AVITEWRITE MODEL OF HANDWRITING LEARNING

1. AVITEWRITE MODEL OF HANDWRITING LEARNING Marym Naghibolhosseini 810185148 A Benchmark for Biological System Modeling

2. Importance of this Topic • Learning of complex movements • Reactive imitation predictive, automatic performance • Error-prone movements correct movements • Multimodal velocity profiles unimodal, bell-shaped velocity profiles • New learning during reactive movements

3. VITE model for reaching • Straight reaching movements with bell-shaped velocity profiles • How synchronous multijoint reaching trajectories could be generated at variable speeds. • How arm movements are influenced by proprioceptive feedback and external forces. • Uses a vectorial representation of movement direction and length. • Limitation: it did not explain how curved movements could be generated

4. VITE model description • population vector hypothesis • motor and parietal cortex compute a vectorial representation of movement direction in motor and/or spatial coordinates. • A population vector is defined as a “weighted vector sum of contributions of directionally tuned neurons” • A Difference Vector (DV) is computed • The DV is multiplied by a gradually increasing GO signal (alters trajectory speed without changing trajectory shape) • The DV times the GO signal is integrated at the PPV until the present position of the hand reaches the target.

5. VITEWRITE • Curved movements are generated using a velocity-dependent stroke-launching rule. • Asynchronous superposition of multiple muscle synergy activations with unimodal, bell-shaped velocity profiles for each synergy. • Synergy-overlap strategy to generate curved movments • Movements is controlled by a predefined sequence of planning vectors. • No address how these planning vectors may be discovered, learned, and stored in a self-organizing process AVITEWRITE

6. What is AVITEWRITE • Cortical VITE + VITEWRITE trajectory generation models + the cerebellar spectral timing model of Fiala et al (1996). • A neural model of corticocerebellar interactions during attentive imitation and predictive learning of sequential handwriting movements. • Model of perception/action cycle of handwriting, which involves vision, attention, learning, and movement.

7. Properties of the model • Decrease in writing time as learning progresses • Generation of unimodal, bell-shaped velocity profiles • Size scaling with isochrony • Speed scaling with preservation of the letter shape and the shapes of the velocity profiles • Inverse relation between curvature and tangential velocity • Two-thirds power law relation between angular velocity and curvature

8. Properties of the model • System capable of • Reactive movements • memory-based movements • Switching between these movements • system of on-line movement error correction

9. Some basic concepts • Movement synergies • Strokes • Via-point movements • Inverse relation between curvature and velocity • Direction reversal • Synergy switching

10. Diagram of the model

11. 1. Visual attention • Focuses on the current hand position. • Moves to select a target position. • Focused within a circular region. • Decreasing the attention radius increases the correlation between the model and the human subjects performance (position, velocity, acceleration) with more learning trials. • Small attention radius may prevent convergence in a reasonable period of time. • large attention radius will yield a poor trajectory, which converges quickly.

13. 2. Difference Vector Representation • Formed between the current hand position (PPV) and the new target position (TPV). • Vision provides direction and amplitude information, in the form of DVvis. • Activates the appropriate muscle synergy. • TPV , DVvis drive the movement and act as teaching signals to train a cerebellar spectral memory via climbing fiber inputs. • DVvis, to a trajectory generator which can combine temporally overlapping muscle synergy activations to generate curved movements whose speed and size are volitionally controlled. • The trajectory generator then starts to integrate the memory-enhanced difference vector, DVS, generating a velocity vector that drives movement to the target

14. Visual difference vector • Memory available and movement is sufficiently accurate then memory directs the movement • Memory signal is too small or an error is made by deviating from the attention radius around the template curve, then vision controls the movement direction.

15. Cerebellar spectral componentCerebellar synaptic weights • Term gi models activation of Purkinje cell i at time t • Each synaptic weight is modified only if its spectral component gi is active and visual target information is available

16. Adaptively timed cerebellar output • Each term hi = gi.zi provides a local view in time of the learned information. • R is generated at a fixed rate in response to a given density of PC spectral components gi through time • Decreasing spectral density allows movement learning at variable speeds.

17. Working memory • Stores learned motor commands to allow them to be executed at decreased speeds as the speed and size of trajectory generation are volitionally controlled through the basal ganglia Execution of movements at variable speeds • Possibly in prefrontal cortex, forms a category representation of each letter, which controls adaptive pathways to all the synergies. • The letter category determines which cerebellar spectra, corresponding to the particular synergies needed to write that letter, are activated via mossy fiber inputs • Only those adaptive pathways that were modified due to prior learning will read-out nonzero values of the cerebellar spectral memory output, R • R is temporarily stored in a working memory buffer, simulated as a discretely sampled set of values from the continuous cerebellar output

18. Memory-enhanced differencevector • Competition between reactive movement and memory-based movement control systems. (the cerebellar motor memory competes for control of movement with prefrontal and premotor areas that guide reactive movements based on visual input) • working memory output, WM, is combined with the visual difference vector, DVvis, and scaled by a size-controlling GRO signal, S, to form the size-scaled, memory-enhanced difference vector, DVS

19. The speed-controlling GO signal Present position vector

20. Memory-modulated target Gating difference vector • It tracks the cumulative DVs through time • DVgate controls readout from the WM buffer • If Dvgate ≤ 0 The next cerebellar command that is read from the WM buffer • WM readout is controlled by the speed of the movement (the shapes of the movement and its velocity profile are preserved as performance speed is changed)

21. Acceleration profileMuscle-filtered acceleration profile

22. Correlation equationCurvature

23. The cerebellar adaptive timing system • Learns the activation pattern of the muscle synergy involved in the movement • Cooperate or compete (6) with reactive visual attention for control of the motor cortical trajectory generator • Adaptively timed learning of strokes may be achieved by Spectral timing in the cerebellum • Cerebellum open a timed gate to express a learned motor gain • Conditioned stimulus trigger a spectrum of phase-delayed depolarizations of the Purkinje cells • Unconditioned stimulus trigger teaching signal in climbing fibers then Long Term Depression (LTD) of active Purkinje cells may occur at that time • Disinhibition of the cerebellar nuclei

24. The cerebellar adaptive timing system+Handwriting • Can learn a continuous response over time in complex tasks like handwriting • Different Purkinje cell spectra are activated by the commands corresponding to different muscle synergies. • The climbing fiber unconditioned stimuli act as error-based signals that train the Purkinje cells to become hyperpolarized in specific temporal patterns that lead to correctly shaped writing movements

25. Voluntary attention control • Assumes that attention can be voluntarily controlled to achieve a desired level of accuracy, or else to complete learning in a limited time at the expense of accuracy. • Directing unpracticed movements • Superior frontal, inferior parietal and superior temporal cortex, Prefrontal cortex was activated in a finger movement-sequence learning task during new learning but not during automatic performance after learning • The left dorsal prefrontal cortex was reactivated when subjects paid attention to the performance of a previously learned movement sequence

26. Cortical working memory • transiently stores the cerebellar output and releases it at a variable rate, allowing speed scaling of learned movements which is limited by the rate of cerebellar memory readout

27. Properties of human handwriting (when AVITEWRITE learns to write a letter) • Size and speed can be volitionally varied after learning while preserving letter shape and the shapes of the velocity profiles • Writing letters of different sizes in the same amount of time (isochrony) • Learning at slower speeds facilitates future learning at faster speeds • Unimodal, bell-shaped velocity profiles as a letter is learned • Inverse relation between curvature and tangential velocity • Two-thirds power law relation between angular velocity and curvature

28. Methods • raw position and time data were collected • velocity, acceleration, and curvature were calculated. • the vertical (y-direction) velocity zero crossings were used to determine strokes and separate adjacent letters • Averaging to enhance signal-to-noise ratio • Each letter prototype was scaled in size • The end of a letter was defined as the falling of both x and y velocities below a threshold (0.006) when within a threshold distance of the end of the letter being traced. • performance was evaluated by calculating the correlations between the model trajectory, velocity, and acceleration with the human data • Model velocity and acceleration were first scaled • correlation between the models tangential velocity and the tangential velocity predicted by the two-thirds power law • The correlation between the human tangential velocity and that predicted by the two-thirds power law

29. Minimum snap model • bottom-up model of trajectory formation • based on dynamic minimization of the square of the third (jerk) or fourth (snap) derivative of hand position • The version which minimizes snap yielded better correlation with human experimental data. • The model assumes that all letters are formed by a concatenation of shape primitives • the model generates each stroke primitive by use of a via-point, an intermediate target prior to the end of the stroke • strong correlations are reported between model-generated position, velocity and acceleration traces and the human counterparts. • The inverse relation between movement velocity and curvature

30. Result • correlations between the model and the human data, averaged over x and y position, velocity and acceleration over all letters for all subjects • Model performance was variable across the subjects • The main differences between the human data and model output are a variable stretching or compression of parts of the model velocity and acceleration profiles relative to the human profiles • Two-thirds power law prediction of tangential velocity has singularities at points where the curvature is zero

31. Refrences • Grossberg S. and Paine R.W.,” A neural model of cortico-cerebellar interactions during attentive imitation and predictive learning of sequential handwriting movements,” Neural Networks, Vol.13, pp.999–1046, 2000. • Paine R.W., Grossberg S., Van Gemmert A.W.A., “A quantitative evaluation of the AVITEWRITE model of handwriting learning,” Human Movement Science, Vol.23, pp. 837–860, 2004.

32. Thank You For Your Kind Patiance