125 points by divelab 6 months ago flag hide 19 comments
deeplearningdave 6 months ago next
Fascinating article! I never thought about using submarine visualizations in neural networks.
mlmentor 6 months ago next
I know right? I initially thought it was a joke but the results are undeniable.
deeplearningdave 6 months ago prev next
MLMentor, I agree. It certainly opens up new avenues for research.
matrixmaker 6 months ago prev next
I wonder how this could be applied to other areas such as image recognition or even natural language processing?
mlmentor 6 months ago next
matrixmaker, neural networks have proven to be incredibly versatile so I'd like to see the outcomes.
quantum_hacker 6 months ago prev next
Great article! Really works well in explaining the concepts. I have one question about the optimization method. Any insights?
deeplearningdave 6 months ago next
quantum_hacker, we actually used the Adam optimizer with learning rate decay. More details in the paper.
supervised_learner 6 months ago prev next
How come nobody mentioned tensor manipulation tools for visualization? Matrix mappers or tensorroyale are great candidates for this task.
deeplearningdave 6 months ago next
supervised_learner, great point! We've been focusing mostly on deep dive demonstations so far - but definitely something to explore further. Our goal was to introduce the concept.
thepracticaljim 6 months ago prev next
Impressive work! Keep up the great work. I'd love to see more examples illustrated in different settings like unsupervised or reinforcement learning.
deeplearningdave 6 months ago next
thepracticaljim, thanks for the suggestion! I'd like to think of this as a starting point and not the final solution. Hoping to discuss many more ideas like these to delve further into neural networks.
neuron_splitter 6 months ago prev next
What justification led to the choice of visualization tools and techniques applied in your work? New visualizers and trends pop up frequently.
deeplearningdave 6 months ago next
neuron_splitter, that's a fair question. We based our choices on a solid theoretical background, Accessibility, and the capacity for further development. We're excited to see how things progress.
data_oracle 6 months ago prev next
Can you share more practical usage scenarios and success stories from industry? It would make the concepts more relatable and applicable in real-world situations.
deeplearningdave 6 months ago next
data_oracle, Absolutely on the case. We've lined up a series of case studies from top firms in the space (Google, Tesla, Amazon) for a deeper dive into and behind-the-scenes look at the application of these techniques.
guru_j 6 months ago prev next
How would this type of visualization help in detecting adversarial attacks and improving robustness?
deeplearningdave 6 months ago next
guru_j, a clever use of submarine visualization could help spot red flags or inconsistencies in the network's inner workings to prevent adversarial attacks. Though it's a delicate balance, the goal is to be vigilant without mitigating the road to discovery.
weightwatcher34 6 months ago prev next
As a long-time follower of neural network development, I'm always eager to come across innovative techniques that offer new angles to uncover network behavior and relationships. This work ticks all the boxes.
deeplearningdave 6 months ago next
weightwatcher34, absolutely! The idea is to share insights gained and fostering discussions to even deeper understandings and further advancements.