PaperlayerView exact revision on arXiv

Source-linked AI summary

Association profile conditioning in a set-temporal transformer for cross-session intracortical motor decoding

Xinyuan Zhang, Handong Mo, Pengfei Wen, Shuang Liang, Jichang Yang, Yan Zeng, Zhongrui Wang, Han Wang

arXiv:2609.39080v1q-bio.NCcs.AI

TL;DR

Intracortical motor decoders degrade across sessions as recorded units and their behavioral associations change, while existing approaches either update weights or rely on unlabeled activity. APST uses a few labeled calibration trials to condition a frozen-weight set-temporal decoder, improving over activity-only decoding on DANDI688 and reporting results on FALCON as well.

  • Problem

    Across sessions, recorded units can change and persisting units can change their associations with behavior; whether a few labeled trials can adapt a decoder with frozen weights remains unclear.

    Source: §1 · p. 1

  • Method

    APST computes a 4-D association profile for each unit from a few labeled calibration trials in closed form and uses the profiles to condition a set-attention encoder and causal temporal transformer with frozen weights.

    Source: §1 · p. 1

  • Results

    APST raises DANDI688 performance over Activity-only from 0.40 to 0.78 on Sub-C and 0.58 to 0.81 on Sub-M, and scores 0.65, 0.42, and 0.44 on FALCON M1, M2, and H1.

    Source: §3.2 · p. 3, §3.3 · p. 4

  • Takeaways & Limitations

    DANDI688 and FALCON results support profile conditioning for cross-session decoding without target-session weight updates.

    Source: §4 · p. 4

    At target deployment, all network parameters remain source-trained and frozen; profiles and unit identities are computed from calibration without target-session backpropagation.

    Source: §2.1 · p. 1

Abstract

from arXiv · show

Intracortical motor decoders degrade across sessions because the set of recorded units changes and persisting units can alter how their firing relates to behavior. Most existing methods update network weights on each new session or rely on unlabeled activity, which does not directly reveal such changes. We present APST, an Association Profile-conditioned Set-Temporal transformer that adapts to new sessions with all network weights frozen. From a few labeled calibration trials, APST summarizes how each unit's firing relates to behavior in a four-dimensional association profile computed in closed form. The profiles condition a set-attention encoder that accepts any number and order of units, followed by a causal transformer for streaming decoding. On held-out DANDI688 sessions from two monkeys, APST reaches velocity $R^2$ of $0.78$ and $0.81$, versus $0.40$ and $0.58$ for a variant that uses neural activity alone, and matches or exceeds an RNN fine-tuned on the same trials. On FALCON private held-out evaluation, it attains $R^2$ of $0.65$, $0.42$, and $0.44$ on M1, M2, and H1.

1. INTRODUCTION

Across sessions, electrode drift changes the recorded units and persisting units can change their behavioral associations, so a Day 0 decoder degrades without target-session calibration. APST addresses the open question of whether a few labeled trials can adapt a decoder while its network weights remain frozen.

1. INTRODUCTION

  • Slight array motion changes which units are recorded across sessions, and a fixed electrode index need not denote the same neural population.A decoder fit on Day 0 degrades on Day N without target calibration.Source: §1 · p. 1, Figure 1 · p. 1
  • Prior methods use latent alignment, target-session fine-tuning, or unlabeled activity, but unlabeled activity cannot resolve a persisting unit’s changed association with behavior.Labeled few-shot fine-tuning uses calibration labels but updates network weights, leaving it unclear whether frozen weights can adapt from those trials.Source: §1 · p. 1
  • A unit’s association profile is a task-conditioned functional identity summarizing its relationship to behavior, complementing activity signatures from the same trials.Profiles can capture directional tuning or behavior-weighted associations projected onto a source-fitted SVD basis for multidimensional kinematics or muscle EMG.Source: §1 · p. 1
  • APST computes each unit’s 4-D association profile from a few labeled trials in closed form, then conditions a permutation-invariant set encoder and causal temporal transformer with frozen weights.The introduction says APST is evaluated on DANDI688 and FALCON.Source: §1 · p. 1
  • On Sub-C, association profiles raise velocity R2 from 0.40 to 0.78 over an activity-only variant, while shuffled profiles score below activity-only.The evaluation includes DANDI688 recordings from Sub-C and Sub-M and FALCON benchmarks M1, M2, and H1.Source: §1 · p. 1

2. METHOD

APST uses a few labeled target-session trials to compute fixed four-dimensional unit association profiles and cached identities, then combines them with live activity for causal, set-based decoding. Its profile estimators and decoder are designed to handle changing unit sets without target-session network updates.

2.1. Notation and calibration boundary

  • All network parameters are trained on source sessions and frozen at deployment; target calibration computes each observed unit’s static 4-D association profile and identity.The calibration set contains M_s labeled trials.Source: §2.1 · p. 1

2.2. Association profiles

  • Association profiles summarize how each unit’s calibration activity relates to behavior, using a closed-form fit whose behavior enters through design covariates or sample weights.The fit requires one calibration pass and no target-session gradients.Source: §2.2 · p. 2

2.3. Directional tuning profiles (M2 and DANDI688)

  • For two-dimensional movement, directional profiles fit first-harmonic tuning from movement direction and collect directional coefficients, modulation depth, and baseline response.The fit uses movement-window mean counts per bin; profile components are standardized with source-session statistics.Source: §2.3 · p. 2

2.4. SVD-compressed association profiles (M1 and H1)

  • For M1 EMG and H1 velocity, behavior-weighted unit associations are projected onto a source-fitted four-dimensional SVD basis to match the shared profile interface.The method standardizes unit responses and shrinks associations toward zero when the total state weight is small; M1 uses K = 16 states and H1 K = 14.Source: §2.4 · p. 2

2.5. Profile-conditioned unit identities

  • APST combines a trial-averaged activity signature with its association profile to form a task-conditioned unit identity, using profile-based FiLM modulation and concatenation.The FiLM scale and shift are zero-initialized, so training starts from concatenation alone; identities are computed once and cached.Source: §2.5 · pp. 2–3

2.6. Set-temporal streaming decoder

  • At each time step, a shared causal convolution maps five bins of live activity to a unit representation, which is combined with cached identity and profile to form a unit token.The token MLP and layer normalization process the projected identity added to activity and the concatenated profile.Source: §2.6 · p. 3
  • Eight learned slot queries aggregate unit tokens with permutation-invariant set attention, and four causal transformer layers decode the population sequence using local windows and cached key/value state.The padding mask supports variable unit counts; the streaming state is bounded by the receptive field rather than the full session history.Source: §2.6 · p. 3

3. EXPERIMENTS

Experiments compare APST with baselines on DANDI688 and FALCON, then examine calibration budgets, held-in accuracy, and conditioning ablations. APST matches or exceeds fine-tuned RNN performance on DANDI688 and performs competitively on FALCON.

3.1. Datasets and protocol

  • The evaluation spans four 20 ms-binned benchmarks, with DANDI688 using sorted single units and FALCON using unsorted threshold crossings on private held-out sessions.DANDI688 RNN comparisons use 4, 8, 16, and 32 calibration trials; ablations and linear baselines use 32. Calibration costs 3.83 to 21.75 million MACs, with profile fitting accounting for < 0.1% (< 15k operations).Source: §3.1 · p. 3, §3.2 · p. 4

3.2. Main results

  • On FALCON private held-out evaluation, APST reaches 0.65 ± 0.11, 0.42 ± 0.10, and 0.44 ± 0.15 R² on M1, M2, and H1.It has the highest mean among methods without target-session gradients on M2 and H1, and is within 0.01 of SPINT on M1.Source: Table 1 · p. 3, §3.2 · p. 3
  • With frozen weights, APST reaches 0.78 on Sub-C and 0.81 on Sub-M, matching or exceeding RNN-FT and outperforming linear baselines.On the most distant Sub-C session, APST retains 0.58 versus 0.48 for RNN-FT and 0.06 for Activity-only.Source: Figure 3 · p. 4, §3.2 · p. 4
  • At 8 trials, APST reaches 0.75 on Sub-C and 0.77 on Sub-M, versus 0.64 and 0.47 for RNN-FT.On Sub-M, APST stays ahead at every 4–32-trial budget by 0.11 to 0.38.Source: Figure 3 · p. 4, §3.2 · p. 4

3.3. Ablation studies

  • Profiles raise performance over Activity-only from 0.40 to 0.78 on Sub-C and from 0.58 to 0.81 on Sub-M, with gains of 0.03, 0.18, and 0.20 on FALCON M1, M2, and H1.Shuffled profiles score below Activity-only in all five settings, supporting their role as unit-specific identities rather than session-level summaries.Source: Figure 3 · p. 4, §3.3 · p. 4
  • On FALCON, token-only and dual conditioning differ by at most 0.01 per task, within seed SD of 0.02, while identity-only falls to 0.40 on H1 versus 0.48.Dual conditioning is retained because cached identity conditioning adds no streaming cost.Source: Figure 3 · p. 4, §3.3 · p. 4

3.2. Main results

  • APST’s held-in R² of 0.77, 0.64, and 0.60 matches the strongest baselines, while its held-in-to-held-out losses versus SPINT are smaller on M2 and H1 and similar on M1.The reported losses are 0.22 versus 0.33 on M2, 0.16 versus 0.18 on H1, and 0.12 versus 0.11 on M1.Source: §3.2 · pp. 3–4, Table 2 · p. 4

4. CONCLUSION AND DISCUSSION

APST uses association profiles from a few labeled target-session trials to adapt without target-session weight updates. DANDI688 and FALCON results support profile conditioning, while the paper identifies task-specific profile design and unresolved conditioning roles as limitations.

4. CONCLUSION AND DISCUSSION

  • APST adapts a source-trained set-temporal decoder using unit–behavior association profiles estimated from a few labeled target-session trials.The same 4-D profile interface is used for planar and high-dimensional tasks.Source: §4 · p. 4
  • Results on DANDI688 and FALCON support profile conditioning without target-session weight updates, while sliding-window causal attention bounds streaming key/value state.Source: §4 · p. 4
  • Profile estimators and source bases remain task-specific, and how the two conditioning sites divide work across tasks remains uncharacterized.Future work will explore profile estimators learned jointly with the decoder across datasets and modalities.Source: §4 · p. 4

Compliance with Ethical Standards

The study used publicly released FALCON and DANDI688 recordings, conducted no new human or animal experiments, and relied on ethical approval obtained by the original data collectors.

Compliance with Ethical Standards

Loading 2609.39080v1…