Symbolic Knowledge about y1, y2

Two Birds, One Neural Network
1

Hi everyone,

maybe I’ve missed it: Is it forbidden to encode symbolic knowledge about y1 and y2 (i.e., that y1(x0, x1) = cos(49*x0 + 42*x1) and that y2(x0, x1) = cos(56*x0 + 63*x1)) into our solution/network?

All the best,
Thomas

2

Hello @wedenigt,
You can incorporate symbolic knowledge about y1 and y2 to build your network wisely, but not as a replacement for the solution. Preprocessing, postprocessing, any adjustments, etc. of results generated by the neural network are not allowed.

Xeek Team

3

Hi Guys
I think it is likely that I can do this analytically. I thought machine learning was for problems too hard to do analytically.
cheers, Eric

4

Hi,
I am curious if it is permissible to perform data augmentation by generating additional data points?

5

Hello @cklwankaggle
You could generate additional samples for gaining insights. However, for the final evaluation, you should provide a reproducible solution that is based on the provided datasets.

Best regards,
Onward Team

6

@discourse-admin

Could you please elaborate more on this? I’m particularly interested in “preprocessing”. What specific operations are considered off-limits in this context?

1 Like
7

Hello @daisuke0530,
Thank you for your question. Regarding preprocessing, it’s important to consider the overall goal of keeping the fundamental nature of the data. Generally, any operation that significantly alters the fundamental nature or inherent characteristics of the data might be considered off-limits. If you have some particular operations in mind, we could discuss them more precisely.

Best regards,
Onward Team.

8

@discourse-admin

Thank you for your response. I have two scenarios in mind that I believe should both be permissible:

  1. Converting ‘x’ into slope and intercept, and vice versa. (For prediction, generate slope and intercept using a neural network, then convert these into a line.)
  2. Converting ‘y’ into period and phase, and vice versa.
1 Like
9

Hello @daisuke0530
You are correct that these types of transformations do not influence the nature of the data, so they could be used.

Good luck in solving the challenge.
Onward Team

1 Like
10

@discourse-admin,

Thank you for your answer; it has clarified for me.
I have another question regarding the definition of a neural network. I’m considering three scenarios, which I belive all should be valid:

  1. Similar to CNNs in image classification, assign a score to each potential option (x) and select the top 30 items. (Variant of this was already answered here, but let me double check)
  2. Similar to R-CNN in object detection, generate a series of proposals (x), score each one, and then choose the top 30 items.
  3. Similar to GPT for text generation, assign a score to each token (representing a possible x in this context) and select the best as a prediction. This process would be repeated 30 times.

Are they all valid?

1 Like
11

Hello, @daisuke0530

All suggested option will be considered as valid.

Best wishes,
Onward Team

2 Likes
12

Can we generate samples and select only which meet the criteria (not to maximize diversity), does it come under post-processing as we are not actually altering the results generated by neural network?

Regards,
VC

1 Like
13

Hello @chepurivishwas360
Your goal should be to design and tune a neural network to generate results that meet the conditions, rather than doing this job yourself. So, the type of selection you described, where only samples meeting the criteria are chosen, would be considered a forced adjustment of results. This type of post-processing is forbidden and could lead to disqualification.
Best regards,
Onward Team

1 Like