def naive_train(data):
    M, N = data.shape
    print("data shape", M, N)
    
    W = np.zeros((M, M))
    rho = np.sum([np.sum(t) for t in data]) / (N * M)
    for k in range(N):
        t = data[:, k] - rho
        W += np.outer(t, t)

    W /= N
    
    # Zeroing the diagonal
    np.fill_diagonal(W, 0)
    return W

# This function approximates the behavior of the MATLAB satlins function 
# (saturating linear transfer) using np.clip().
def satlins(x):
    return np.clip(x, 0, 1)

def predict(data, W):
    CY = []

    M, N = data.shape
    # Iterating over each column of t
    for k in range(N):
        u = data[:, k]
        while np.linalg.norm(satlins(W @ u) - u) > 0:
            u = satlins(W @ u)
        CY.append(u)

    CY = np.array(CY).T  # Convert list of arrays to 2D array
    return CY