what are the techniques used to reduce the recall or precision if overfit in logistic regression?

Veeru(VeeraNancharaiah Javvaji)

jsriveeru@gmail.com

brusuvain@gmail.com

hello, i had missed the last doubt class and the last session due to work, and im still lagging behind in the projects, could you help me?

karankarnik47@gmail.com

Veeru(VeeraNancharaiah Javvaji)

jsriveeru@gmail.com

Are models always available for free or there is any licensing as well? for e.g. by corporates?

Shreyas Phatak

hishreyas@gmail.com

Is it possible to have few layers in one language and few in some other language. Eg. can we create a model where 3 layers are in C++ but rest 2 in Python.

manoj.gupta.91@gmail.com

in deep learning what is the alternative for ML clustering (un-supervised)? wanted to know any technique for anomaly detection in deep learning ?

sudhir.m.shetty@gmail.com

arpitvw16@gmail.com

arpitvw16@gmail.com

siliconfish@yahoo.com

Veeru(VeeraNancharaiah Javvaji)

jsriveeru@gmail.com

followup question - Why larger weight is equated to overfitting? how did we arrive at that?

Puneet Rastogi

puneetrstg@gmail.com

parmarvedpro5@gmail.com

The workflow can be broken down into following basic steps:
Training a machine learning model on a local system.
Wrapping the inference logic into a flask application.
Using docker to containerize the flask application.
Hosting the docker container on an AWS ec2 instance and consuming the web-service.

SGD (two statements from book)
(1) Note that since instances are picked randomly, some instances may be picked several times per epoch, while others may not be picked at all.
(2) If you want to be sure that the algorithm goes through every instance at each epoch, another approach is to shuffle the training set.
Does second statement says about going through all instances or not to repeat an instance in single epoch because SGD does not compute all instances it works with one random instance in each step. So not to repeat an instance in single epoch is what I understood from this statement.

manoj.gupta.91@gmail.com

I believe L1, L2 and Elastic net regularization technique also works with Neural Networksas they are conceptually related to Gradient Descent? Is that correct

Puneet Rastogi

puneetrstg@gmail.com

Logistic Regression is a classification problem ... so can we use confusion matrix

dev.feb88@gmail.com

I am confused with the larger weight resulting in overfitting, when we apply gradient descent, weights will be learnt so that appropriate model which represent the data. In the query regarding the larger weights are we referring to the initilization to larger values or larger step size. We only perform normalization on the features not on weights. So my doubt is how is larger weight resulting in overfitting.

vinods.kumar@gmail.com

vinods.kumar@gmail.com

vinods.kumar@gmail.com

One qq -so rate of change, we need to keep as low as possible and almost consistent?

krajiv.2018@gmail.com

vinods.kumar@gmail.com

you explanation regarding higher weight move towards overfitting , if python or any other language is having round off till 5 digit or any number of digit , after decimal in that case also there will bemore degree of freedom ,

arpitvw16@gmail.com

kupadhy@gmail.com

arpitvw16@gmail.com

arpitvw16@gmail.com

arpitvw16@gmail.com

arpitvw16@gmail.com

arpitvw16@gmail.com

as we going to deploy any model to mobile , example : face filters that we have in mobile ?

arpitvw16@gmail.com

Sorry, one dumb question. Is Optimizers all about weight? and Gradient Descent furhter optimize?

krajiv.2018@gmail.com

Is it better to use SGD or Mini-batch SGD to reach to global minima. Since both of these may not settle down to perfect global minima they may settle somewhere global minima. Once we are done we can use GD to reach to exact global minima.

manoj.gupta.91@gmail.com

lalit.kathuria76@gmail.com

qq - Rolling ball once reached surface & with friction, the accelaration will come down? Slope will be less

krajiv.2018@gmail.com

Can’t we obtain this just my having a dynamic learning rate which start with a high value and then decreases as we get closer to the global minima?

Domenico Fioravanti

nicodom@gmail.com

Domenico Fioravanti

nicodom@gmail.com

Sandeep, I saw equation like m<-Bm- (minus)n.. but in your case it's positive. can you please brief on plus vs minus?

krajiv.2018@gmail.com

in Gradient descent, when learning rate was higher, we could jump over the minima, so how is the momentum handling this.

vinods.kumar@gmail.com

should default optimizer to be "momentum"? May be you will be covering next. if yes, please ignore my question.

krajiv.2018@gmail.com

Pushkraj Gaikwad

gaikwad.pushkraj@gmail.com

If we are using GD for Linear Regression how can we handle case of local minima?

lalit.kathuria76@gmail.com

NAG would work only if momentum vector pointing in right direction and then only NAG would work perfect? So, assuming we need to know momentum vector is pointing in right direction before picking NAG?

krajiv.2018@gmail.com

karankarnik47@gmail.com

shalini.cse16@nituk.ac.in

parmarvedpro5@gmail.com

Sorry Sandeep, not able to catch much over AdaGrad. Can you please make it simpler to understand? May be it's just a problem for me.

krajiv.2018@gmail.com

2 doubts Sandeep: What happens if data is dense, and also the learning rate will shrink over time ...how does ADagrad deal with it

anmolck@gmail.com

so what is the summary in terms of which optimizer should be used in the first place ?

Prakhar Prasad

prakhar.prasad@gmail.com

What I meant -> It has lower learning rate for frequent features and higher learning rate for infrequent features

anmolck@gmail.com

anmolck@gmail.com