Update README.md

README.md

@@ -17,9 +17,9 @@ metrics:
 
 ## NULA 
 
-NULA is an anti-aliased residual convolutional neural network for CIFAR-10 image classification. It was designed to improve robustness against perturbations that exploit downsampling.
+NULA, an anti-aliased residual convolutional neural network for CIFAR-10 image classification, trained to be robust against perturbations that exploit downsampling operations.
 
-Standard image models often rely on fragile high-frequency ques, which downsampling operators destroy or alias exactly to those components.
+Classical image models rely on fragile high-frequency cues, which downsampling operators destroy or alias exactly to those components.
 
 NULA is trained to reduce this dependence and instead form representations that remain stable under information-destroying transformations such as resizing, decimations, and aliasing-style perturbation.
 
@@ -28,7 +28,7 @@ NULA is trained to reduce this dependence and instead form representations that
 
 Downsampling operations are linear maps from a high-dimensional space to a lower-dimensional one.
 
-By the Rank-Nullity theorem, this matrix has a massive NULL space. 
+By the Rank-Nullity theorem, this matrix has a massive NULL space ;). 
 
 An attacker can exploit this: they utilize the discarded samples of these downsampling operations as extra degrees of freedom. 
 
@@ -36,20 +36,14 @@ By sculpting perturbations with components in the null space of the downsampling
 
 The result is an image perceptually identical to the original, with a manipulated activation pattern.
 
-```python
-
-```
 ## Approach
 
 [BlurPool: what it does mechanically and why it addresses the problem]
 [SE blocks: what they add]
+The augmentation functions used during adversarial training are in [`augmentations.py`](augmentations.py). 
+The training script is [`train_robust.py`](train_robust.py).
 
 
-## Architecture
-
-[stem → s1 → s2 → s3 → head, with channel dims]
-[BlurPool replaces strided conv — note this explicitly]
-
 ### FIRST EVALUATION (Base)
 
 The first evaluation was trained for clean classification performance without the robust training procedure described above.