[ Prev ]
2017-10-12

-- Oct 11 In-Class Exercise
Hypothesis 1: There would not be any change in accuracy as we increase the number of training sets beyond a certain number. This number would be small, as we are dealing with just 25*25 pixel images Hypothesis 2: It would be easy to train the plus symbol as it is distinct from the other symbols Hypothesis 3: Efficiency of training would decrease after adding around x number of stray marks
Hypothesis 1: There would not be any change in accuracy as we increase the number of training sets beyond a certain number. This number would be small, as we are dealing with just 25*25 pixel images Hypothesis 2: It would be easy to train the plus symbol as it is distinct from the other symbols Hypothesis 3: Efficiency of training would decrease after adding around x number of stray marks

-- Oct 11 In-Class Exercise
Hypothesis for exp 2: Given a certain no. of training examples, the model predicts all the symbols with equal accuracy
Hypothesis for exp 2: Given a certain no. of training examples, the model predicts all the symbols with equal accuracy

-- Oct 11 In-Class Exercise
Name: Kunal Deshmukh Hypothesis1 : With lesser images to train, error rate would be greater. Hypothesis2 : test data and train data should be of same quality. Otherwise model won't be able to predict correct label name from test data.
Name: Kunal Deshmukh Hypothesis1 : With lesser images to train, error rate would be greater. Hypothesis2 : test data and train data should be of same quality. Otherwise model won't be able to predict correct label name from test data.
2017-10-15

-- Oct 11 In-Class Exercise
Hypothesis: There exists a set of hyperparameters and trained weights that can accurately classify the images on Zener cards given sufficient number of examples and time to train.
Description:
Test cases for the hypothesis would scan the following properties:
• Varying training set size (how much ground truth does the model need?)
• Varying classification letter (are some easier to learn than others?)
• Vary the hyperparameter epsilon to see how increasing or decreasing it alters the model (how much leeway do we have for increased accuracy before overfitting, or reduced runtime without failure to find the maximal separator?)
(Edited: 2017-10-15)
Hypothesis: There exists a set of hyperparameters and trained weights that can accurately classify the images on Zener cards given sufficient number of examples and time to train. Description: Test cases for the hypothesis would scan the following properties: • Varying training set size (how much ground truth does the model need?) • Varying classification letter (are some easier to learn than others?) • Vary the hyperparameter epsilon to see how increasing or decreasing it alters the model (how much leeway do we have for increased accuracy before overfitting, or reduced runtime without failure to find the maximal separator?)

-- Oct 11 In-Class Exercise
Hypotheses
  1. A larger training set will produce more accurate models. a. test this hypothesis by varying the training set size by increments, have a fixed test set size, and compare the accuracy of each run. If our hypothesis is correct, we would expect a positive correlation between training set and accuracy.
  2. The star shape will be the easiest shape to recognize a. we will test this hypothesis by keeping the training and test set sizes fixed, and compare the accuracy of each model trained for a specific shape.
  3. generating noise on the images will make the model have more difficulty to train a. we will tweak the parameters in image generation, making the stray marks larger and more frequent. we will train models based on this noisy data and compare them to models trained on clean data. b. our hypothesis will be correct if the models on clean data converge faster, and are more accurate to their noisy counter parts.
(Edited: 2017-10-16)
Hypotheses # A larger training set will produce more accurate models. a. test this hypothesis by varying the training set size by increments, have a fixed test set size, and compare the accuracy of each run. If our hypothesis is correct, we would expect a positive correlation between training set and accuracy. #The star shape will be the easiest shape to recognize a. we will test this hypothesis by keeping the training and test set sizes fixed, and compare the accuracy of each model trained for a specific shape. #generating noise on the images will make the model have more difficulty to train a. we will tweak the parameters in image generation, making the stray marks larger and more frequent. we will train models based on this noisy data and compare them to models trained on clean data. b. our hypothesis will be correct if the models on clean data converge faster, and are more accurate to their noisy counter parts.
2017-10-30

-- Oct 11 In-Class Exercise
Hypothesis: large number of training samples results in better prediction for SVM.
Test: Increase the number of training samples and compare accuracy
(Edited: 2017-10-30)
Hypothesis: large number of training samples results in better prediction for SVM. Test: Increase the number of training samples and compare accuracy
X