Project 2: How many dataset we need?

Project2.How Many Dataset We Need?

It is very comfortable to make the model with a lot of dataset.
But we already know, gathering enough dataset is very hard even we don’t know that how many dataset we need.
We just can assume that could be related with data by itself.
As a recommendation, someone said 1,000 images each class.

Dataset: MNIST fashion dataset
dataset shape: (60000, 28, 28)
dataset labels: (10, 60000),
label 0:9, 6000 images by label

Hypothesis: there is an optimal number of dataset for the model
+@ checking the result of accuracy and loss,
it would be gaussian distribution.
so, we can be estimating that best performance.

• Sampling the specific number of images by class
• 100 trials for each experiments
• gaussian fitting for them, train accuray / validation accuracy

Number of dataset by class	dataset split ratio	batch size	epoch	train mean	train std	validation mean	validation std
10	5:5	1	1	0.25	0.05	0.47	0.08
10	5:5	1	2	0.63	0.04	0.57	0.06
10	5:5	1	3	0.74	-0.06	0.6	0.06
10	5:5	1	4	0.81	0.07	0.63	-0.05
10	5:5	1	5	0.87	0.06	0.64	0.05
10	5:5	1	6	0.91	0.04	0.66	0.04
10	5:5	1	7	0.96	0.03	0.68	0.01
10	5:5	1	8	0.98	0.03	0.68	0.02
10	5:5	2	1	0.22	-0.05	0.47	-0.08
10	5:5	2	2	0.64	-0.04	0.58	0.06
10	5:5	2	3	0.77	-0.05	0.62	0.05
10	5:5	2	4	0.82	0.07	0.65	0.04
10	5:5	2	5	0.91	0.05	0.65	0.04
10	5:5	2	6	0.93	0.05	0.66	0.03
10	5:5	2	7	0.96	0.03	0.68	0.01
10	5:5	2	8	0.99	0.00	0.69	-0.03
10	5:5	3	1	0.2	-0.08	0.46	-0.06
10	5:5	3	2	0.64	-0.04	0.58	0.06
10	5:5	3	3	0.77	0.05	0.61	-0.07
10	5:5	3	4	0.83	0.06	0.64	-0.03
10	5:5	3	5	0.92	0.06	0.66	0.03
10	5:5	3	6	0.92	0.01	0.69	0.04
10	5:5	3	7	0.97	0.03	0.68	0.03
10	5:5	3	8	1.06	0.05	0.68	-0.03