What we will do next?

Some thoughts about future research questions and our final goals towards the unknown.

1. What works are scientifically meaningful for computational neuroscience? Specially, what can we do for network neuroscience?

2. Why are we doing spiking neural networks or neuromorphic computing? What will it finally lead to?

3. What can we do to integrate artificial intelligence and neuroscience? What are the critical problems in neuroscience that can be solved by artificial intelligence? What are the phenomenon and mechanism in artificial neuroscience that requires the integration of neuroscience?

4. What makes human beings in the current stages? In the previous research path, our paradigm is to investigate a specific task through the reductionism. However, the LLM’s motivation may be the opposite. It is not because of the complex infrastructure of the brain so that it can execute complex tasks, but the complex tasks that shapes the high intelligence.

5. Some big thoughts?

   1. The AI develops from decision-making to generative AI. Reasoning from the first-principle, what can be the next step?
   2. The physics and AI development has several key time points. Regarding the physics, the emergence of Newton’s three principles for mechanics, the quantum mechanics and relative theory permanently change the landscape of the physics. Regarding the AI, the prosperity of DNN and associated hardware boost its development waves by waves. What can be the changing point for neuroscience?
   3. The next-word prediction serves as a good pre-train task for the LLMs. What types of properties does this task capture so that it is valid and efficient ? Regarding the general artificial intelligence, what kinds of pre-train tasks are appropriate? For the vision-LM or multimodal-LM, what kinds of pre-train tasks are appropriate?
   4. Transformers are good for predicting the next word. Yet is it the proper design for intelligence? Relatively, what type of data structure matches the network infrastructure?

6. Suppose we have a model that can predict brain’s response to the given stimulus. What kind of questions about encoding and decoding can we ask? What can serve as a fundamental model for neural encoding?

7. If we take the SNN as a special type of quantization, can we derive its advantages on the representation and parallelism? For the representation, we mean that compared to the quantization of the low-bit setup, what benefit it brings when we divide the 1-time expression into multiple steps? For the parallelism, we mean that what kind of computational benefit it brings with the fixed size spikes?

8. The brain uses simple and imprecise mechanism to support the learning. Moreover, the underline infrastructure appears to be dominated by the genes rather than the data, which suggests that they might be a generic initialization for deep neural networks. Or, the pre-training part of a large ANN might be realized through simple learning rules.

9. Quick Thoughts:

   1. Combine Transformer pretraining and quantization approaches in a spike manner. Use the spiking mechanism to control masks.
   2. Regarding the behavioral experiments, we can provide detailed description of the subject and then has the character to perform the task. Can we have an example here?
   3. What are the common space for vision-language? It should be not only defined by similarity in the representation (CLIP), but also by certain tasks. The CLIP maps an image into a space aligned with the caption, which is typically a subspace of the information that can be represented by the image. So our question can be: how to generate questions/dialogue for the given images? What are the questions that can expands the image space?

10. What are good research works in AI?

    1. It established new learning paradigm, like RL, generative AI, CLIP, etc…
    2. It proposes new fundamental learning infrastructures, like CNN, ResNet, Transformer, stableDiff, etc…

11. The bitter lesson: what should we believe to generate intelligence.

    1. Our brain is a trajectory full of randomness in its current stage. Not all mechanisms in the brain are useful. Many mechanisms appear only because of the compromise to the underline infrastructure and only works in that way.
    2. Considering the evolution theory, maybe it is random for the brain to appear in its current state, but it could be of high probability to show up under some scalable rules. Designing scalable infrastructure and identifying scalable principles may be more effective in building more intelligent agents.
    3. The human’s utilization of information and energy is efficient. However, we should interpret it as a distinct scalable solution rather than a perfect model to follow. It makes no sense in the computation to say something is brain-inspired without digging into the underline computational rules.
    4. What’s the difference between the scalability of current AIs and those in the previous decades? How should be design research plan? Or more broadly, how should we design our intelligent agent? Is there any scalable plan that works for all tasks?
    5. The evolution leads to diversities in the function. For example, there are different types of eyes. These eyes are represented by genes. How can we build the similar “design” or “encode” system? Or even more, can we talk whether such a system is efficient? If it is, how can we design one with similar efficiency?
    6. We can either build AI through connection and symbols. BUT, can we do both?

12. How to realize the scaling law for brain mapping?

    1. Scan time, resolution, task?
    2. What does the brain mapping tell us from the view of machine learning?
    3. What will be the more effective paradigm for neural intervention?

13. An interesting idea of abstraction for AI: we can have AI to draw a human body or other figures, then have some professional painters to judge the body-science and highlight-shade.