brain-activity pattern as input and outputs a class label, such as “cat” or “dog,” to identify the eliciting stimulus.It can compute a weighted sum of measured responses and apply a threshold to select the returned label.
- Decoding models: revealing information and its format: In a two-stimulus example, successful linear decoding shows that the stimuli elicit distinct response patterns, demonstrating mutual information between stimulus and response.This same mutual information could also be tested with an encoding model or a multivariate test of response-pattern differences.
- Decoding models: revealing information and its format: Univariate encoding models generally have less sensitivity because they do not account for noise correlations between different response channels.Multivariate analyses can account for noise correlations but may have less specificity and may fail to control false positives if their distributional assumptions are violated.
Linear decodability indicates “explicit” information
Decoding can establish that particular information is present and linearly exploitable in brain responses, but it does not by itself characterize the representation or computational mechanism. Encoding models instead predict representational spaces and can serve as brain-computational models, provided their fitted parameters are evaluated for generalization and overfitting.
- Decoding: Above-chance but imperfect decoding indicates linearly decodable information, yet errors cannot generally be attributed to a lack of linear separability because measurement noise and subsampling also limit performance.
- Stimulus reconstruction: Stimulus reconstruction from novel stimuli can indicate rich stimulus information, but reconstruction quality must be compared with the presented stimulus and interpreted relative to output-space complexity and priors.A complicated prior can make good-looking reconstructions reflect constraints on possible outputs rather than details encoded in the brain region.
- Decoding: Decoding analyses test whether particular information is present in brain responses using independent test data, while shifting analyses toward information relevant to computational function.Independent testing makes significant results less likely when model assumptions are violated.
- Decoding: Decoder weight maps are difficult to interpret because large weights can cancel noise, while priors and fitting procedures can select among essentially equally informative predictors.
- Encoding models: Encoding models predict brain responses from sensory stimuli and representational spaces, making them suitable brain-computational models, whereas decoding generally provides only weak constraints on computational mechanisms.Encoding-model inference must account for fitted parameters and overfitting, often through independent test sets when priors are used.
The single-model-significance fallacy
Significant variance explained by a single encoding or decoding model demonstrates information or dependency, not that the model captures the brain’s computation. Theoretical insight requires comparing multiple plausible models at matched levels of generalization and explainable variance.
- The single-model-significance fallacy: Computational interpretation requires models that capture brain computations at an appropriate abstraction while operating in the causal direction and remaining neurobiologically plausible.This requirement distinguishes computational models from statistical tools used merely to detect information in a brain region.
- The single-model-significance fallacy: A single model’s significant explained variance shows information or dependency, but does not establish that the model’s specific computational form is correct.Significant linear correlation demonstrates dependency without demonstrating linearity; likewise, significant variance from a complex encoding model does not validate its computational interpretation.
- The single-model-significance fallacy: Even a bad model can explain significant variance, particularly when many parameters are fitted to the data.Parameter-rich fits can achieve significance without providing evidence about the underlying brain computations.
- The single-model-significance fallacy: To learn about underlying brain computations, researchers should assess multiple models’ proportions of explainable variance at a given generalization level and compare them inferentially.The relevant variance is nonnoise variance, and generalization level must be held in view when evaluating competing models.