Discussion for Pop Quiz

deep_learning

#1

This thread here is for discussions of approaches asked in article

Feel free to post your answers here.


#3

Scenario-1
In case of machine learning what are the features captured by the camera after that on the basis of previous experience that at what angle they have steered the car but in case of deep learning they will consider the past work but after that they will analyse what sort of impact it has done previously and consider more and more aspect .According to me here machine learning is better than deep learning because data is less so we can move to machine learning

Scenario-2
In this by seeing his credentials we can,t say exactly whether we can grant loan or not, because in case of machine learning it consider the previous records of the people who are having the same credentials and whether the loan has granted to them or not, if they received then according to machine learning we will grant the loan. But in case of deep the above procedure will be also done the consider the above information but at the same time the people who has received the loan whether they return the loan amount within time or whether they have defaulted or not. Even it will do the analysis on the behaviour of the people
Here I think we have to apply deep learning.

Scenario -3
Here we will apply machine learning


#5

Hi @RAJIT Can you explain a bit for scenario 3 as to why you would apply ML over DL?


#6

I think for scenario-3 we apply machine learning because in this we need processing fast and in case of translation we need the meaning of the words which are already available.


#7

Scenario 1.
We must be using Deep Learning to solve this case. In this scenario, every tiny detail need to be taken care of. Since it’s a software we’re building for the automatic steering of the wheel, extracting the features manually may leave us with various loop holes. Also, the abstraction of the essential features such as the pixel value, the road type, the object in front etc, which are hidden and cannot be coded directly into the program, isn’t a job of a regular machine learning algorithm. A deep leaning algorithm would be the most efficient here to extract out the perfect features.

Scenario 2.
This is a normal classification problem which can be solved using the traditional ML algorithms. Since, the past history of the person and the credentials is known a traditional ML algorithm can be set using the features to classify the person as defaulter or not. Setting up features for this case is much easier as compared to the first case.

Scenario 3.
This is a Natural Language Processing (NLP) problem which can be solved using deep learning technique.


#8
  1. For first scenario, deep learning need to be used as minute details will be involved in future. We can’t predict each and every minute detail beforehand(as done in machine learning). Hence, even though the training of model takes time, deep learning would be best for future.

  2. Having domain knowledge of banking and the factors to be considered for bank loan, we can use machine learning. As the factors to be considered for it will remain same most of the time.

  3. Deep learning


#9

Hi everybody! Here are my two pence:

Scenario 1.
Considering the quantity of data you are receiving every moment, and the fact that you need to be very precise in analyzing them otherwise you could have an accident, DL is the way to go IMHO. Training with classical ML techniques cannot garantee us that we took every detail into account: it is faster, but it’s more dangerous. Maybe if we train our car to drive in a city we could have a better performance, but when we leave city, or even that specific city, performance would drop drastically. DL should perform in average better.

Scenario 2.
This is the typical True/False regression problem, not unlike fraud detection or Titanic survival in kaggle. After preprocessing the data (i.e. oversampling can be needed in these cases), I would apply logistic regression and decision trees (or an ensemble of those): these are proven trusty methods that reliably perform very well.

Scenario 3.
Again, in this case Natural Language Processing (NLP) is a very good technique used everyday, but maybe many “hues” of any language speech can be lost in translation. DL could retrieve them, but I’m actually more concerned about how to explain to a Russian candidate why the people think the way they think if DL does not offer an interpretation of the results. I think in the end I would opt for ML.


#10

Hello @RAJIT.

I think language translation is not only translating word by word. It is more than that. It includes Linguistic topology and many more things. The sentence should sound correct in translated language. for example, take this simple example for understanding.

Hindi to English
H: आप क्या करते हैं?
Its correct english translation would be
E: What do you do?

Whereas if we do word by translation ,it would result in :

You what do ?

Hope it explains.


#11

Scenario-1
I will be using deep learning as its a real time scenario and during testing it executes much faster then neural nets when having large no of training data. Here we don’t have any predefined features.

Scenario-2
I will be applying machine learning since I have got predefined features and one may get to know on what basis the person has got loan.

Scenario -3
Not sure.


#12

In alignment wth the explanation from the article:

Senario 1 & 3 : Object Recognition and NLP
We need high accuracy
Manually handcrafting features is difficult
Interpretability of model isnt required.
Huge amount of data could be made available.
Hence deep learning.

Scenario 2. Loan prediction
In this case feature engineering is quite possible.
Model interpretation is very much required.
Associated features and their values may keep changing from time to time so we need to remodel at every fixed intervals. And this should be done in timely manner.
All of this votes for ML.


#13

Scenario 1:
Using deep learning, the system will identify parameters those are important to make the perfect turn. There could be numerous combination of parameters that could lead to a perfect turn.
Using machine learning, it will need several hours of work to configure the parameters for a successful turn. These rules will have to be manually fed for the model to run.

Considering the volume of data will be high, and self-driving cars need real-time performance of the software, deep learning will best suit

Scenario 2:
Using deep learning, once all data is fed system will identify the parameters and a provide the result for eligibility.

Using machine learning, all parameters can will have to be defined to provide eligibility results.

ML will best suit here considering the limited amout of parameters and need for an expertise to define.

Scenario 3:
Using deep learning, the system will have to be trained with several results to improve accuracy

Using ML, lot of rules need to be build and fed into the system

Deep Learning would be suggested considering the ease for the system to identify translation logic and the amount of variations a language can have. While ML can also be used, however it will need continuous monotoring to improve accuracy


#14

Scenario 1 - so in machine learning it will take in the data and then recognise the objects as shown in problem solving approach. So first you will do is object detection and then object recognition. After that one can apply feature engineering and thus detect the angle by which it should rotate. In case of deep learning, it’ll try to learn high-level features from data and hence predict the angle. I feel that deep learning will be more effective as it’ll learn from the high level features and predict the output

Scenario 2 - This again can be done by machine learning and will be better than deep learning as it will predict the output from the manually entered data given a person’s credentials and background information.

Scenario 3- it is similar to google translator. So when machine learning is used, it translates phrase by phrase conversion whereas using deep learning the translator is much more efficient and closer to the human translation.