some results

readme.md

@@ -1,20 +1,26 @@
 MNIST is the hello world of machine learning. When we learn out optimizer moves through peaks and valleys. But how bumpy are those valleys and how many are there?
 
-This project tries to map a slice of the MNIST classification problem. We reduce mnist to a binary problem: is the number > 5? 
+This project tries to map a slice of the MNIST classification problem. We reduce mnist to a binary problem: is the number > 5?
 
 - Then we pre-train a model
-- freeze all but two parameters. We call those x and y. 
-- Height, or z, is our loss. 
+- freeze all but two parameters. We call those x and y.
+- Height, or z, is our loss.
 - Then we peform a grid search over x and y to find the loss at differen't points.
 - finally we countour the landscape
-- and show the path of our model as it learns 
+- and show the path of our model as it learns
 
+Results:
 
-![](docs/countours.png)
+![](docs/loss.png)
+![](docs/logloss.png)
+![](docs/std.png)
+![](docs/logstd.png)
 
 
 TODO
 
-- [ ] run a finer grid
+- [x] run a finer grid
+- [ ] run a wider search
+- [ ] try on an untrained model
 - [ ] show path of a model as it learns
 - [ ] make gif