Information transfer rate
Accuracy alone cannot compare a fast, error-prone speller with a slow, precise one. Information transfer rate (ITR) folds accuracy, the number of choices, and the time per selection into a single bits-per-minute figure by modelling the BCI as a noisy communication channel.
B=\log_2 N+P\log_2 P+(1-P)\log_2\!\frac{1-P}{N-1}Wolpaw bits per selection for N equiprobable targets at accuracy P. At P=1 it is log₂N; as P→1/N it falls to 0.
How much information a single selection carries depends on both how many choices there are and how often you are right. Perfect accuracy gives the full \log_2 N bits; pure guessing gives 0.
- N
- Number of possible targets.
- P
- Accuracy — the probability of a correct selection.
- B
- Bits of information conveyed per selection.
With 4 targets at 100% accuracy B is 2 bits; as accuracy falls to 25% it drops to 0. This feeds the information transfer rate.
\mathrm{ITR}=B\cdot\frac{60}{T_{\text{sel}}}\quad[\text{bits/min}]Bits per selection times selection rate gives throughput in bits per minute — the number to quote when comparing systems.
Multiply the bits per selection by how many selections you make each minute to get the system's true communication speed — the fair number for comparing BCIs.
- B
- Bits of information per selection.
- T_{\text{sel}}
- Seconds it takes to make one selection.
- \mathrm{ITR}
- Throughput in bits per minute.
2 bits every 4 seconds is 30 bits per minute.
ITR's fine print
ITR is a model, and its assumptions are routinely broken. It assumes a memoryless channel with stationary accuracy, equiprobable and independent targets, and no benefit from context. Real spellers violate all four — a language model exploits letter statistics, errors cluster, and users pause. Two systems with identical Wolpaw ITR can feel completely different to use.
Reading the decoder: representational analysis
A high ITR tells you that a decoder works, not what it learned. Representational similarity analysis (RSA) opens the box without touching the classifier. Build a representational dissimilarity matrix (RDM) — the pairwise distance between neural response patterns for every condition — then correlate it with the RDM predicted by a candidate model of the underlying code.
d_{ij}=1-\operatorname{corr}\big(\mathbf r_i,\mathbf r_j\big),\qquad r_{\mathrm{RSA}}=\operatorname{corr}\big(\mathrm{RDM}_{\text{neural}},\ \mathrm{RDM}_{\text{model}}\big)RDM entries as pattern dissimilarity; RSA score as the correlation between the neural and model dissimilarity structures. It compares geometries, not raw activity, so it survives across sessions and subjects.
Instead of comparing raw brain activity, compare the pattern of which conditions look alike versus different. If the brain and a model group the same conditions together, they share a representation — a comparison of geometry, not raw signals.
- \mathbf r_i
- The activity pattern evoked by condition i.
- d_{ij}
- Dissimilarity between conditions i and j (1 minus their correlation).
- \mathrm{RDM}
- The full table of pairwise dissimilarities.
- r_{\mathrm{RSA}}
- Correlation between the neural and model dissimilarity tables.
Because it compares geometries, RSA survives across sessions and even across subjects.
RSA reframes decoding as a question about geometry: is the structure a decoder exploits the one a task-relevant model predicts, or an artefact-driven shortcut? It connects the interpretability of deep decoders to a classical, model-comparison logic, and is a bridge to the population-dynamics view of the neighbouring track.
The open problems
The frontier of statistical decoding is not a better classifier on a fixed dataset — that problem is largely solved. It is stability over time. Neural non-stationarity and covariate shift mean a decoder frozen on Monday degrades by Friday, so the real research targets are recalibration and adaptive decoding that track drift without daily supervised setup.
The second frontier is data scarcity. Because calibration is expensive and non-stationary, the winning strategies borrow strength across sessions, subjects and datasets: transfer learning, domain adaptation, and the recent push toward neural foundation models pretrained on large unlabelled corpora and fine-tuned per user. Whether such models truly generalise, or merely memorise dataset idiosyncrasies, is exactly the question the validation discipline of Guides 3–4 exists to settle — honestly.