Description of problem
I am working on implementing the paper: Pattern recognition using Spiking Neural Networks. This is also posted under the Projects category. As per the paper, STDP is the learning rule applied. I have used the Brian2 STDP example from the official Brian2 tutorials. The problem is on running the script, the weight converges to 0 despite getting the correct firing pattern of neurons as given in the paper
Minimal code to reproduce problem
taupre = taupost = 20*ms wmax = 1 Apre = 0.01 Apost = -Apre*taupre/taupost*1.05 stdp_eqs = ''' w : 1 dapre/dt = -apre/taupre : 1 (event-driven) dapost/dt = -apost/taupost : 1 (event-driven) ''' stdp_pre = ''' v_post += w * mV apre += Apre w = clip(w+apost, 0, wmax * mV) ''' stdp_post = ''' apost += Apost w = clip(w+apre, 0, wmax * mV) ''' # Inp -> Hidden Syn_inp_hid = b2.Synapses(izh , izh_hidden , model = stdp_eqs , on_pre = stdp_pre , on_post = stdp_post , method = 'euler' , name = 'Syn_inp_hid') # Hidden -> Output Syn_hid_out = b2.Synapses(izh_hidden , izh_output , ''' w : 1''' , on_pre = 'v += w * mV' , method = 'euler' , name = 'Syn_hid_out') # MAKE CONNECTIONS Syn_inp_hid.connect() Syn_hid_out.connect() Syn_inp_hid.w = 'rand() * 0.8' Syn_hid_out.w = 'rand() * 0.8'
What you have aready tried
On setting wmax to very high values like say 500, I get the following actiivty from the weights
Expected output (if relevant)
This is the graph taken from the above paper, the weights can be seen climbing to higher values and 0.8 is considered to be the max value specified by the authors.
Actual output (if relevant)
The STDP plot is at the bottom. This is the graph generated when I run the script.
The architecture is feedforward (i.e.) Input -> Hidden Layer -> Output. I have made the Synapses connections accordingly. Any suggestion on what to look for would be appreciated.