diff options
Diffstat (limited to 'sdil/core.py')
| -rw-r--r-- | sdil/core.py | 53 |
1 files changed, 49 insertions, 4 deletions
diff --git a/sdil/core.py b/sdil/core.py index 8b3f923..e66dc39 100644 --- a/sdil/core.py +++ b/sdil/core.py @@ -102,7 +102,8 @@ class SDILNet: def __init__(self, sizes, act="tanh", device="cpu", seed=0, w_scale=1.0, a_scale=1.0, feedback="error", dtype=torch.float32, nuis_rho=0.0, nuis_seed=1234, residual=False, - predictor_mode="diagonal", traffic_mode="soma"): + predictor_mode="diagonal", traffic_mode="soma", + vectorizer_mode="linear"): # sizes = [n_in, n_h1, ..., n_hk, n_out] self.sizes = list(sizes) self.L = len(sizes) - 1 # number of weight layers @@ -132,8 +133,24 @@ class SDILNet: self.W = [he((sizes[i + 1], sizes[i]), w_scale) for i in range(self.L)] self.b = [torch.zeros(sizes[i + 1], device=device, dtype=dtype) for i in range(self.L)] - # apical vectorizer A_l : c (n_classes) -> hidden velocity (n_l), hidden layers only (l=1..L-1) + # Apical vectorizer A_l : c (n_classes) -> hidden velocity (n_l), + # hidden layers only. A fixed linear map cannot represent credit that + # changes across activation regions. Optional zero-initialized gated + # terms preserve the initial linear model while letting perturbation + # calibration learn state-dependent feedback locally. + if vectorizer_mode not in ("linear", "soma_gated", "context_gated"): + raise ValueError(f"unknown vectorizer_mode: {vectorizer_mode}") + self.vectorizer_mode = vectorizer_mode self.A = [he((sizes[i + 1], self.n_classes), a_scale) for i in range(self.L - 1)] + if vectorizer_mode == "linear": + self.A_gate = None + elif vectorizer_mode == "soma_gated": + self.A_gate = [torch.zeros_like(weight) for weight in self.A] + else: + context_dim = sizes[-2] + self.A_gate = [torch.zeros((sizes[i + 1], self.n_classes * context_dim), + device=device, dtype=dtype) + for i in range(self.L - 1)] if predictor_mode not in ("diagonal", "full"): raise ValueError(f"unknown predictor_mode: {predictor_mode}") self.predictor_mode = predictor_mode @@ -245,9 +262,29 @@ class SDILNet: traffic = topdown if self.traffic_mode == "topdown" else (soma + topdown) / _SQRT2 return self.nuis_rho * traffic + def vectorizer(self, l, c, h_l=None, context=None): + """Return the learned instructional component of apical activity. + + ``soma_gated`` adds ``h_i G_i c`` independently in every cell. + ``context_gated`` adds a linear readout of ``c outer context``. Both + vanish when c=0, so neutral-period predictor identification remains + unchanged. + """ + value = c @ self.A[l].t() + if self.vectorizer_mode == "linear": + return value + if self.vectorizer_mode == "soma_gated": + if h_l is None: + raise ValueError("soma-gated vectorizer requires somatic activity") + return value + h_l * (c @ self.A_gate[l].t()) + if context is None: + raise ValueError("context-gated vectorizer requires a context state") + features = (c.unsqueeze(2) * context.unsqueeze(1)).flatten(1) + return value + features @ self.A_gate[l].t() + def apical(self, l, c, h_l=None, context=None): """a_l = A_l c + ordinary non-teaching apical traffic.""" - a = c @ self.A[l].t() + a = self.vectorizer(l, c, h_l, context) if h_l is not None: a = a + self.apical_traffic(l, h_l, context) return a @@ -510,8 +547,16 @@ def sdil_step(net, x, y, y_onehot, cfg, step, prev_error=None): # ================= apical vectorizer A via node perturbation ======= if did_pert: for l in range(net.L - 1): - dA = (qs[l] - r_list[l]).t() @ c / B # (n_l, n_classes) + calibration_error = qs[l] - r_list[l] + dA = calibration_error.t() @ c / B # (n_l, n_classes) net.A[l] += cfg.eta_A * dA + if net.vectorizer_mode == "soma_gated": + dgate = (calibration_error * h[l + 1]).t() @ c / B + net.A_gate[l] += cfg.eta_A * dgate + elif net.vectorizer_mode == "context_gated": + features = (c.unsqueeze(2) * context.unsqueeze(1)).flatten(1) + dgate = calibration_error.t() @ features / B + net.A_gate[l] += cfg.eta_A * dgate # ================= predictor P (neutral) ========================== # KEY identification condition. P must learn the soma->apical coupling |