Project 1

Report content (10):

Data description and preprocessing

(if necessary normalization, feature selection, transformation, etc.). Motivation for choosing the particular problem.

Because of the nature of our problem (image classification) we need to aprouch this with that in mind

• We choose to use normalization : Normalizing the pixel values of the images can help improve convergence during training. This typically involves scaling the pixel values to a specific range, such as [0, 1] or [-1, 1]. we did [0, 1]

thats it

Data visualization

(histograms, box plots, other plots).

Imagens msm com as labels

Short description of the implemented ML models.

-> Logistic Regression I guess -> MLP feed forward nn

Model training

(data splitting – train, validate, test, k-fold Cross validation). Visualize graphically the cost function trajectory over iterations. Training with regularized and nonregularized cost function.

show cost function tajectory and accuracy

Using a form of validation:

Best validation accuracy: 0.9534930417187959 Best learning rate: 0.001 Best momentum coefficient: 0.99

Model hyper-parameter selection

regularization parameter lambda, number of NN hidden layer units, number of hidden layers (if necessary), sigma, C, k, etc.. Systematic approach instead of just one or several randomly chosen values.

get the hyoer parammeters -> greed search cv

learning rate=0.001 Training accuracy: 0.9582252135365985 Testing accuracy: 0.9538174011090002

learning rate=0.01 Training accuracy: 0.22367282949508055 Testing accuracy: 0.22297391223761642

learning rate=0.0001 Training accuracy: 0.9484365105107888 Testing accuracy: 0.9454381168620546

For a classification problem, you need to present the confusion Matrix

(accuracy, precision, recall, F1 score, etc.).

    apple   axe     book    house   sword

apple 28078 320 215 128 131 axe 122 23692 315 140 521 book 181 522 22846 203 115 house 327 353 677 25550 121 sword 110 784 123 120 23792

apple Precision: 0.974321605 (28078 / (28078 + 122 + 181 + 327 + 110)) Recall: 0.972499307 (28078 / (28078 + 320 + 215 + 128 + 131)) F1 Score: 0.9734096031300774

axe Precision: 0.922909119 ( 23692/(320 + 23692 + 522 + 353 + 784)) Recall: 0.955707947 ( 23692 / (122 + 23692 + 315 + 140 + 521) ) F1 Score: 0.9390222151714112 book Precision: 0.94498676373 = 22846 / (215 + 315 + 22846 + 677 + 123) Recall: 0.95722126785 = 22846 /(181 + 522 + 22846 + 203 + 115) F1 Score: 0.9510646712260572

house Precision: 0.97739183657 = 25550 / (128 + 140 + 203 + 25550 + 120) Recall: 0.94531596862 = 25550 / (327 + 353 + 677 + 25550 + 121) F1 Score: 0.9610863473008548

sword Precision: 0.96401944894 = 23792 / (131 + 521 + 115 + 121 + 23792) Recall: 0.95439046893 = 23792 / (110 + 784 + 123 + 120 + 23792) F1 Score: 0.9591807938034582

Performance comparison between the models.

• logistic regression
• MLP nn

Results in graphical or table formats.

• Grafico de evolucao de accuracy
• Neural Network accuracy: 0.9737770794157801 for neurons = (784,300,60,50,40,10), 784 é bom por causa de ser 28*28 qualquer outro numero vai ser pior

Conclusions.

Problem complexity.

iamgens n podemos k fold crossvalidation

divide train, validacao, teste

ToDos

Notas

• Correct data spliting stupid Implementation