Recap and Conclusion

Recap

• What is Computer Vision? building a system that can understand and process visual information

• Main approaches to Computer Vision

• semantic segmentation using Fully Convolutional Networks (FCN) 🤔

What’s next?

1. we will discover interesting new types of deep models Neural ODE NIPS 2018 $$ h_{t+1} = h_t + f(h_t, \theta_t) \Rightarrow \frac{dh}{dt} = f(h(t), t, \theta) $$

2. Deep Learning will find new applications

• AI4S
• DL4CS, like using DL training a hash function to improve the hash table
• DL4Mathematics, auto theorem proving

3. Deep Learning will use more data and more computional power

Problems and Challenges

Models are biased

eg: Vector Arithmetic with Word Vectors, fit biased dataset to a model leads to poor performance; Economic Bias (usually train model on the wealthier western family?); Racial Bias

maybe making ML work for everyone is a big challenge

New theory?

Empirical Mysteries

• Good Subnetwork 😲

• Generalization, traditional Ai STAT says that if a model can fit on random dataset, maybe the model is too large to generalize to new dataset 🤔

• Double Descent 😱, seems like as we scale up, the overfitting curve does not behave as expected

Need for more labeled data

• low-shot learning
• using self supervised learning to learn from unlabeled data?
  • jigzaw puzzles ECCV 2016
  • colorization
  • Inpainting

Deep Learning does not “understand” the world

• LLM lack of common sense, at the age of GPT2 😆, now we are at 2025…… Can we say that those LLMs are “smart” enough to understand the world?

• “The Elephant in the Room”

I really want to see those problems in 2025! 😋