A filter model for MT pattern and component cells
- To study a feedforward model for building MT cells from V1 DS units.
- To understand how pattern cell responses might be generated in MT
- Rust NC, Mante V, Simoncelli EP, Movshon JA (2006) How MT cells analyze the motion of visual patterns. Nature Neurosci 9:1421--1431.
- The MT model units of Rust et al.
Rust et al. have proposed a model of MT neurons that can produce units with both component and pattern cell responses using the same circuitry and mechanisms. In this model, there are 12 V1 direction columns whose responses are modeled as direction tuning curves. These responses are then scaled to simulate two types of V1 response normalization, one which is dependent on activity within each channel itself (tuned normalization) and the other where the normalization is driven by the pooled activity across all channels (untuned normalization). The normalized V1 outputs are then linearly combined with a given MT input weight distribution, with both excitatory and inhibitory (positive and negative) weights. Finally this summed signal is transformed through an output nonlinearity to generate the final MT unit firing rate.
The me_v5 model type has been developed in iModel to implement an image-based version of the Rust et al. model. Visit the model homepage. Open one of the versions of the model parameter file (use the link in the Summary paragraph on the model homepage). Each V1 channel in the me_v5 model is a non-opponent motion energy filter unit (see the filter object). Normalization is applied to the outputs of these units using the equations and parameters of Rust et al. (see the v1 object). Input weights for all V1 channels are specified directly in the .moo file and the equation and parameters for the output nonlinearity of the model MT unit are also specified as in the Rust model (see the v5 object). We have generated 5 model MT units with parameters corresponding to the 5 example neurons shown in Figures 4, 6 and 8 in the Rust et al. paper (see all 5 versions on the model homepage).
MT neurons are classified as pattern or component cells based on their responses to plaid stimuli consisting of two superimposed sinusoidal gratings with a fixed offset in drift direction. View this stimulus on iModel for a 120 degree direction offset at plaid120.stm To examine the responses of a model pattern MT unit to single sinusoids and plaids drifting in different directions, run the following simulation:
- Model: V5_ME
- Response: mt.rsp
- Stimulus: plaid120_v5_me.stm
- tn: 1024
- sf: 2.4
- stim_nrpt: 10
When the simulation finishes, follow the "Results available" link and click the button to "View zz.nd". Set contrast1 to 0.17 and contrast2 to 0, and select the "Tuning Curve" analysis. The display plots the mean firing rate of the MT unit as a function of direction for a single sinusoidal grating. You should notice the following:
- If you now set contrast2 to 0.17, the tuning curve will show the mean firing rate response of the MT cell to plaid stimuli. As for single gratings, this tuning curve has one peak. The pattern cell has the same direction preference for single gratings and plaid stimuli consisting of two gratings whose motion is separated by 120 degrees.
- Run the same simulation but using the component-type model V5_ME.b. Compare the single grating and plaid tuning curves for this unit with the plaid unit you saw above.
- Open the model parameter files (the default V5_ME model and V5_ME.b). Make a note of the parameters that differ between these models. Many of these were specified by the fits of Rust et al. to their example cells, but there are additional differences in parameters. What do the parameter values that differ between the models control?
- V1 tuning and MT output
One of the advantages of implementing the model with motion energy units for V1 channels is that we can alter and test the tuning of V1 responses in a principled way (e.g. changing key RF parameters rather than just generating different tuning curves), and see how this affects the tuning of the MT unit.
- Run a contrast tuning curve for V1 and MT signals in the V5_ME pattern cell:
- Model: V5_ME
- Response: mt_v1.rsp
- Stimulus: sine_con_v5_me.stm
- tn: 512
- sf: 2.4
- direction: 180
- stim_nrpt: 5
- Look at the tuning curve for the recorded raw V1 signal v1 and the signal after normalization, v1_norm. What is the effect of normalization on the contrast response of V1 units? Is this a reasonable approximation of contrast response in a V1 cell?
- What parameters could you change if you wanted to re-tune the V1 contrast response?
- Also check the contrast response of the MT unit, channel mt. How is this different from the V1 units? Is this reasonable for an MT output?
- Run another simulation with the sine_con_v5_me stimulus, but this time use the anti-preferred direction for the MT and V1 units, direction 0. What does the contrast response tuning for V1 and MT signals look like now? How is this different or similar to the direction dependence of the contrast response functions in actual V1 and MT units?
In summary, the Rust et al. model is one of the simplest feedforward models for building MT cells from V1 DS inputs, with only a few free parameters to specify. However the model requires elements at intermediate stages that are not consistent with physiological responses in the V1 cells that give input to MT.