Merge pull request #13 from EleutherAI/non-linear

Add fleace experiment

experiments/fleace.py

@@ -0,0 +1,69 @@
+import torch
+import torch.nn.functional as F
+
+
+def v_transform(K: torch.Tensor):
+    """ReLU(Kernel) V-transform as described in Tensor Programs I: Wide Feedforward or
+    Recurrent Neural Networks of Any Architecture are Gaussian Processes https://arxiv.org/abs/1910.12478.
+    """
+    diag = torch.diagonal(K)
+    scale = torch.sqrt(diag.unsqueeze(1) * diag.unsqueeze(0))
+    c = K / scale
+
+    return (
+        (1 / (2 * torch.pi))
+        * ((1 - c.pow(2)).sqrt() + (torch.pi - torch.acos(c)) * c)
+        * scale
+    )
+
+
+def g(x: torch.Tensor, k: int, device="cpu"):
+    """Apply `ReLU(x) + ε` to x with ε ~ N(0, 1) and return the first k dimensions."""
+    e = torch.randn(x.shape, device=device)
+
+    return (F.relu(x) + e)[..., :k]
+
+
+def relu_eraser(x: torch.Tensor, n: int, k: int):
+    """Use closed-form solution for free-form LEACE for ReLU to remove linearly accessible
+    information about ReLU(x) from x as described in Non-Linear Least-Squares Concept Erasure.
+    """
+    # E[Z | X] where Z = ReLU(X) + ε and ε ~ N(0, 1)
+    f = F.relu(x)[:, :k]
+
+    # Closed form solution for E[X ReLU(X).T] in R^(n k)
+    # Note: cross_cov is not centered
+    cross_cov = torch.eye(n)[:, :k] * 0.5
+
+    # Closed form solution for E[ReLU(X) ReLU(X).T] in R^(k k)
+    V = v_transform(torch.eye(k))
+
+    A = cross_cov @ torch.linalg.pinv(V)
+
+    return x - (A @ f.T).T
+
+
+def test_v_transform_monte_carlo():
+    num_samples = 100_000
+    dim = 10
+    X = torch.randn((num_samples, dim))
+
+    cov = (F.relu(X).T @ F.relu(X)) / (num_samples - 1)
+
+    torch.testing.assert_close(cov, v_transform(torch.eye(10)), rtol=0.01, atol=0.01)
+
+
+def test_relu_linear_erasure():
+    batch_size = 2_000_000
+    n, k = 16, 8
+
+    x = torch.randn((batch_size, n))
+    z = g(x, k)
+    r_x = relu_eraser(x, n, k)
+
+    assert torch.norm((r_x.T @ z) / (batch_size - 1)) < 0.01
+
+
+if __name__ == "__main__":
+    test_v_transform_monte_carlo()
+    test_relu_linear_erasure()