The Binner

[chongxi]

• All decoders take bins of data as input, while the BMI output individual spikes.
• A binner is supposed to take each individual spike and update its states, including the current bin number, the real-time updated spike count vector of N neurons etc.
• Once it moves to the next state (next bin), it will trigger the decoder by emitting an event.
• The decoder will respond when the minimum required #bin is there, which includes the predefined pre_bins and current_bin.
• The length of the bin should be configurable, e.g. 125ms (8Hz), 66.67ms (15Hz), 33.33ms (30Hz), 16.67ms (60Hz) etc.

[chongxi]

Binner API

binner = Binner(bin_size=33.33, n_id=N, n_bin=B)    # binner initialization (space and time)
binner.input(bmi_output, type='individual_spike')   # internal state update triggered

[chongxi]

Binner Process

For LSTM:

[chongxi]

Binner Internal Parameters:

• N N neurons give rise to N spike count in each bin
• B B bins
• bin_size The time span (ms) to compute the spike count of each bin

Internal States:

• count_vec (+1 on with the input spike_id)
• nbins (+1 when the input timestamps goes to the next bin, its number is the current bin)
• output (emitted variable to the decoder, N neuron's spike count in previous B bins)

when nbins grows, the binner emits the decode event with its property output. IMPORTANT: make sure that the bmi_output.spk_id is orthogonal

def input(self, bmi_output, type='individual_spike'):
    current_bin = bmi_output.timestamp//self.bin_size
    if current_bin == self.nbins:   # state integrate
        self.count_vec[bmi_output.spk_id, self.nbins] += 1
    elif current_bin > self.nbins:   # state output
        self.nbins += 1
        self.count_vec[bmi_output.spk_id, self.nbins] += 1
        self.count_vec = np.hstack((self.count_vec, self.new_empty_bin))
        self.emit('decode', var=self.output)

@property
def output(self):
    self._output = self.count_vec[:N, -B:]
    return self._output

@property
def new_empty_bin(self):
    return np.zeros((N,1))