[mlpack] GSoC 2020 Ideas

[Andrei M]

Hello,

Thank you for the feedback. I submitted a draft proposal based on it.

Best,
Andrei

On Sat, Mar 21, 2020, 00:43 Marcus Edel <marcus.edel at fu-berlin.de> wrote:

> Hello Andrei,
>
> thanks for the update, I don't have anything to add, sounds totally
> reasonable
> to me. As an overall timeline, this could definitely work.
>
> Thanks,
> Marcus
>
> On 19. Mar 2020, at 13:07, Andrei M <mihalea.andrei at gmail.com> wrote:
>
> Hello again,
>
> Thank you for the feedback.
>
> After a longer though process, I decided I would like to implement the
> DeepLabv3+ model for semantic segmentation as part of the ANN models
> project. This implies several phases of implementation and this is the
> split I propose:
>
> Step 1:
> Implement a dataloader for a semantic segmentation dataset: This will be
> either Pascal VOC 2012 or ADE20K.
>
> Step 2:
> Implement an Xception model backbone, with atrous depth-wise separable
> convolutions. This is the backbone that makes the model yield the best
> performance, according to the original paper, overpassing the ResNet-101
> backbone.
>
> Step 3:
> a. Implement the encoder architecture of the model, which is a DeepLabv3,
> that uses the previously mentioned Xception backbone. This task also
> implies the building of the atrous spatial pyramid pooling module.
> b. Implement the decoder architecture, which is a simple architecture
> based on convolutions, which refine the segmentation results of the encoder
>
> Step 4:
> Train and test the implemented model on the selected dataset, then compare
> the results with the ones obtained in the paper. Visualize the results and
> create relevant plots and statistics.
>
> That would be a shorter version of my proposal.
>
> Best,
> Andrei
>
> On Wed, 11 Mar 2020 at 23:35, Marcus Edel <marcus.edel at fu-berlin.de>
> wrote:
>
>> Hello Andrei,
>>
>> 1. RL: I've taken a more in-depth look on the reinforcement learning
>> module. The
>> DQN, Double DQN and prioritized replay are already implemented, so as
>> part of
>> the rainbow the remaining components are Dueling networks, Multi-step
>> learning,
>> Distributional RL, Noisy. Therefore, I suggest finishing the
>> implementation of
>> the Rainbow DQN and then an implementation of the ACKTR algorithm.
>>
>> Sounds totally reasonable to me.
>>
>> 2. Applications of ANN: Implementing a U-Net or DeepLabv3 architecture for
>> semantic segmentation.
>>
>> I like both models, also good that you mentioned you like to focus on
>> either
>> U-Net or DeepLabv3.
>>
>> I would like to know if the ideas above would make enough for a summer
>> project
>> for each of the two sections.
>>
>> Definitely, a big part of each project is documentation and testing,
>> writing
>> good tests takes time.
>>
>> Let me know if I should clarify anything further.
>>
>> Thanks,
>> Marcus
>>
>> On 10. Mar 2020, at 15:50, Andrei M <mihalea.andrei at gmail.com> wrote:
>>
>> Hello,
>>
>> Thank you for the response.
>>
>> I've been thinking more about the ideas for the GSoC and I've established
>> a top 2 I'd like to work on: Reinforcement learning or Applications of ANN.
>> (I'll only select one for the final proposal)
>>
>> 1. RL: I've taken a more in-depth look on the reinforcement learning
>> module. The DQN, Double DQN and prioritized replay are already implemented,
>> so as part of the rainbow the remaining components are Dueling networks,
>> Multi-step learning, Distributional RL, Noisy. Therefore, I suggest
>> finishing the implementation of the Rainbow DQN and then an implementation
>> of the ACKTR algorithm.
>>
>> 2. Applications of ANN: Implementing a U-Net or DeepLabv3 architecture
>> for semantic segmentation.
>>
>> I would like to know if the ideas above would make enough for a summer
>> project for each of the two sections.
>>
>> Thank you,
>> Andrei
>>
>> On Mon, Mar 9, 2020 at 1:22 AM Marcus Edel <marcus.edel at fu-berlin.de>
>> wrote:
>>
>>> Hello Andrei,
>>>
>>> welcome and thanks for you interest. Looks like you already brainstormed
>>> about
>>> the ideas, that great. I think each method you proposed made sense,
>>> there is
>>> alrady a PR open for PPO (https://github.com/mlpack/mlpack/pull/1912)
>>> which is
>>> very close to being merged, so I think you can remove that from the list.
>>>
>>> Also, I think both ideas could be combined, like if you add a new layer
>>> to the
>>> codebase. That said, we don't have project priorities, so you a free to
>>> go with
>>> anything you find interesting.
>>>
>>> Let me know if I should clarify anything.
>>>
>>> Thanks,
>>> Marcus
>>>
>>> On 6. Mar 2020, at 15:47, Andrei M <mihalea.andrei at gmail.com> wrote:
>>>
>>> Hello,
>>>
>>> I'm a second year master's degree student in the field of artificial
>>> intelligence and I've been thinking about applying to Google Summer of Code
>>> for this summer and mlpack is the project I want to work on.
>>>
>>> I've spent the last few weeks to get familiar with the code base and
>>> write some code for a new feature (a loss function that wasn't
>>> implemented). There are several ideas in the list that peaked my interest
>>> and I consider them equally interesting: reinforcement learning, essential
>>> deep learning modules, application of ANN algorithms implemented in mlpack
>>> and improvisation and implementation of ANN modules.
>>>
>>> I think these ideas would fit well for me since I've been implementing
>>> neural networks such as DQN, Double DQN, Dueling networks, GANS and several
>>> others in PyTorch and I've also been in touch with the state-of-art
>>> research in various fields, like the ones mentioned above. Therefore, I
>>> think I would equally enjoy working on the reinforcement learning path and
>>> working on bringing features and modules that are present in other
>>> libraries, like PyTorch.
>>>
>>> Below are some summaries of the ideas I'm thinking about:
>>> *Reinforcement learning*: Here I would like to work on Rainbow and
>>> Proximal Policy Optimization Algorithms, train and test them on different
>>> environments and empirically show their advantages and disadvantages (for
>>> example how double DQN can reduce the overestimation problem that appears
>>> in DQN).
>>> *Application of ANN algorithms implemented in mlpack*: For this idea, I
>>> have two options that come to my mind: first one is implementing a sequence
>>> to sequence model for language translation and the other consists of
>>> implementing U-Net like architectures which are usually employed for
>>> segmentation tasks or depth prediction.
>>> *Essential deep learning modules*: The plan I propose for this idea is
>>> implementing some of the GAN architectures that aren't yet implemented,
>>> starting from the first types of GANs that appeared, like conditional GANs
>>> and info GANs, then advancing to more modern ones, trying to obtain and
>>> visualize the results shown in the papers they've been presented on.
>>>
>>> I would also like to know what are the features with high priority for
>>> mlpack to have and if you have any suggestion on what I should propose to
>>> match these priorities.
>>>
>>> Also, can more ideas be proposed in a single application?
>>> Any feedback and suggestions are appreciated.
>>>
>>> Best,
>>> Andrei
>>> _______________________________________________
>>> mlpack mailing list
>>> mlpack at lists.mlpack.org
>>> http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
>>>
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/20200327/cc3a67ab/attachment-0001.htm>