Need for an alternative training method to predictive coding that avoids backpropagation
Problem
Current deep learning models rely on backpropagation for gradient computation, which is problematic for discrete components such as argmax, MoE, and VQ‑VAE, as well as for biologically plausible learning mechanisms; users seek alternatives that do not depend on traditional backpropagation.
Pain points
- Lack of differentiable training methods for discrete layers or model components
- Desire for a more biologically plausible credit assignment mechanism
- Desire to achieve effective gradient estimation without using backpropagation
Questions to verify
- Have you encountered situations in real projects where discrete layers or model components made gradient computation difficult?
- Do you prefer using biologically plausible learning rules over traditional backpropagation?
- What are your expectations regarding the feasibility of using predictive coding as an alternative training method?
Evidence
quote
Can it be better than existing training methods for discrete components like argmax, MoE or VQ-VAE?
quote
The brain doesn't need to do back propagation. Back propagation is a simulation of gradient descent which the universe achieved naturally by its own devices.
Disclaimer: this card is distilled by AI from community discussions as “pain point evidence”. It is a starting point for need analysis and validation, not a “validated market need”. AI has analytical limits; please open the original discussion above to manually verify before key decisions.