GQ1. DL is part of the solution GQ2. "vectors, not symbols” is false dichotomy diff functions: yes, in part Operations a la logic, we do need (contra your view) GQ3. Outputs of deep learning may serve as input to reasoning; symbolic techniques needed for some inferences.
-
The ML community is working hard on the robustness question. We know that stochastic gradient descent sometimes confers robustness and other times fails miserably. We are exploring many avenues: causal models, adversarial training, improved models, etc.
Note of course difference between deep learning and ML; IMHO deep learning is useful, but just a subset of final ML toolbox. Ultimately ML will need to encompass many techniques (some not invented yet), and integrate them all to achieve robustness.
-
-
DL = { models are modular & non-linear, minimizes an objective function, computes gradient "analytically", uses gradients to minimize }. (This applies to all learning paradigms: supervised, unsupervised,...) Which of these 4 pillars do you propose to replace, and by what?
-
@AvilaGarcez said what i would have said: not replace but supplement with tools for symbolic operations - 3 more replies
New conversation -
-
-
This is an important hypothesis, but the DL community keeps demonstrating that anything we can do in machine learning they can usually do better and in a more unified way. (At least if "better" means "accurate".) As a decision tree ensembles person, I find this frustrating :-)
- 7 more replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.