Why inputs to M1 are hard to separate
M1 activity mixes two sources: patterns the cortex generates on its own, and signals arriving from elsewhere. Dynamical-systems accounts of reaching treat M1 as a transiently autonomous system whose rotational dynamics are set up by an initial state and triggered by a condition-invariant signal. That picture fits brief, largely open-loop reaches, but it leaves little room for the intermittent external signals a monkey needs when a target moves and the movement must be revised mid-flight.
The paper's problem is therefore an inference problem. If a population response could be explained either by intrinsic dynamics or by an external input, the data alone may not say which. The authors call this nonidentifiability: different models, or different hyperparameter settings, can fit the recorded firing rates indistinguishably while implying different input structures. Any claim about how M1 encodes corrective instructions depends on first showing that the inputs can be recovered at all.
LFADS splits intrinsic dynamics from inferred inputs
The authors modelled spiking data with Latent Factor Analysis via Dynamical Systems (LFADS), a recurrent-network autoencoder. For each trial it estimates an initial state, runs a learned dynamical system (the generator) forward to produce low-dimensional factors, and maps those factors back to spike counts through a generalized linear model. A side controller adds a time-varying input stream whenever the autonomous generator cannot reconstruct the data, so an inferred input marks a moment when population activity is not parsimoniously explained by the cortex's own dynamics.
The data came from Utah arrays in M1 of three male macaques performing a Random Target Pursuit task, with movement constrained to the horizontal plane by a two-link exoskeleton. A new target appeared as soon as the previous one was contacted, and monkeys were not required to stop or hold, so they produced continuous looping trajectories that demanded frequent online revision. That design is what makes corrective inputs observable in the first place.
Inputs earn their place by improving decoding
Because LFADS is trained on neural data alone, without behaviour, agreement with behaviour is an external check on the fit. Including external inputs improved decoding of movement kinematics relative to models without inputs: r² = 0.75 ± 0.08 with inputs versus 0.39 ± 0.08 without, reported as mean ± s.d. across monkeys. Decoding from LFADS-denoised single-trial rates also beat decoding from the raw firing rates (r² = 0.46), indicating that the model recovered usable single-trial trajectories rather than only average structure.
The inferred inputs themselves were sparse: high Gini coefficients indicated that activity was concentrated in time, and sparsity peaked near the input-penalty values that gave the best kinematic decoding. The authors therefore focused on the large, brief excursions rather than the low baseline. This matters for interpretation, since a sparse transient is a different object from a sustained drive, and the penalty used to obtain it shapes what the model reports.
Input timing locks to the target, not the reach
Averaged across target presentations, inferred input magnitude peaked about 50 ms after the target appeared, in all three monkeys. The peak coincided with previously reported latencies from a visual movement cue to the start of M1 activity changes. When the authors raised the threshold for counting an input, the peaks became sharper around that 50 ms mark, meaning the largest transients were also the most consistently timed.
The decisive comparison was between two lockings of the same transient: latency to target presentation versus latency to movement initiation. Latencies relative to target presentation had both smaller variance and a smaller width at half-height than latencies relative to the initial movement (p < 0.001 for both, permutation test). The authors read this as evidence that the transients relate to visual input, and that variable reaction times are not explained by variability in input timing but may arise from how those inputs interact with the current state of M1's dynamical system. That is an interpretation of the timing pattern, not a direct measurement of the interaction.
Inputs carry target direction in hand-centred coordinates
To test whether the transients were meaningful, the authors binned targets by the direction from hand to target and averaged inputs for each bin, producing tuning-like curves. Decoding target position from inputs worked well only in a narrow window: performance fell sharply once the decoding window started more than 50 ms after target presentation. Firing rates, by contrast, predicted target position stably across the whole trial. The authors' explanation is that inputs briefly encode the movement but persistently alter the system's state, so the encoding is sustained downstream in M1 activity.
The inputs also carried a coordinate frame. A nonlinear decoder predicting target position performed better with the hand as origin than with the shoulder: r² = 0.70 versus 0.53 for monkey RS, 0.63 versus 0.39 for MK, and 0.46 versus 0.21 for RJ (p < 0.001 for each monkey, corrected t test), with similar results in a second session per monkey. The authors compare this to hand-centred representations upstream of M1 in the fronto-parietal reach network, while cautioning that a binary hand-versus-shoulder comparison is a coarse description of reference frames.
Model ensembles decide when inputs can be trusted
The identifiability problem is addressed with simulations where ground truth is known. LFADS recovered single-trial trajectories accurately and reliably, but its separation of dynamics from inputs was identifiable in some simulated cases and not in others. Ensembles of models with varied hyperparameters served as the diagnostic: when ensemble members agreed, the inputs were identifiable; when they disagreed, the fit was inconclusive. Networks with transient inputs gave better ensemble consensus and better agreement with ground truth than networks with sustained inputs.
Applied to the monkeys, ensembles built from decorrelated-input, fast-decay, and slow-decay settings agreed consistently for two animals, supporting meaningful inference of input coding. For the third monkey, RJ, some ensemble results reached significance only at p < 0.05, indicating that those data were only marginally identifiable, and the initial-versus-corrective comparison was inconclusive. The authors could not determine the cause; they list lower trial counts, array placement, which neurons were sampled, input time constants, and RJ's substantially slower movements as possibilities, and stress that any such explanation remains speculative without ground-truth inputs.
Sources used in this answer
Combatting nonidentifiability to infer motor cortex inputs yields similar encoding of initial and corrective movements
LFADS model ensembles inferred sparse M1 inputs locked to target appearance and tuned similarly for initial and corrective movements in two of three macaques, while a third monkey's data were only marginally identifiable.
